diff options
Diffstat (limited to 'drivers/lguest')
| -rw-r--r-- | drivers/lguest/Makefile | 12 | ||||
| -rw-r--r-- | drivers/lguest/README | 47 | ||||
| -rw-r--r-- | drivers/lguest/core.c | 357 | ||||
| -rw-r--r-- | drivers/lguest/hypercalls.c | 144 | ||||
| -rw-r--r-- | drivers/lguest/interrupts_and_traps.c | 212 | ||||
| -rw-r--r-- | drivers/lguest/io.c | 265 | ||||
| -rw-r--r-- | drivers/lguest/lg.h | 47 | ||||
| -rw-r--r-- | drivers/lguest/lguest.c | 535 | ||||
| -rw-r--r-- | drivers/lguest/lguest_asm.S | 71 | ||||
| -rw-r--r-- | drivers/lguest/lguest_bus.c | 75 | ||||
| -rw-r--r-- | drivers/lguest/lguest_user.c | 166 | ||||
| -rw-r--r-- | drivers/lguest/page_tables.c | 329 | ||||
| -rw-r--r-- | drivers/lguest/segments.c | 126 | ||||
| -rw-r--r-- | drivers/lguest/switcher.S | 284 |
14 files changed, 2433 insertions, 237 deletions
diff --git a/drivers/lguest/Makefile b/drivers/lguest/Makefile index 55382c7..e504747 100644 --- a/drivers/lguest/Makefile +++ b/drivers/lguest/Makefile @@ -5,3 +5,15 @@ obj-$(CONFIG_LGUEST_GUEST) += lguest.o lguest_asm.o lguest_bus.o obj-$(CONFIG_LGUEST) += lg.o lg-y := core.o hypercalls.o page_tables.o interrupts_and_traps.o \ segments.o io.o lguest_user.o switcher.o + +Preparation Preparation!: PREFIX=P +Guest: PREFIX=G +Drivers: PREFIX=D +Launcher: PREFIX=L +Host: PREFIX=H +Switcher: PREFIX=S +Mastery: PREFIX=M +Beer: + @for f in Preparation Guest Drivers Launcher Host Switcher Mastery; do echo "{==- $$f -==}"; make -s $$f; done; echo "{==-==}" +Preparation Preparation! Guest Drivers Launcher Host Switcher Mastery: + @sh ../../Documentation/lguest/extract $(PREFIX) `find ../../* -name '*.[chS]' -wholename '*lguest*'` diff --git a/drivers/lguest/README b/drivers/lguest/README new file mode 100644 index 0000000..b7db39a --- /dev/null +++ b/drivers/lguest/README @@ -0,0 +1,47 @@ +Welcome, friend reader, to lguest. + +Lguest is an adventure, with you, the reader, as Hero. I can't think of many +5000-line projects which offer both such capability and glimpses of future +potential; it is an exciting time to be delving into the source! + +But be warned; this is an arduous journey of several hours or more! And as we +know, all true Heroes are driven by a Noble Goal. Thus I offer a Beer (or +equivalent) to anyone I meet who has completed this documentation. + +So get comfortable and keep your wits about you (both quick and humorous). +Along your way to the Noble Goal, you will also gain masterly insight into +lguest, and hypervisors and x86 virtualization in general. + +Our Quest is in seven parts: (best read with C highlighting turned on) + +I) Preparation + - In which our potential hero is flown quickly over the landscape for a + taste of its scope. Suitable for the armchair coders and other such + persons of faint constitution. + +II) Guest + - Where we encounter the first tantalising wisps of code, and come to + understand the details of the life of a Guest kernel. + +III) Drivers + - Whereby the Guest finds its voice and become useful, and our + understanding of the Guest is completed. + +IV) Launcher + - Where we trace back to the creation of the Guest, and thus begin our + understanding of the Host. + +V) Host + - Where we master the Host code, through a long and tortuous journey. + Indeed, it is here that our hero is tested in the Bit of Despair. + +VI) Switcher + - Where our understanding of the intertwined nature of Guests and Hosts + is completed. + +VII) Mastery + - Where our fully fledged hero grapples with the Great Question: + "What next?" + +make Preparation! +Rusty Russell. diff --git a/drivers/lguest/core.c b/drivers/lguest/core.c index ce909ec..0a46e88 100644 --- a/drivers/lguest/core.c +++ b/drivers/lguest/core.c @@ -1,5 +1,8 @@ -/* World's simplest hypervisor, to test paravirt_ops and show - * unbelievers that virtualization is the future. Plus, it's fun! */ +/*P:400 This contains run_guest() which actually calls into the Host<->Guest + * Switcher and analyzes the return, such as determining if the Guest wants the + * Host to do something. This file also contains useful helper routines, and a + * couple of non-obvious setup and teardown pieces which were implemented after + * days of debugging pain. :*/ #include <linux/module.h> #include <linux/stringify.h> #include <linux/stddef.h> @@ -61,11 +64,33 @@ static struct lguest_pages *lguest_pages(unsigned int cpu) (SWITCHER_ADDR + SHARED_SWITCHER_PAGES*PAGE_SIZE))[cpu]); } +/*H:010 We need to set up the Switcher at a high virtual address. Remember the + * Switcher is a few hundred bytes of assembler code which actually changes the + * CPU to run the Guest, and then changes back to the Host when a trap or + * interrupt happens. + * + * The Switcher code must be at the same virtual address in the Guest as the + * Host since it will be running as the switchover occurs. + * + * Trying to map memory at a particular address is an unusual thing to do, so + * it's not a simple one-liner. We also set up the per-cpu parts of the + * Switcher here. + */ static __init int map_switcher(void) { int i, err; struct page **pagep; + /* + * Map the Switcher in to high memory. + * + * It turns out that if we choose the address 0xFFC00000 (4MB under the + * top virtual address), it makes setting up the page tables really + * easy. + */ + + /* We allocate an array of "struct page"s. map_vm_area() wants the + * pages in this form, rather than just an array of pointers. */ switcher_page = kmalloc(sizeof(switcher_page[0])*TOTAL_SWITCHER_PAGES, GFP_KERNEL); if (!switcher_page) { @@ -73,6 +98,8 @@ static __init int map_switcher(void) goto out; } + /* Now we actually allocate the pages. The Guest will see these pages, + * so we make sure they're zeroed. */ for (i = 0; i < TOTAL_SWITCHER_PAGES; i++) { unsigned long addr = get_zeroed_page(GFP_KERNEL); if (!addr) { @@ -82,6 +109,9 @@ static __init int map_switcher(void) switcher_page[i] = virt_to_page(addr); } + /* Now we reserve the "virtual memory area" we want: 0xFFC00000 + * (SWITCHER_ADDR). We might not get it in theory, but in practice + * it's worked so far. */ switcher_vma = __get_vm_area(TOTAL_SWITCHER_PAGES * PAGE_SIZE, VM_ALLOC, SWITCHER_ADDR, VMALLOC_END); if (!switcher_vma) { @@ -90,49 +120,105 @@ static __init int map_switcher(void) goto free_pages; } + /* This code actually sets up the pages we've allocated to appear at + * SWITCHER_ADDR. map_vm_area() takes the vma we allocated above, the + * kind of pages we're mapping (kernel pages), and a pointer to our + * array of struct pages. It increments that pointer, but we don't + * care. */ pagep = switcher_page; err = map_vm_area(switcher_vma, PAGE_KERNEL, &pagep); if (err) { printk("lguest: map_vm_area failed: %i\n", err); goto free_vma; } + + /* Now the switcher is mapped at the right address, we can't fail! + * Copy in the compiled-in Switcher code (from switcher.S). */ memcpy(switcher_vma->addr, start_switcher_text, end_switcher_text - start_switcher_text); - /* Fix up IDT entries to point into copied text. */ + /* Most of the switcher.S doesn't care that it's been moved; on Intel, + * jumps are relative, and it doesn't access any references to external + * code or data. + * + * The only exception is the interrupt handlers in switcher.S: their + * addresses are placed in a table (default_idt_entries), so we need to + * update the table with the new addresses. switcher_offset() is a + * convenience function which returns the distance between the builtin + * switcher code and the high-mapped copy we just made. */ for (i = 0; i < IDT_ENTRIES; i++) default_idt_entries[i] += switcher_offset(); + /* + * Set up the Switcher's per-cpu areas. + * + * Each CPU gets two pages of its own within the high-mapped region + * (aka. "struct lguest_pages"). Much of this can be initialized now, + * but some depends on what Guest we are running (which is set up in + * copy_in_guest_info()). + */ for_each_possible_cpu(i) { + /* lguest_pages() returns this CPU's two pages. */ struct lguest_pages *pages = lguest_pages(i); + /* This is a convenience pointer to make the code fit one + * statement to a line. */ struct lguest_ro_state *state = &pages->state; - /* These fields are static: rest done in copy_in_guest_info */ + /* The Global Descriptor Table: the Host has a different one + * for each CPU. We keep a descriptor for the GDT which says + * where it is and how big it is (the size is actually the last + * byte, not the size, hence the "-1"). */ state->host_gdt_desc.size = GDT_SIZE-1; state->host_gdt_desc.address = (long)get_cpu_gdt_table(i); + + /* All CPUs on the Host use the same Interrupt Descriptor + * Table, so we just use store_idt(), which gets this CPU's IDT + * descriptor. */ store_idt(&state->host_idt_desc); + + /* The descriptors for the Guest's GDT and IDT can be filled + * out now, too. We copy the GDT & IDT into ->guest_gdt and + * ->guest_idt before actually running the Guest. */ state->guest_idt_desc.size = sizeof(state->guest_idt)-1; state->guest_idt_desc.address = (long)&state->guest_idt; state->guest_gdt_desc.size = sizeof(state->guest_gdt)-1; state->guest_gdt_desc.address = (long)&state->guest_gdt; + + /* We know where we want the stack to be when the Guest enters + * the switcher: in pages->regs. The stack grows upwards, so + * we start it at the end of that structure. */ state->guest_tss.esp0 = (long)(&pages->regs + 1); + /* And this is the GDT entry to use for the stack: we keep a + * couple of special LGUEST entries. */ state->guest_tss.ss0 = LGUEST_DS; - /* No I/O for you! */ + + /* x86 can have a finegrained bitmap which indicates what I/O + * ports the process can use. We set it to the end of our + * structure, meaning "none". */ state->guest_tss.io_bitmap_base = sizeof(state->guest_tss); + + /* Some GDT entries are the same across all Guests, so we can + * set them up now. */ setup_default_gdt_entries(state); + /* Most IDT entries are the same for all Guests, too.*/ setup_default_idt_entries(state, default_idt_entries); - /* Setup LGUEST segments on all cpus */ + /* The Host needs to be able to use the LGUEST segments on this + * CPU, too, so put them in the Host GDT. */ get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT; get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT; } - /* Initialize entry point into switcher. */ + /* In the Switcher, we want the %cs segment register to use the + * LGUEST_CS GDT entry: we've put that in the Host and Guest GDTs, so + * it will be undisturbed when we switch. To change %cs and jump we + * need this structure to feed to Intel's "lcall" instruction. */ lguest_entry.offset = (long)switch_to_guest + switcher_offset(); lguest_entry.segment = LGUEST_CS; printk(KERN_INFO "lguest: mapped switcher at %p\n", switcher_vma->addr); + /* And we succeeded... */ return 0; free_vma: @@ -146,35 +232,58 @@ free_some_pages: out: return err; } +/*:*/ +/* Cleaning up the mapping when the module is unloaded is almost... + * too easy. */ static void unmap_switcher(void) { unsigned int i; + /* vunmap() undoes *both* map_vm_area() and __get_vm_area(). */ vunmap(switcher_vma->addr); + /* Now we just need to free the pages we copied the switcher into */ for (i = 0; i < TOTAL_SWITCHER_PAGES; i++) __free_pages(switcher_page[i], 0); } -/* IN/OUT insns: enough to get us past boot-time probing. */ +/*H:130 Our Guest is usually so well behaved; it never tries to do things it + * isn't allowed to. Unfortunately, "struct paravirt_ops" isn't quite + * complete, because it doesn't contain replacements for the Intel I/O + * instructions. As a result, the Guest sometimes fumbles across one during + * the boot process as it probes for various things which are usually attached + * to a PC. + * + * When the Guest uses one of these instructions, we get trap #13 (General + * Protection Fault) and come here. We see if it's one of those troublesome + * instructions and skip over it. We return true if we did. */ static int emulate_insn(struct lguest *lg) { u8 insn; unsigned int insnlen = 0, in = 0, shift = 0; + /* The eip contains the *virtual* address of the Guest's instruction: + * guest_pa just subtracts the Guest's page_offset. */ unsigned long physaddr = guest_pa(lg, lg->regs->eip); - /* This only works for addresses in linear mapping... */ + /* The guest_pa() function only works for Guest kernel addresses, but + * that's all we're trying to do anyway. */ if (lg->regs->eip < lg->page_offset) return 0; + + /* Decoding x86 instructions is icky. */ lgread(lg, &insn, physaddr, 1); - /* Operand size prefix means it's actually for ax. */ + /* 0x66 is an "operand prefix". It means it's using the upper 16 bits + of the eax register. */ if (insn == 0x66) { shift = 16; + /* The instruction is 1 byte so far, read the next byte. */ insnlen = 1; lgread(lg, &insn, physaddr + insnlen, 1); } + /* We can ignore the lower bit for the moment and decode the 4 opcodes + * we need to emulate. */ switch (insn & 0xFE) { case 0xE4: /* in <next byte>,%al */ insnlen += 2; @@ -191,9 +300,13 @@ static int emulate_insn(struct lguest *lg) insnlen += 1; break; default: + /* OK, we don't know what this is, can't emulate. */ return 0; } + /* If it was an "IN" instruction, they expect the result to be read + * into %eax, so we change %eax. We always return all-ones, which + * traditionally means "there's nothing there". */ if (in) { /* Lower bit tells is whether it's a 16 or 32 bit access */ if (insn & 0x1) @@ -201,28 +314,46 @@ static int emulate_insn(struct lguest *lg) else lg->regs->eax |= (0xFFFF << shift); } + /* Finally, we've "done" the instruction, so move past it. */ lg->regs->eip += insnlen; + /* Success! */ return 1; } - +/*:*/ + +/*L:305 + * Dealing With Guest Memory. + * + * When the Guest gives us (what it thinks is) a physical address, we can use + * the normal copy_from_user() & copy_to_user() on that address: remember, + * Guest physical == Launcher virtual. + * + * But we can't trust the Guest: it might be trying to access the Launcher + * code. We have to check that the range is below the pfn_limit the Launcher + * gave us. We have to make sure that addr + len doesn't give us a false + * positive by overflowing, too. */ int lguest_address_ok(const struct lguest *lg, unsigned long addr, unsigned long len) { return (addr+len) / PAGE_SIZE < lg->pfn_limit && (addr+len >= addr); } -/* Just like get_user, but don't let guest access lguest binary. */ +/* This is a convenient routine to get a 32-bit value from the Guest (a very + * common operation). Here we can see how useful the kill_lguest() routine we + * met in the Launcher can be: we return a random value (0) instead of needing + * to return an error. */ u32 lgread_u32(struct lguest *lg, unsigned long addr) { u32 val = 0; - /* Don't let them access lguest binary */ + /* Don't let them access lguest binary. */ if (!lguest_address_ok(lg, addr, sizeof(val)) || get_user(val, (u32 __user *)addr) != 0) kill_guest(lg, "bad read address %#lx", addr); return val; } +/* Same thing for writing a value. */ void lgwrite_u32(struct lguest *lg, unsigned long addr, u32 val) { if (!lguest_address_ok(lg, addr, sizeof(val)) @@ -230,6 +361,9 @@ void lgwrite_u32(struct lguest *lg, unsigned long addr, u32 val) kill_guest(lg, "bad write address %#lx", addr); } +/* This routine is more generic, and copies a range of Guest bytes into a + * buffer. If the copy_from_user() fails, we fill the buffer with zeroes, so + * the caller doesn't end up using uninitialized kernel memory. */ void lgread(struct lguest *lg, void *b, unsigned long addr, unsigned bytes) { if (!lguest_address_ok(lg, addr, bytes) @@ -240,6 +374,7 @@ void lgread(struct lguest *lg, void *b, unsigned long addr, unsigned bytes) } } +/* Similarly, our generic routine to copy into a range of Guest bytes. */ void lgwrite(struct lguest *lg, unsigned long addr, const void *b, unsigned bytes) { @@ -247,6 +382,7 @@ void lgwrite(struct lguest *lg, unsigned long addr, const void *b, || copy_to_user((void __user *)addr, b, bytes) != 0) kill_guest(lg, "bad write address %#lx len %u", addr, bytes); } +/* (end of memory access helper routines) :*/ static void set_ts(void) { @@ -257,54 +393,108 @@ static void set_ts(void) write_cr0(cr0|8); } +/*S:010 + * We are getting close to the Switcher. + * + * Remember that each CPU has two pages which are visible to the Guest when it + * runs on that CPU. This has to contain the state for that Guest: we copy the + * state in just before we run the Guest. + * + * Each Guest has "changed" flags which indicate what has changed in the Guest + * since it last ran. We saw this set in interrupts_and_traps.c and + * segments.c. + */ static void copy_in_guest_info(struct lguest *lg, struct lguest_pages *pages) { + /* Copying all this data can be quite expensive. We usually run the + * same Guest we ran last time (and that Guest hasn't run anywhere else + * meanwhile). If that's not the case, we pretend everything in the + * Guest has changed. */ if (__get_cpu_var(last_guest) != lg || lg->last_pages != pages) { __get_cpu_var(last_guest) = lg; lg->last_pages = pages; lg->changed = CHANGED_ALL; } - /* These are pretty cheap, so we do them unconditionally. */ + /* These copies are pretty cheap, so we do them unconditionally: */ + /* Save the current Host top-level page directory. */ pages->state.host_cr3 = __pa(current->mm->pgd); + /* Set up the Guest's page tables to see this CPU's pages (and no + * other CPU's pages). */ map_switcher_in_guest(lg, pages); + /* Set up the two "TSS" members which tell the CPU what stack to use + * for traps which do directly into the Guest (ie. traps at privilege + * level 1). */ pages->state.guest_tss.esp1 = lg->esp1; pages->state.guest_tss.ss1 = lg->ss1; - /* Copy direct trap entries. */ + /* Copy direct-to-Guest trap entries. */ if (lg->changed & CHANGED_IDT) copy_traps(lg, pages->state.guest_idt, default_idt_entries); - /* Copy all GDT entries but the TSS. */ + /* Copy all GDT entries which the Guest can change. */ if (lg->changed & CHANGED_GDT) copy_gdt(lg, pages->state.guest_gdt); /* If only the TLS entries have changed, copy them. */ else if (lg->changed & CHANGED_GDT_TLS) copy_gdt_tls(lg, pages->state.guest_gdt); + /* Mark the Guest as unchanged for next time. */ lg->changed = 0; } +/* Finally: the code to actually call into the Switcher to run the Guest. */ static void run_guest_once(struct lguest *lg, struct lguest_pages *pages) { + /* This is a dummy value we need for GCC's sake. */ unsigned int clobber; + /* Copy the guest-specific information into this CPU's "struct + * lguest_pages". */ copy_in_guest_info(lg, pages); - /* Put eflags on stack, lcall does rest: suitable for iret return. */ + /* Now: we push the "eflags" register on the stack, then do an "lcall". + * This is how we change from using the kernel code segment to using + * the dedicated lguest code segment, as well as jumping into the + * Switcher. + * + * The lcall also pushes the old code segment (KERNEL_CS) onto the + * stack, then the address of this call. This stack layout happens to + * exactly match the stack of an interrupt... */ asm volatile("pushf; lcall *lguest_entry" + /* This is how we tell GCC that %eax ("a") and %ebx ("b") + * are changed by this routine. The "=" means output. */ : "=a"(clobber), "=b"(clobber) + /* %eax contains the pages pointer. ("0" refers to the + * 0-th argument above, ie "a"). %ebx contains the + * physical address of the Guest's top-level page + * directory. */ : "0"(pages), "1"(__pa(lg->pgdirs[lg->pgdidx].pgdir)) + /* We tell gcc that all these registers could change, + * which means we don't have to save and restore them in + * the Switcher. */ : "memory", "%edx", "%ecx", "%edi", "%esi"); } +/*:*/ +/*H:030 Let's jump straight to the the main loop which runs the Guest. + * Remember, this is called by the Launcher reading /dev/lguest, and we keep + * going around and around until something interesting happens. */ int run_guest(struct lguest *lg, unsigned long __user *user) { + /* We stop running once the Guest is dead. */ while (!lg->dead) { + /* We need to initialize this, otherwise gcc complains. It's + * not (yet) clever enough to see that it's initialized when we + * need it. */ unsigned int cr2 = 0; /* Damn gcc */ - /* Hypercalls first: we might have been out to userspace */ + /* First we run any hypercalls the Guest wants done: either in + * the hypercall ring in "struct lguest_data", or directly by + * using int 31 (LGUEST_TRAP_ENTRY). */ do_hypercalls(lg); + /* It's possible the Guest did a SEND_DMA hypercall to the + * Launcher, in which case we return from the read() now. */ if (lg->dma_is_pending) { if (put_user(lg->pending_dma, user) || put_user(lg->pending_key, user+1)) @@ -312,6 +502,7 @@ int run_guest(struct lguest *lg, unsigned long __user *user) return sizeof(unsigned long)*2; } + /* Check for signals */ if (signal_pending(current)) return -ERESTARTSYS; @@ -319,77 +510,154 @@ int run_guest(struct lguest *lg, unsigned long __user *user) if (lg->break_out) return -EAGAIN; + /* Check if there are any interrupts which can be delivered + * now: if so, this sets up the hander to be executed when we + * next run the Guest. */ maybe_do_interrupt(lg); + /* All long-lived kernel loops need to check with this horrible + * thing called the freezer. If the Host is trying to suspend, + * it stops us. */ try_to_freeze(); + /* Just make absolutely sure the Guest is still alive. One of + * those hypercalls could have been fatal, for example. */ if (lg->dead) break; + /* If the Guest asked to be stopped, we sleep. The Guest's + * clock timer or LHCALL_BREAK from the Waker will wake us. */ if (lg->halted) { set_current_state(TASK_INTERRUPTIBLE); schedule(); continue; } + /* OK, now we're ready to jump into the Guest. First we put up + * the "Do Not Disturb" sign: */ local_irq_disable(); - /* Even if *we* don't want FPU trap, guest might... */ + /* Remember the awfully-named TS bit? If the Guest has asked + * to set it we set it now, so we can trap and pass that trap + * to the Guest if it uses the FPU. */ if (lg->ts) set_ts(); - /* Don't let Guest do SYSENTER: we can't handle it. */ + /* SYSENTER is an optimized way of doing system calls. We + * can't allow it because it always jumps to privilege level 0. + * A normal Guest won't try it because we don't advertise it in + * CPUID, but a malicious Guest (or malicious Guest userspace + * program) could, so we tell the CPU to disable it before + * running the Guest. */ if (boot_cpu_has(X86_FEATURE_SEP)) wrmsr(MSR_IA32_SYSENTER_CS, 0, 0); + /* Now we actually run the Guest. It will pop back out when + * something interesting happens, and we can examine its + * registers to see what it was doing. */ run_guest_once(lg, lguest_pages(raw_smp_processor_id())); - /* Save cr2 now if we page-faulted. */ + /* The "regs" pointer contains two extra entries which are not + * really registers: a trap number which says what interrupt or + * trap made the switcher code come back, and an error code + * which some traps set. */ + + /* If the Guest page faulted, then the cr2 register will tell + * us the bad virtual address. We have to grab this now, + * because once we re-enable interrupts an interrupt could + * fault and thus overwrite cr2, or we could even move off to a + * different CPU. */ if (lg->regs->trapnum == 14) cr2 = read_cr2(); + /* Similarly, if we took a trap because the Guest used the FPU, + * we have to restore the FPU it expects to see. */ else if (lg->regs->trapnum == 7) math_state_restore(); + /* Restore SYSENTER if it's supposed to be on. */ if (boot_cpu_has(X86_FEATURE_SEP)) wrmsr(MSR_IA32_SYSENTER_CS, __KERNEL_CS, 0); + + /* Now we're ready to be interrupted or moved to other CPUs */ local_irq_enable(); + /* OK, so what happened? */ switch (lg->regs->trapnum) { case 13: /* We've intercepted a GPF. */ + /* Check if this was one of those annoying IN or OUT + * instructions which we need to emulate. If so, we + * just go back into the Guest after we've done it. */ if (lg->regs->errcode == 0) { if (emulate_insn(lg)) continue; } break; case 14: /* We've intercepted a page fault. */ + /* The Guest accessed a virtual address that wasn't + * mapped. This happens a lot: we don't actually set + * up most of the page tables for the Guest at all when + * we start: as it runs it asks for more and more, and + * we set them up as required. In this case, we don't + * even tell the Guest that the fault happened. + * + * The errcode tells whether this was a read or a + * write, and whether kernel or userspace code. */ if (demand_page(lg, cr2, lg->regs->errcode)) continue; - /* If lguest_data is NULL, this won't hurt. */ + /* OK, it's really not there (or not OK): the Guest + * needs to know. We write out the cr2 value so it + * knows where the fault occurred. + * + * Note that if the Guest were really messed up, this + * could happen before it's done the INITIALIZE + * hypercall, so lg->lguest_data will be NULL, so + * &lg->lguest_data->cr2 will be address 8. Writing + * into that address won't hurt the Host at all, + * though. */ if (put_user(cr2, &lg->lguest_data->cr2)) kill_guest(lg, "Writing cr2"); break; case 7: /* We've intercepted a Device Not Available fault. */ - /* If they don't want to know, just absorb it. */ + /* If the Guest doesn't want to know, we already + * restored the Floating Point Unit, so we just + * continue without telling it. */ if (!lg->ts) continue; break; - case 32 ... 255: /* Real interrupt, fall thru */ + case 32 ... 255: + /* These values mean a real interrupt occurred, in + * which case the Host handler has already been run. + * We just do a friendly check if another process + * should now be run, then fall through to loop + * around: */ cond_resched(); case LGUEST_TRAP_ENTRY: /* Handled at top of loop */ continue; } + /* If we get here, it's a trap the Guest wants to know + * about. */ if (deliver_trap(lg, lg->regs->trapnum)) continue; + /* If the Guest doesn't have a handler (either it hasn't + * registered any yet, or it's one of the faults we don't let + * it handle), it dies with a cryptic error message. */ kill_guest(lg, "unhandled trap %li at %#lx (%#lx)", lg->regs->trapnum, lg->regs->eip, lg->regs->trapnum == 14 ? cr2 : lg->regs->errcode); } + /* The Guest is dead => "No such file or directory" */ return -ENOENT; } +/* Now we can look at each of the routines this calls, in increasing order of + * complexity: do_hypercalls(), emulate_insn(), maybe_do_interrupt(), + * deliver_trap() and demand_page(). After all those, we'll be ready to + * examine the Switcher, and our philosophical understanding of the Host/Guest + * duality will be complete. :*/ + int find_free_guest(void) { unsigned int i; @@ -407,55 +675,96 @@ static void adjust_pge(void *on) write_cr4(read_cr4() & ~X86_CR4_PGE); } +/*H:000 + * Welcome to the Host! + * + * By this point your brain has been tickled by the Guest code and numbed by + * the Launcher code; prepare for it to be stretched by the Host code. This is + * the heart. Let's begin at the initialization routine for the Host's lg + * module. + */ static int __init init(void) { int err; + /* Lguest can't run under Xen, VMI or itself. It does Tricky Stuff. */ if (paravirt_enabled()) { printk("lguest is afraid of %s\n", paravirt_ops.name); return -EPERM; } + /* First we put the Switcher up in very high virtual memory. */ err = map_switcher(); if (err) return err; + /* Now we set up the pagetable implementation for the Guests. */ err = init_pagetables(switcher_page, SHARED_SWITCHER_PAGES); if (err) { unmap_switcher(); return err; } + + /* The I/O subsystem needs some things initialized. */ lguest_io_init(); + /* /dev/lguest needs to be registered. */ err = lguest_device_init(); if (err) { free_pagetables(); unmap_switcher(); return err; } + + /* Finally, we need to turn off "Page Global Enable". PGE is an + * optimization where page table entries are specially marked to show + * they never change. The Host kernel marks all the kernel pages this + * way because it's always present, even when userspace is running. + * + * Lguest breaks this: unbeknownst to the rest of the Host kernel, we + * switch to the Guest kernel. If you don't disable this on all CPUs, + * you'll get really weird bugs that you'll chase for two days. + * + * I used to turn PGE off every time we switched to the Guest and back + * on when we return, but that slowed the Switcher down noticibly. */ + + /* We don't need the complexity of CPUs coming and going while we're + * doing this. */ lock_cpu_hotplug(); if (cpu_has_pge) { /* We have a broader idea of "global". */ + /* Remember that this was originally set (for cleanup). */ cpu_had_pge = 1; + /* adjust_pge is a helper function which sets or unsets the PGE + * bit on its CPU, depending on the argument (0 == unset). */ on_each_cpu(adjust_pge, (void *)0, 0, 1); + /* Turn off the feature in the global feature set. */ clear_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); } unlock_cpu_hotplug(); + + /* All good! */ return 0; } +/* Cleaning up is just the same code, backwards. With a little French. */ static void __exit fini(void) { lguest_device_remove(); free_pagetables(); unmap_switcher(); + + /* If we had PGE before we started, turn it back on now. */ lock_cpu_hotplug(); if (cpu_had_pge) { set_bit(X86_FEATURE_PGE, boot_cpu_data.x86_capability); + /* adjust_pge's argument "1" means set PGE. */ on_each_cpu(adjust_pge, (void *)1, 0, 1); } unlock_cpu_hotplug(); } +/* The Host side of lguest can be a module. This is a nice way for people to + * play with it. */ module_init(init); module_exit(fini); MODULE_LICENSE("GPL"); diff --git a/drivers/lguest/hypercalls.c b/drivers/lguest/hypercalls.c index ea52ca4..db6caac 100644 --- a/drivers/lguest/hypercalls.c +++ b/drivers/lguest/hypercalls.c @@ -1,5 +1,10 @@ -/* Actual hypercalls, which allow guests to actually do something. - Copyright (C) 2006 Rusty Russell IBM Corporation +/*P:500 Just as userspace programs request kernel operations through a system + * call, the Guest requests Host operations through a "hypercall". You might + * notice this nomenclature doesn't really follow any logic, but the name has + * been around for long enough that we're stuck with it. As you'd expect, this + * code is basically a one big switch statement. :*/ + +/* Copyright (C) 2006 Rusty Russell IBM Corporation This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -23,37 +28,55 @@ #include <irq_vectors.h> #include "lg.h" +/*H:120 This is the core hypercall routine: where the Guest gets what it + * wants. Or gets killed. Or, in the case of LHCALL_CRASH, both. + * + * Remember from the Guest: %eax == which call to make, and the arguments are + * packed into %edx, %ebx and %ecx if needed. */ static void do_hcall(struct lguest *lg, struct lguest_regs *regs) { switch (regs->eax) { case LHCALL_FLUSH_ASYNC: + /* This call does nothing, except by breaking out of the Guest + * it makes us process all the asynchronous hypercalls. */ break; case LHCALL_LGUEST_INIT: + /* You can't get here unless you're already initialized. Don't + * do that. */ kill_guest(lg, "already have lguest_data"); break; case LHCALL_CRASH: { + /* Crash is such a trivial hypercall that we do it in four + * lines right here. */ char msg[128]; + /* If the lgread fails, it will call kill_guest() itself; the + * kill_guest() with the message will be ignored. */ lgread(lg, msg, regs->edx, sizeof(msg)); msg[sizeof(msg)-1] = '\0'; kill_guest(lg, "CRASH: %s", msg); break; } case LHCALL_FLUSH_TLB: + /* FLUSH_TLB comes in two flavors, depending on the + * argument: */ if (regs->edx) guest_pagetable_clear_all(lg); else guest_pagetable_flush_user(lg); break; - case LHCALL_GET_WALLCLOCK: { - struct timespec ts; - ktime_get_real_ts(&ts); - regs->eax = ts.tv_sec; - break; - } case LHCALL_BIND_DMA: + /* BIND_DMA really wants four arguments, but it's the only call + * which does. So the Guest packs the number of buffers and + * the interrupt number into the final argument, and we decode + * it here. This can legitimately fail, since we currently + * place a limit on the number of DMA pools a Guest can have. + * So we return true or false from this call. */ regs->eax = bind_dma(lg, regs->edx, regs->ebx, regs->ecx >> 8, regs->ecx & 0xFF); break; + + /* All these calls simply pass the arguments through to the right + * routines. */ case LHCALL_SEND_DMA: send_dma(lg, regs->edx, regs->ebx); break; @@ -81,10 +104,13 @@ static void do_hcall(struct lguest *lg, struct lguest_regs *regs) case LHCALL_SET_CLOCKEVENT: guest_set_clockevent(lg, regs->edx); break; + case LHCALL_TS: + /* This sets the TS flag, as we saw used in run_guest(). */ lg->ts = regs->edx; break; case LHCALL_HALT: + /* Similarly, this sets the halted flag for run_guest(). */ lg->halted = 1; break; default: @@ -92,25 +118,42 @@ static void do_hcall(struct lguest *lg, struct lguest_regs *regs) } } -/* We always do queued calls before actual hypercall. */ +/* Asynchronous hypercalls are easy: we just look in the array in the Guest's + * "struct lguest_data" and see if there are any new ones marked "ready". + * + * We are careful to do these in order: obviously we respect the order the + * Guest put them in the ring, but we also promise the Guest that they will + * happen before any normal hypercall (which is why we check this before + * checking for a normal hcall). */ static void do_async_hcalls(struct lguest *lg) { unsigned int i; u8 st[LHCALL_RING_SIZE]; + /* For simplicity, we copy the entire call status array in at once. */ if (copy_from_user(&st, &lg->lguest_data->hcall_status, sizeof(st))) return; + + /* We process "struct lguest_data"s hcalls[] ring once. */ for (i = 0; i < ARRAY_SIZE(st); i++) { struct lguest_regs regs; + /* We remember where we were up to from last time. This makes + * sure that the hypercalls are done in the order the Guest + * places them in the ring. */ unsigned int n = lg->next_hcall; + /* 0xFF means there's no call here (yet). */ if (st[n] == 0xFF) break; + /* OK, we have hypercall. Increment the "next_hcall" cursor, + * and wrap back to 0 if we reach the end. */ if (++lg->next_hcall == LHCALL_RING_SIZE) lg->next_hcall = 0; + /* We copy the hypercall arguments into a fake register + * structure. This makes life simple for do_hcall(). */ if (get_user(regs.eax, &lg->lguest_data->hcalls[n].eax) || get_user(regs.edx, &lg->lguest_data->hcalls[n].edx) || get_user(regs.ecx, &lg->lguest_data->hcalls[n].ecx) @@ -119,74 +162,139 @@ static void do_async_hcalls(struct lguest *lg) break; } + /* Do the hypercall, same as a normal one. */ do_hcall(lg, ®s); + + /* Mark the hypercall done. */ if (put_user(0xFF, &lg->lguest_data->hcall_status[n])) { kill_guest(lg, "Writing result for async hypercall"); break; } + /* Stop doing hypercalls if we've just done a DMA to the + * Launcher: it needs to service this first. */ if (lg->dma_is_pending) break; } } +/* Last of all, we look at what happens first of all. The very first time the + * Guest makes a hypercall, we end up here to set things up: */ static void initialize(struct lguest *lg) { u32 tsc_speed; + /* You can't do anything until you're initialized. The Guest knows the + * rules, so we're unforgiving here. */ if (lg->regs->eax != LHCALL_LGUEST_INIT) { kill_guest(lg, "hypercall %li before LGUEST_INIT", lg->regs->eax); return; } - /* We only tell the guest to use the TSC if it's reliable. */ + /* We insist that the Time Stamp Counter exist and doesn't change with + * cpu frequency. Some devious chip manufacturers decided that TSC + * changes could be handled in software. I decided that time going + * backwards might be good for benchmarks, but it's bad for users. + * + * We also insist that the TSC be stable: the kernel detects unreliable + * TSCs for its own purposes, and we use that here. */ if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) && !check_tsc_unstable()) tsc_speed = tsc_khz; else tsc_speed = 0; + /* The pointer to the Guest's "struct lguest_data" is the only + * argument. */ lg->lguest_data = (struct lguest_data __user *)lg->regs->edx; - /* We check here so we can simply copy_to_user/from_user */ + /* If we check the address they gave is OK now, we can simply + * copy_to_user/from_user from now on rather than using lgread/lgwrite. + * I put this in to show that I'm not immune to writing stupid + * optimizations. */ if (!lguest_address_ok(lg, lg->regs->edx, sizeof(*lg->lguest_data))) { kill_guest(lg, "bad guest page %p", lg->lguest_data); return; } + /* The Guest tells us where we're not to deliver interrupts by putting + * the range of addresses into "struct lguest_data". */ if (get_user(lg->noirq_start, &lg->lguest_data->noirq_start) || get_user(lg->noirq_end, &lg->lguest_data->noirq_end) - /* We reserve the top pgd entry. */ + /* We tell the Guest that it can't use the top 4MB of virtual + * addresses used by the Switcher. */ || put_user(4U*1024*1024, &lg->lguest_data->reserve_mem) || put_user(tsc_speed, &lg->lguest_data->tsc_khz) + /* We also give the Guest a unique id, as used in lguest_net.c. */ || put_user(lg->guestid, &lg->lguest_data->guestid)) kill_guest(lg, "bad guest page %p", lg->lguest_data); - /* This is the one case where the above accesses might have - * been the first write to a Guest page. This may have caused - * a copy-on-write fault, but the Guest might be referring to - * the old (read-only) page. */ + /* We write the current time into the Guest's data page once now. */ + write_timestamp(lg); + + /* This is the one case where the above accesses might have been the + * first write to a Guest page. This may have caused a copy-on-write + * fault, but the Guest might be referring to the old (read-only) + * page. */ guest_pagetable_clear_all(lg); } +/* Now we've examined the hypercall code; our Guest can make requests. There + * is one other way we can do things for the Guest, as we see in + * emulate_insn(). */ -/* Even if we go out to userspace and come back, we don't want to do - * the hypercall again. */ +/*H:110 Tricky point: we mark the hypercall as "done" once we've done it. + * Normally we don't need to do this: the Guest will run again and update the + * trap number before we come back around the run_guest() loop to + * do_hypercalls(). + * + * However, if we are signalled or the Guest sends DMA to the Launcher, that + * loop will exit without running the Guest. When it comes back it would try + * to re-run the hypercall. */ static void clear_hcall(struct lguest *lg) { lg->regs->trapnum = 255; } +/*H:100 + * Hypercalls + * + * Remember from the Guest, hypercalls come in two flavors: normal and + * asynchronous. This file handles both of types. + */ void do_hypercalls(struct lguest *lg) { + /* Not initialized yet? */ if (unlikely(!lg->lguest_data)) { + /* Did the Guest make a hypercall? We might have come back for + * some other reason (an interrupt, a different trap). */ if (lg->regs->trapnum == LGUEST_TRAP_ENTRY) { + /* Set up the "struct lguest_data" */ initialize(lg); + /* The hypercall is done. */ clear_hcall(lg); } return; } + /* The Guest has initialized. + * + * Look in the hypercall ring for the async hypercalls: */ do_async_hcalls(lg); + + /* If we stopped reading the hypercall ring because the Guest did a + * SEND_DMA to the Launcher, we want to return now. Otherwise if the + * Guest asked us to do a hypercall, we do it. */ if (!lg->dma_is_pending && lg->regs->trapnum == LGUEST_TRAP_ENTRY) { do_hcall(lg, lg->regs); + /* The hypercall is done. */ clear_hcall(lg); } } + +/* This routine supplies the Guest with time: it's used for wallclock time at + * initial boot and as a rough time source if the TSC isn't available. */ +void write_timestamp(struct lguest *lg) +{ + struct timespec now; + ktime_get_real_ts(&now); + if (put_user(now, &lg->lguest_data->time)) + kill_guest(lg, "Writing timestamp"); +} diff --git a/drivers/lguest/interrupts_and_traps.c b/drivers/lguest/interrupts_and_traps.c index bee029b..49787e9 100644 --- a/drivers/lguest/interrupts_and_traps.c +++ b/drivers/lguest/interrupts_and_traps.c @@ -1,100 +1,160 @@ +/*P:800 Interrupts (traps) are complicated enough to earn their own file. + * There are three classes of interrupts: + * + * 1) Real hardware interrupts which occur while we're running the Guest, + * 2) Interrupts for virtual devices attached to the Guest, and + * 3) Traps and faults from the Guest. + * + * Real hardware interrupts must be delivered to the Host, not the Guest. + * Virtual interrupts must be delivered to the Guest, but we make them look + * just like real hardware would deliver them. Traps from the Guest can be set + * up to go directly back into the Guest, but sometimes the Host wants to see + * them first, so we also have a way of "reflecting" them into the Guest as if + * they had been delivered to it directly. :*/ #include <linux/uaccess.h> #include "lg.h" +/* The address of the interrupt handler is split into two bits: */ static unsigned long idt_address(u32 lo, u32 hi) { return (lo & 0x0000FFFF) | (hi & 0xFFFF0000); } +/* The "type" of the interrupt handler is a 4 bit field: we only support a + * couple of types. */ static int idt_type(u32 lo, u32 hi) { return (hi >> 8) & 0xF; } +/* An IDT entry can't be used unless the "present" bit is set. */ static int idt_present(u32 lo, u32 hi) { return (hi & 0x8000); } +/* We need a helper to "push" a value onto the Guest's stack, since that's a + * big part of what delivering an interrupt does. */ static void push_guest_stack(struct lguest *lg, unsigned long *gstack, u32 val) { + /* Stack grows upwards: move stack then write value. */ *gstack -= 4; lgwrite_u32(lg, *gstack, val); } +/*H:210 The set_guest_interrupt() routine actually delivers the interrupt or + * trap. The mechanics of delivering traps and interrupts to the Guest are the + * same, except some traps have an "error code" which gets pushed onto the + * stack as well: the caller tells us if this is one. + * + * "lo" and "hi" are the two parts of the Interrupt Descriptor Table for this + * interrupt or trap. It's split into two parts for traditional reasons: gcc + * on i386 used to be frightened by 64 bit numbers. + * + * We set up the stack just like the CPU does for a real interrupt, so it's + * identical for the Guest (and the standard "iret" instruction will undo + * it). */ static void set_guest_interrupt(struct lguest *lg, u32 lo, u32 hi, int has_err) { unsigned long gstack; u32 eflags, ss, irq_enable; - /* If they want a ring change, we use new stack and push old ss/esp */ + /* There are two cases for interrupts: one where the Guest is already + * in the kernel, and a more complex one where the Guest is in + * userspace. We check the privilege level to find out. */ if ((lg->regs->ss&0x3) != GUEST_PL) { + /* The Guest told us their kernel stack with the SET_STACK + * hypercall: both the virtual address and the segment */ gstack = guest_pa(lg, lg->esp1); ss = lg->ss1; + /* We push the old stack segment and pointer onto the new + * stack: when the Guest does an "iret" back from the interrupt + * handler the CPU will notice they're dropping privilege + * levels and expect these here. */ push_guest_stack(lg, &gstack, lg->regs->ss); push_guest_stack(lg, &gstack, lg->regs->esp); } else { + /* We're staying on the same Guest (kernel) stack. */ gstack = guest_pa(lg, lg->regs->esp); ss = lg->regs->ss; } - /* We use IF bit in eflags to indicate whether irqs were enabled - (it's always 1, since irqs are enabled when guest is running). */ + /* Remember that we never let the Guest actually disable interrupts, so + * the "Interrupt Flag" bit is always set. We copy that bit from the + * Guest's "irq_enabled" field into the eflags word: the Guest copies + * it back in "lguest_iret". */ eflags = lg->regs->eflags; if (get_user(irq_enable, &lg->lguest_data->irq_enabled) == 0 && !(irq_enable & X86_EFLAGS_IF)) eflags &= ~X86_EFLAGS_IF; + /* An interrupt is expected to push three things on the stack: the old + * "eflags" word, the old code segment, and the old instruction + * pointer. */ push_guest_stack(lg, &gstack, eflags); push_guest_stack(lg, &gstack, lg->regs->cs); push_guest_stack(lg, &gstack, lg->regs->eip); + /* For the six traps which supply an error code, we push that, too. */ if (has_err) push_guest_stack(lg, &gstack, lg->regs->errcode); - /* Change the real stack so switcher returns to trap handler */ + /* Now we've pushed all the old state, we change the stack, the code + * segment and the address to execute. */ lg->regs->ss = ss; lg->regs->esp = gstack + lg->page_offset; lg->regs->cs = (__KERNEL_CS|GUEST_PL); lg->regs->eip = idt_address(lo, hi); - /* Disable interrupts for an interrupt gate. */ + /* There are two kinds of interrupt handlers: 0xE is an "interrupt + * gate" which expects interrupts to be disabled on entry. */ if (idt_type(lo, hi) == 0xE) if (put_user(0, &lg->lguest_data->irq_enabled)) kill_guest(lg, "Disabling interrupts"); } +/*H:200 + * Virtual Interrupts. + * + * maybe_do_interrupt() gets called before every entry to the Guest, to see if + * we should divert the Guest to running an interrupt handler. */ void maybe_do_interrupt(struct lguest *lg) { unsigned int irq; DECLARE_BITMAP(blk, LGUEST_IRQS); struct desc_struct *idt; + /* If the Guest hasn't even initialized yet, we can do nothing. */ if (!lg->lguest_data) return; - /* Mask out any interrupts they have blocked. */ + /* Take our "irqs_pending" array and remove any interrupts the Guest + * wants blocked: the result ends up in "blk". */ if (copy_from_user(&blk, lg->lguest_data->blocked_interrupts, sizeof(blk))) return; bitmap_andnot(blk, lg->irqs_pending, blk, LGUEST_IRQS); + /* Find the first interrupt. */ irq = find_first_bit(blk, LGUEST_IRQS); + /* None? Nothing to do */ if (irq >= LGUEST_IRQS) return; + /* They may be in the middle of an iret, where they asked us never to + * deliver interrupts. */ if (lg->regs->eip >= lg->noirq_start && lg->regs->eip < lg->noirq_end) return; - /* If they're halted, we re-enable interrupts. */ + /* If they're halted, interrupts restart them. */ if (lg->halted) { /* Re-enable interrupts. */ if (put_user(X86_EFLAGS_IF, &lg->lguest_data->irq_enabled)) kill_guest(lg, "Re-enabling interrupts"); lg->halted = 0; } else { - /* Maybe they have interrupts disabled? */ + /* Otherwise we check if they have interrupts disabled. */ u32 irq_enabled; if (get_user(irq_enabled, &lg->lguest_data->irq_enabled)) irq_enabled = 0; @@ -102,112 +162,218 @@ void maybe_do_interrupt(struct lguest *lg) return; } + /* Look at the IDT entry the Guest gave us for this interrupt. The + * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip + * over them. */ idt = &lg->idt[FIRST_EXTERNAL_VECTOR+irq]; + /* If they don't have a handler (yet?), we just ignore it */ if (idt_present(idt->a, idt->b)) { + /* OK, mark it no longer pending and deliver it. */ clear_bit(irq, lg->irqs_pending); + /* set_guest_interrupt() takes the interrupt descriptor and a + * flag to say whether this interrupt pushes an error code onto + * the stack as well: virtual interrupts never do. */ set_guest_interrupt(lg, idt->a, idt->b, 0); } + + /* Every time we deliver an interrupt, we update the timestamp in the + * Guest's lguest_data struct. It would be better for the Guest if we + * did this more often, but it can actually be quite slow: doing it + * here is a compromise which means at least it gets updated every + * timer interrupt. */ + write_timestamp(lg); } +/*H:220 Now we've got the routines to deliver interrupts, delivering traps + * like page fault is easy. The only trick is that Intel decided that some + * traps should have error codes: */ static int has_err(unsigned int trap) { return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17); } +/* deliver_trap() returns true if it could deliver the trap. */ int deliver_trap(struct lguest *lg, unsigned int num) { u32 lo = lg->idt[num].a, hi = lg->idt[num].b; + /* Early on the Guest hasn't set the IDT entries (or maybe it put a + * bogus one in): if we fail here, the Guest will be killed. */ if (!idt_present(lo, hi)) return 0; set_guest_interrupt(lg, lo, hi, has_err(num)); return 1; } +/*H:250 Here's the hard part: returning to the Host every time a trap happens + * and then calling deliver_trap() and re-entering the Guest is slow. + * Particularly because Guest userspace system calls are traps (trap 128). + * + * So we'd like to set up the IDT to tell the CPU to deliver traps directly + * into the Guest. This is possible, but the complexities cause the size of + * this file to double! However, 150 lines of code is worth writing for taking + * system calls down from 1750ns to 270ns. Plus, if lguest didn't do it, all + * the other hypervisors would tease it. + * + * This routine determines if a trap can be delivered directly. */ static int direct_trap(const struct lguest *lg, const struct desc_struct *trap, unsigned int num) { - /* Hardware interrupts don't go to guest (except syscall). */ + /* Hardware interrupts don't go to the Guest at all (except system + * call). */ if (num >= FIRST_EXTERNAL_VECTOR && num != SYSCALL_VECTOR) return 0; - /* We intercept page fault (demand shadow paging & cr2 saving) - protection fault (in/out emulation) and device not - available (TS handling), and hypercall */ + /* The Host needs to see page faults (for shadow paging and to save the + * fault address), general protection faults (in/out emulation) and + * device not available (TS handling), and of course, the hypercall + * trap. */ if (num == 14 || num == 13 || num == 7 || num == LGUEST_TRAP_ENTRY) return 0; - /* Interrupt gates (0xE) or not present (0x0) can't go direct. */ + /* Only trap gates (type 15) can go direct to the Guest. Interrupt + * gates (type 14) disable interrupts as they are entered, which we + * never let the Guest do. Not present entries (type 0x0) also can't + * go direct, of course 8) */ return idt_type(trap->a, trap->b) == 0xF; } - +/*:*/ + +/*M:005 The Guest has the ability to turn its interrupt gates into trap gates, + * if it is careful. The Host will let trap gates can go directly to the + * Guest, but the Guest needs the interrupts atomically disabled for an + * interrupt gate. It can do this by pointing the trap gate at instructions + * within noirq_start and noirq_end, where it can safely disable interrupts. */ + +/*M:006 The Guests do not use the sysenter (fast system call) instruction, + * because it's hardcoded to enter privilege level 0 and so can't go direct. + * It's about twice as fast as the older "int 0x80" system call, so it might + * still be worthwhile to handle it in the Switcher and lcall down to the + * Guest. The sysenter semantics are hairy tho: search for that keyword in + * entry.S :*/ + +/*H:260 When we make traps go directly into the Guest, we need to make sure + * the kernel stack is valid (ie. mapped in the page tables). Otherwise, the + * CPU trying to deliver the trap will fault while trying to push the interrupt + * words on the stack: this is called a double fault, and it forces us to kill + * the Guest. + * + * Which is deeply unfair, because (literally!) it wasn't the Guests' fault. */ void pin_stack_pages(struct lguest *lg) { unsigned int i; + /* Depending on the CONFIG_4KSTACKS option, the Guest can have one or + * two pages of stack space. */ for (i = 0; i < lg->stack_pages; i++) + /* The stack grows *upwards*, hence the subtraction */ pin_page(lg, lg->esp1 - i * PAGE_SIZE); } +/* Direct traps also mean that we need to know whenever the Guest wants to use + * a different kernel stack, so we can change the IDT entries to use that + * stack. The IDT entries expect a virtual address, so unlike most addresses + * the Guest gives us, the "esp" (stack pointer) value here is virtual, not + * physical. + * + * In Linux each process has its own kernel stack, so this happens a lot: we + * change stacks on each context switch. */ void guest_set_stack(struct lguest *lg, u32 seg, u32 esp, unsigned int pages) { - /* You cannot have a stack segment with priv level 0. */ + /* You are not allowd have a stack segment with privilege level 0: bad + * Guest! */ if ((seg & 0x3) != GUEST_PL) kill_guest(lg, "bad stack segment %i", seg); + /* We only expect one or two stack pages. */ if (pages > 2) kill_guest(lg, "bad stack pages %u", pages); + /* Save where the stack is, and how many pages */ lg->ss1 = seg; lg->esp1 = esp; lg->stack_pages = pages; + /* Make sure the new stack pages are mapped */ pin_stack_pages(lg); } -/* Set up trap in IDT. */ +/* All this reference to mapping stacks leads us neatly into the other complex + * part of the Host: page table handling. */ + +/*H:235 This is the routine which actually checks the Guest's IDT entry and + * transfers it into our entry in "struct lguest": */ static void set_trap(struct lguest *lg, struct desc_struct *trap, unsigned int num, u32 lo, u32 hi) { u8 type = idt_type(lo, hi); + /* We zero-out a not-present entry */ if (!idt_present(lo, hi)) { trap->a = trap->b = 0; return; } + /* We only support interrupt and trap gates. */ if (type != 0xE && type != 0xF) kill_guest(lg, "bad IDT type %i", type); + /* We only copy the handler address, present bit, privilege level and + * type. The privilege level controls where the trap can be triggered + * manually with an "int" instruction. This is usually GUEST_PL, + * except for system calls which userspace can use. */ trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF); trap->b = (hi&0xFFFFEF00); } +/*H:230 While we're here, dealing with delivering traps and interrupts to the + * Guest, we might as well complete the picture: how the Guest tells us where + * it wants them to go. This would be simple, except making traps fast + * requires some tricks. + * + * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the + * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here. */ void load_guest_idt_entry(struct lguest *lg, unsigned int num, u32 lo, u32 hi) { - /* Guest never handles: NMI, doublefault, hypercall, spurious irq. */ + /* Guest never handles: NMI, doublefault, spurious interrupt or + * hypercall. We ignore when it tries to set them. */ if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY) return; + /* Mark the IDT as changed: next time the Guest runs we'll know we have + * to copy this again. */ lg->changed |= CHANGED_IDT; + + /* The IDT which we keep in "struct lguest" only contains 32 entries + * for the traps and LGUEST_IRQS (32) entries for interrupts. We + * ignore attempts to set handlers for higher interrupt numbers, except + * for the system call "interrupt" at 128: we have a special IDT entry + * for that. */ if (num < ARRAY_SIZE(lg->idt)) set_trap(lg, &lg->idt[num], num, lo, hi); else if (num == SYSCALL_VECTOR) set_trap(lg, &lg->syscall_idt, num, lo, hi); } +/* The default entry for each interrupt points into the Switcher routines which + * simply return to the Host. The run_guest() loop will then call + * deliver_trap() to bounce it back into the Guest. */ static void default_idt_entry(struct desc_struct *idt, int trap, const unsigned long handler) { + /* A present interrupt gate. */ u32 flags = 0x8e00; - /* They can't "int" into any of them except hypercall. */ + /* Set the privilege level on the entry for the hypercall: this allows + * the Guest to use the "int" instruction to trigger it. */ if (trap == LGUEST_TRAP_ENTRY) flags |= (GUEST_PL << 13); + /* Now pack it into the IDT entry in its weird format. */ idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF); idt->b = (handler&0xFFFF0000) | flags; } +/* When the Guest first starts, we put default entries into the IDT. */ void setup_default_idt_entries(struct lguest_ro_state *state, const unsigned long *def) { @@ -217,19 +383,25 @@ void setup_default_idt_entries(struct lguest_ro_state *state, default_idt_entry(&state->guest_idt[i], i, def[i]); } +/*H:240 We don't use the IDT entries in the "struct lguest" directly, instead + * we copy them into the IDT which we've set up for Guests on this CPU, just + * before we run the Guest. This routine does that copy. */ void copy_traps(const struct lguest *lg, struct desc_struct *idt, const unsigned long *def) { unsigned int i; - /* All hardware interrupts are same whatever the guest: only the - * traps might be different. */ + /* We can simply copy the direct traps, otherwise we use the default + * ones in the Switcher: they will return to the Host. */ for (i = 0; i < FIRST_EXTERNAL_VECTOR; i++) { if (direct_trap(lg, &lg->idt[i], i)) idt[i] = lg->idt[i]; else default_idt_entry(&idt[i], i, def[i]); } + + /* Don't forget the system call trap! The IDT entries for other + * interupts never change, so no need to copy them. */ i = SYSCALL_VECTOR; if (direct_trap(lg, &lg->syscall_idt, i)) idt[i] = lg->syscall_idt; diff --git a/drivers/lguest/io.c b/drivers/lguest/io.c index c8eb792..ea68613 100644 --- a/drivers/lguest/io.c +++ b/drivers/lguest/io.c @@ -1,5 +1,9 @@ -/* Simple I/O model for guests, based on shared memory. - * Copyright (C) 2006 Rusty Russell IBM Corporation +/*P:300 The I/O mechanism in lguest is simple yet flexible, allowing the Guest + * to talk to the Launcher or directly to another Guest. It uses familiar + * concepts of DMA and interrupts, plus some neat code stolen from + * futexes... :*/ + +/* Copyright (C) 2006 Rusty Russell IBM Corporation * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -23,8 +27,36 @@ #include <linux/uaccess.h> #include "lg.h" +/*L:300 + * I/O + * + * Getting data in and out of the Guest is quite an art. There are numerous + * ways to do it, and they all suck differently. We try to keep things fairly + * close to "real" hardware so our Guest's drivers don't look like an alien + * visitation in the middle of the Linux code, and yet make sure that Guests + * can talk directly to other Guests, not just the Launcher. + * + * To do this, the Guest gives us a key when it binds or sends DMA buffers. + * The key corresponds to a "physical" address inside the Guest (ie. a virtual + * address inside the Launcher process). We don't, however, use this key + * directly. + * + * We want Guests which share memory to be able to DMA to each other: two + * Launchers can mmap memory the same file, then the Guests can communicate. + * Fortunately, the futex code provides us with a way to get a "union + * futex_key" corresponding to the memory lying at a virtual address: if the + * two processes share memory, the "union futex_key" for that memory will match + * even if the memory is mapped at different addresses in each. So we always + * convert the keys to "union futex_key"s to compare them. + * + * Before we dive into this though, we need to look at another set of helper + * routines used throughout the Host kernel code to access Guest memory. + :*/ static struct list_head dma_hash[61]; +/* An unfortunate side effect of the Linux double-linked list implementation is + * that there's no good way to statically initialize an array of linked + * lists. */ void lguest_io_init(void) { unsigned int i; @@ -56,6 +88,19 @@ kill: return 0; } +/*L:330 This is our hash function, using the wonderful Jenkins hash. + * + * The futex key is a union with three parts: an unsigned long word, a pointer, + * and an int "offset". We could use jhash_2words() which takes three u32s. + * (Ok, the hash functions are great: the naming sucks though). + * + * It's nice to be portable to 64-bit platforms, so we use the more generic + * jhash2(), which takes an array of u32, the number of u32s, and an initial + * u32 to roll in. This is uglier, but breaks down to almost the same code on + * 32-bit platforms like this one. + * + * We want a position in the array, so we modulo ARRAY_SIZE(dma_hash) (ie. 61). + */ static unsigned int hash(const union futex_key *key) { return jhash2((u32*)&key->both.word, @@ -64,6 +109,9 @@ static unsigned int hash(const union futex_key *key) % ARRAY_SIZE(dma_hash); } +/* This is a convenience routine to compare two keys. It's a much bemoaned C + * weakness that it doesn't allow '==' on structures or unions, so we have to + * open-code it like this. */ static inline int key_eq(const union futex_key *a, const union futex_key *b) { return (a->both.word == b->both.word @@ -71,22 +119,36 @@ static inline int key_eq(const union futex_key *a, const union futex_key *b) && a->both.offset == b->both.offset); } -/* Must hold read lock on dmainfo owner's current->mm->mmap_sem */ +/*L:360 OK, when we need to actually free up a Guest's DMA array we do several + * things, so we have a convenient function to do it. + * + * The caller must hold a read lock on dmainfo owner's current->mm->mmap_sem + * for the drop_futex_key_refs(). */ static void unlink_dma(struct lguest_dma_info *dmainfo) { + /* You locked this too, right? */ BUG_ON(!mutex_is_locked(&lguest_lock)); + /* This is how we know that the entry is free. */ dmainfo->interrupt = 0; + /* Remove it from the hash table. */ list_del(&dmainfo->list); + /* Drop the references we were holding (to the inode or mm). */ drop_futex_key_refs(&dmainfo->key); } +/*L:350 This is the routine which we call when the Guest asks to unregister a + * DMA array attached to a given key. Returns true if the array was found. */ static int unbind_dma(struct lguest *lg, const union futex_key *key, unsigned long dmas) { int i, ret = 0; + /* We don't bother with the hash table, just look through all this + * Guest's DMA arrays. */ for (i = 0; i < LGUEST_MAX_DMA; i++) { + /* In theory it could have more than one array on the same key, + * or one array on multiple keys, so we check both */ if (key_eq(key, &lg->dma[i].key) && dmas == lg->dma[i].dmas) { unlink_dma(&lg->dma[i]); ret = 1; @@ -96,51 +158,91 @@ static int unbind_dma(struct lguest *lg, return ret; } +/*L:340 BIND_DMA: this is the hypercall which sets up an array of "struct + * lguest_dma" for receiving I/O. + * + * The Guest wants to bind an array of "struct lguest_dma"s to a particular key + * to receive input. This only happens when the Guest is setting up a new + * device, so it doesn't have to be very fast. + * + * It returns 1 on a successful registration (it can fail if we hit the limit + * of registrations for this Guest). + */ int bind_dma(struct lguest *lg, unsigned long ukey, unsigned long dmas, u16 numdmas, u8 interrupt) { unsigned int i; int ret = 0; union futex_key key; + /* Futex code needs the mmap_sem. */ struct rw_semaphore *fshared = ¤t->mm->mmap_sem; + /* Invalid interrupt? (We could kill the guest here). */ if (interrupt >= LGUEST_IRQS) return 0; + /* We need to grab the Big Lguest Lock, because other Guests may be + * trying to look through this Guest's DMAs to send something while + * we're doing this. */ mutex_lock(&lguest_lock); down_read(fshared); if (get_futex_key((u32 __user *)ukey, fshared, &key) != 0) { kill_guest(lg, "bad dma key %#lx", ukey); goto unlock; } + + /* We want to keep this key valid once we drop mmap_sem, so we have to + * hold a reference. */ get_futex_key_refs(&key); + /* If the Guest specified an interrupt of 0, that means they want to + * unregister this array of "struct lguest_dma"s. */ if (interrupt == 0) ret = unbind_dma(lg, &key, dmas); else { + /* Look through this Guest's dma array for an unused entry. */ for (i = 0; i < LGUEST_MAX_DMA; i++) { + /* If the interrupt is non-zero, the entry is already + * used. */ if (lg->dma[i].interrupt) continue; + /* OK, a free one! Fill on our details. */ lg->dma[i].dmas = dmas; lg->dma[i].num_dmas = numdmas; lg->dma[i].next_dma = 0; lg->dma[i].key = key; lg->dma[i].guestid = lg->guestid; lg->dma[i].interrupt = interrupt; + + /* Now we add it to the hash table: the position + * depends on the futex key that we got. */ list_add(&lg->dma[i].list, &dma_hash[hash(&key)]); + /* Success! */ ret = 1; goto unlock; } } + /* If we didn't find a slot to put the key in, drop the reference + * again. */ drop_futex_key_refs(&key); unlock: + /* Unlock and out. */ up_read(fshared); mutex_unlock(&lguest_lock); return ret; } -/* lgread from another guest */ +/*L:385 Note that our routines to access a different Guest's memory are called + * lgread_other() and lgwrite_other(): these names emphasize that they are only + * used when the Guest is *not* the current Guest. + * + * The interface for copying from another process's memory is called + * access_process_vm(), with a final argument of 0 for a read, and 1 for a + * write. + * + * We need lgread_other() to read the destination Guest's "struct lguest_dma" + * array. */ static int lgread_other(struct lguest *lg, void *buf, u32 addr, unsigned bytes) { @@ -153,7 +255,8 @@ static int lgread_other(struct lguest *lg, return 1; } -/* lgwrite to another guest */ +/* "lgwrite()" to another Guest: used to update the destination "used_len" once + * we've transferred data into the buffer. */ static int lgwrite_other(struct lguest *lg, u32 addr, const void *buf, unsigned bytes) { @@ -166,6 +269,15 @@ static int lgwrite_other(struct lguest *lg, u32 addr, return 1; } +/*L:400 This is the generic engine which copies from a source "struct + * lguest_dma" from this Guest into another Guest's "struct lguest_dma". The + * destination Guest's pages have already been mapped, as contained in the + * pages array. + * + * If you're wondering if there's a nice "copy from one process to another" + * routine, so was I. But Linux isn't really set up to copy between two + * unrelated processes, so we have to write it ourselves. + */ static u32 copy_data(struct lguest *srclg, const struct lguest_dma *src, const struct lguest_dma *dst, @@ -174,33 +286,59 @@ static u32 copy_data(struct lguest *srclg, unsigned int totlen, si, di, srcoff, dstoff; void *maddr = NULL; + /* We return the total length transferred. */ totlen = 0; + + /* We keep indexes into the source and destination "struct lguest_dma", + * and an offset within each region. */ si = di = 0; srcoff = dstoff = 0; + + /* We loop until the source or destination is exhausted. */ while (si < LGUEST_MAX_DMA_SECTIONS && src->len[si] && di < LGUEST_MAX_DMA_SECTIONS && dst->len[di]) { + /* We can only transfer the rest of the src buffer, or as much + * as will fit into the destination buffer. */ u32 len = min(src->len[si] - srcoff, dst->len[di] - dstoff); + /* For systems using "highmem" we need to use kmap() to access + * the page we want. We often use the same page over and over, + * so rather than kmap() it on every loop, we set the maddr + * pointer to NULL when we need to move to the next + * destination page. */ if (!maddr) maddr = kmap(pages[di]); - /* FIXME: This is not completely portable, since - archs do different things for copy_to_user_page. */ + /* Copy directly from (this Guest's) source address to the + * destination Guest's kmap()ed buffer. Note that maddr points + * to the start of the page: we need to add the offset of the + * destination address and offset within the buffer. */ + + /* FIXME: This is not completely portable. I looked at + * copy_to_user_page(), and some arch's seem to need special + * flushes. x86 is fine. */ if (copy_from_user(maddr + (dst->addr[di] + dstoff)%PAGE_SIZE, (void __user *)src->addr[si], len) != 0) { + /* If a copy failed, it's the source's fault. */ kill_guest(srclg, "bad address in sending DMA"); totlen = 0; break; } + /* Increment the total and src & dst offsets */ totlen += len; srcoff += len; dstoff += len; + + /* Presumably we reached the end of the src or dest buffers: */ if (srcoff == src->len[si]) { + /* Move to the next buffer at offset 0 */ si++; srcoff = 0; } if (dstoff == dst->len[di]) { + /* We need to unmap that destination page and reset + * maddr ready for the next one. */ kunmap(pages[di]); maddr = NULL; di++; @@ -208,13 +346,15 @@ static u32 copy_data(struct lguest *srclg, } } + /* If we still had a page mapped at the end, unmap now. */ if (maddr) kunmap(pages[di]); return totlen; } -/* Src is us, ie. current. */ +/*L:390 This is how we transfer a "struct lguest_dma" from the source Guest + * (the current Guest which called SEND_DMA) to another Guest. */ static u32 do_dma(struct lguest *srclg, const struct lguest_dma *src, struct lguest *dstlg, const struct lguest_dma *dst) { @@ -222,23 +362,31 @@ static u32 do_dma(struct lguest *srclg, const struct lguest_dma *src, u32 ret; struct page *pages[LGUEST_MAX_DMA_SECTIONS]; + /* We check that both source and destination "struct lguest_dma"s are + * within the bounds of the source and destination Guests */ if (!check_dma_list(dstlg, dst) || !check_dma_list(srclg, src)) return 0; - /* First get the destination pages */ + /* We need to map the pages which correspond to each parts of + * destination buffer. */ for (i = 0; i < LGUEST_MAX_DMA_SECTIONS; i++) { if (dst->len[i] == 0) break; + /* get_user_pages() is a complicated function, especially since + * we only want a single page. But it works, and returns the + * number of pages. Note that we're holding the destination's + * mmap_sem, as get_user_pages() requires. */ if (get_user_pages(dstlg->tsk, dstlg->mm, dst->addr[i], 1, 1, 1, pages+i, NULL) != 1) { + /* This means the destination gave us a bogus buffer */ kill_guest(dstlg, "Error mapping DMA pages"); ret = 0; goto drop_pages; } } - /* Now copy until we run out of src or dst. */ + /* Now copy the data until we run out of src or dst. */ ret = copy_data(srclg, src, dst, pages); drop_pages: @@ -247,6 +395,11 @@ drop_pages: return ret; } +/*L:380 Transferring data from one Guest to another is not as simple as I'd + * like. We've found the "struct lguest_dma_info" bound to the same address as + * the send, we need to copy into it. + * + * This function returns true if the destination array was empty. */ static int dma_transfer(struct lguest *srclg, unsigned long udma, struct lguest_dma_info *dst) @@ -255,15 +408,23 @@ static int dma_transfer(struct lguest *srclg, struct lguest *dstlg; u32 i, dma = 0; + /* From the "struct lguest_dma_info" we found in the hash, grab the + * Guest. */ dstlg = &lguests[dst->guestid]; - /* Get our dma list. */ + /* Read in the source "struct lguest_dma" handed to SEND_DMA. */ lgread(srclg, &src_dma, udma, sizeof(src_dma)); - /* We can't deadlock against them dmaing to us, because this - * is all under the lguest_lock. */ + /* We need the destination's mmap_sem, and we already hold the source's + * mmap_sem for the futex key lookup. Normally this would suggest that + * we could deadlock if the destination Guest was trying to send to + * this source Guest at the same time, which is another reason that all + * I/O is done under the big lguest_lock. */ down_read(&dstlg->mm->mmap_sem); + /* Look through the destination DMA array for an available buffer. */ for (i = 0; i < dst->num_dmas; i++) { + /* We keep a "next_dma" pointer which often helps us avoid + * looking at lots of previously-filled entries. */ dma = (dst->next_dma + i) % dst->num_dmas; if (!lgread_other(dstlg, &dst_dma, dst->dmas + dma * sizeof(struct lguest_dma), @@ -273,30 +434,46 @@ static int dma_transfer(struct lguest *srclg, if (!dst_dma.used_len) break; } + + /* If we found a buffer, we do the actual data copy. */ if (i != dst->num_dmas) { unsigned long used_lenp; unsigned int ret; ret = do_dma(srclg, &src_dma, dstlg, &dst_dma); - /* Put used length in src. */ + /* Put used length in the source "struct lguest_dma"'s used_len + * field. It's a little tricky to figure out where that is, + * though. */ lgwrite_u32(srclg, udma+offsetof(struct lguest_dma, used_len), ret); + /* Tranferring 0 bytes is OK if the source buffer was empty. */ if (ret == 0 && src_dma.len[0] != 0) goto fail; - /* Make sure destination sees contents before length. */ + /* The destination Guest might be running on a different CPU: + * we have to make sure that it will see the "used_len" field + * change to non-zero *after* it sees the data we copied into + * the buffer. Hence a write memory barrier. */ wmb(); + /* Figuring out where the destination's used_len field for this + * "struct lguest_dma" in the array is also a little ugly. */ used_lenp = dst->dmas + dma * sizeof(struct lguest_dma) + offsetof(struct lguest_dma, used_len); lgwrite_other(dstlg, used_lenp, &ret, sizeof(ret)); + /* Move the cursor for next time. */ dst->next_dma++; } up_read(&dstlg->mm->mmap_sem); - /* Do this last so dst doesn't simply sleep on lock. */ + /* We trigger the destination interrupt, even if the destination was + * empty and we didn't transfer anything: this gives them a chance to + * wake up and refill. */ set_bit(dst->interrupt, dstlg->irqs_pending); + /* Wake up the destination process. */ wake_up_process(dstlg->tsk); + /* If we passed the last "struct lguest_dma", the receive had no + * buffers left. */ return i == dst->num_dmas; fail: @@ -304,6 +481,8 @@ fail: return 0; } +/*L:370 This is the counter-side to the BIND_DMA hypercall; the SEND_DMA + * hypercall. We find out who's listening, and send to them. */ void send_dma(struct lguest *lg, unsigned long ukey, unsigned long udma) { union futex_key key; @@ -313,31 +492,43 @@ void send_dma(struct lguest *lg, unsigned long ukey, unsigned long udma) again: mutex_lock(&lguest_lock); down_read(fshared); + /* Get the futex key for the key the Guest gave us */ if (get_futex_key((u32 __user *)ukey, fshared, &key) != 0) { kill_guest(lg, "bad sending DMA key"); goto unlock; } - /* Shared mapping? Look for other guests... */ + /* Since the key must be a multiple of 4, the futex key uses the lower + * bit of the "offset" field (which would always be 0) to indicate a + * mapping which is shared with other processes (ie. Guests). */ if (key.shared.offset & 1) { struct lguest_dma_info *i; + /* Look through the hash for other Guests. */ list_for_each_entry(i, &dma_hash[hash(&key)], list) { + /* Don't send to ourselves. */ if (i->guestid == lg->guestid) continue; if (!key_eq(&key, &i->key)) continue; + /* If dma_transfer() tells us the destination has no + * available buffers, we increment "empty". */ empty += dma_transfer(lg, udma, i); break; } + /* If the destination is empty, we release our locks and + * give the destination Guest a brief chance to restock. */ if (empty == 1) { /* Give any recipients one chance to restock. */ up_read(¤t->mm->mmap_sem); mutex_unlock(&lguest_lock); + /* Next time, we won't try again. */ empty++; goto again; } } else { - /* Private mapping: tell our userspace. */ + /* Private mapping: Guest is sending to its Launcher. We set + * the "dma_is_pending" flag so that the main loop will exit + * and the Launcher's read() from /dev/lguest will return. */ lg->dma_is_pending = 1; lg->pending_dma = udma; lg->pending_key = ukey; @@ -346,6 +537,7 @@ unlock: up_read(fshared); mutex_unlock(&lguest_lock); } +/*:*/ void release_all_dma(struct lguest *lg) { @@ -361,7 +553,18 @@ void release_all_dma(struct lguest *lg) up_read(&lg->mm->mmap_sem); } -/* Userspace wants a dma buffer from this guest. */ +/*M:007 We only return a single DMA buffer to the Launcher, but it would be + * more efficient to return a pointer to the entire array of DMA buffers, which + * it can cache and choose one whenever it wants. + * + * Currently the Launcher uses a write to /dev/lguest, and the return value is + * the address of the DMA structure with the interrupt number placed in + * dma->used_len. If we wanted to return the entire array, we need to return + * the address, array size and interrupt number: this seems to require an + * ioctl(). :*/ + +/*L:320 This routine looks for a DMA buffer registered by the Guest on the + * given key (using the BIND_DMA hypercall). */ unsigned long get_dma_buffer(struct lguest *lg, unsigned long ukey, unsigned long *interrupt) { @@ -370,15 +573,29 @@ unsigned long get_dma_buffer(struct lguest *lg, struct lguest_dma_info *i; struct rw_semaphore *fshared = ¤t->mm->mmap_sem; + /* Take the Big Lguest Lock to stop other Guests sending this Guest DMA + * at the same time. */ mutex_lock(&lguest_lock); + /* To match between Guests sharing the same underlying memory we steal + * code from the futex infrastructure. This requires that we hold the + * "mmap_sem" for our process (the Launcher), and pass it to the futex + * code. */ down_read(fshared); + + /* This can fail if it's not a valid address, or if the address is not + * divisible by 4 (the futex code needs that, we don't really). */ if (get_futex_key((u32 __user *)ukey, fshared, &key) != 0) { kill_guest(lg, "bad registered DMA buffer"); goto unlock; } + /* Search the hash table for matching entries (the Launcher can only + * send to its own Guest for the moment, so the entry must be for this + * Guest) */ list_for_each_entry(i, &dma_hash[hash(&key)], list) { if (key_eq(&key, &i->key) && i->guestid == lg->guestid) { unsigned int j; + /* Look through the registered DMA array for an + * available buffer. */ for (j = 0; j < i->num_dmas; j++) { struct lguest_dma dma; @@ -387,6 +604,8 @@ unsigned long get_dma_buffer(struct lguest *lg, if (dma.used_len == 0) break; } + /* Store the interrupt the Guest wants when the buffer + * is used. */ *interrupt = i->interrupt; break; } @@ -396,4 +615,12 @@ unlock: mutex_unlock(&lguest_lock); return ret; } +/*:*/ +/*L:410 This really has completed the Launcher. Not only have we now finished + * the longest chapter in our journey, but this also means we are over halfway + * through! + * + * Enough prevaricating around the bush: it is time for us to dive into the + * core of the Host, in "make Host". + */ diff --git a/drivers/lguest/lg.h b/drivers/lguest/lg.h index 3e2ddfb..64f0abe 100644 --- a/drivers/lguest/lg.h +++ b/drivers/lguest/lg.h @@ -58,9 +58,18 @@ struct lguest_dma_info u8 interrupt; /* 0 when not registered */ }; -/* We have separate types for the guest's ptes & pgds and the shadow ptes & - * pgds. Since this host might use three-level pagetables and the guest and - * shadow pagetables don't, we can't use the normal pte_t/pgd_t. */ +/*H:310 The page-table code owes a great debt of gratitude to Andi Kleen. He + * reviewed the original code which used "u32" for all page table entries, and + * insisted that it would be far clearer with explicit typing. I thought it + * was overkill, but he was right: it is much clearer than it was before. + * + * We have separate types for the Guest's ptes & pgds and the shadow ptes & + * pgds. There's already a Linux type for these (pte_t and pgd_t) but they + * change depending on kernel config options (PAE). */ + +/* Each entry is identical: lower 12 bits of flags and upper 20 bits for the + * "page frame number" (0 == first physical page, etc). They are different + * types so the compiler will warn us if we mix them improperly. */ typedef union { struct { unsigned flags:12, pfn:20; }; struct { unsigned long val; } raw; @@ -77,8 +86,12 @@ typedef union { struct { unsigned flags:12, pfn:20; }; struct { unsigned long val; } raw; } gpte_t; + +/* We have two convenient macros to convert a "raw" value as handed to us by + * the Guest into the correct Guest PGD or PTE type. */ #define mkgpte(_val) ((gpte_t){.raw.val = _val}) #define mkgpgd(_val) ((gpgd_t){.raw.val = _val}) +/*:*/ struct pgdir { @@ -243,7 +256,32 @@ unsigned long get_dma_buffer(struct lguest *lg, unsigned long key, /* hypercalls.c: */ void do_hypercalls(struct lguest *lg); - +void write_timestamp(struct lguest *lg); + +/*L:035 + * Let's step aside for the moment, to study one important routine that's used + * widely in the Host code. + * + * There are many cases where the Guest does something invalid, like pass crap + * to a hypercall. Since only the Guest kernel can make hypercalls, it's quite + * acceptable to simply terminate the Guest and give the Launcher a nicely + * formatted reason. It's also simpler for the Guest itself, which doesn't + * need to check most hypercalls for "success"; if you're still running, it + * succeeded. + * + * Once this is called, the Guest will never run again, so most Host code can + * call this then continue as if nothing had happened. This means many + * functions don't have to explicitly return an error code, which keeps the + * code simple. + * + * It also means that this can be called more than once: only the first one is + * remembered. The only trick is that we still need to kill the Guest even if + * we can't allocate memory to store the reason. Linux has a neat way of + * packing error codes into invalid pointers, so we use that here. + * + * Like any macro which uses an "if", it is safely wrapped in a run-once "do { + * } while(0)". + */ #define kill_guest(lg, fmt...) \ do { \ if (!(lg)->dead) { \ @@ -252,6 +290,7 @@ do { \ (lg)->dead = ERR_PTR(-ENOMEM); \ } \ } while(0) +/* (End of aside) :*/ static inline unsigned long guest_pa(struct lguest *lg, unsigned long vaddr) { diff --git a/drivers/lguest/lguest.c b/drivers/lguest/lguest.c index 18dade0..1bc1546 100644 --- a/drivers/lguest/lguest.c +++ b/drivers/lguest/lguest.c @@ -1,6 +1,32 @@ -/* - * Lguest specific paravirt-ops implementation +/*P:010 + * A hypervisor allows multiple Operating Systems to run on a single machine. + * To quote David Wheeler: "Any problem in computer science can be solved with + * another layer of indirection." + * + * We keep things simple in two ways. First, we start with a normal Linux + * kernel and insert a module (lg.ko) which allows us to run other Linux + * kernels the same way we'd run processes. We call the first kernel the Host, + * and the others the Guests. The program which sets up and configures Guests + * (such as the example in Documentation/lguest/lguest.c) is called the + * Launcher. + * + * Secondly, we only run specially modified Guests, not normal kernels. When + * you set CONFIG_LGUEST to 'y' or 'm', this automatically sets + * CONFIG_LGUEST_GUEST=y, which compiles this file into the kernel so it knows + * how to be a Guest. This means that you can use the same kernel you boot + * normally (ie. as a Host) as a Guest. * + * These Guests know that they cannot do privileged operations, such as disable + * interrupts, and that they have to ask the Host to do such things explicitly. + * This file consists of all the replacements for such low-level native + * hardware operations: these special Guest versions call the Host. + * + * So how does the kernel know it's a Guest? The Guest starts at a special + * entry point marked with a magic string, which sets up a few things then + * calls here. We replace the native functions in "struct paravirt_ops" + * with our Guest versions, then boot like normal. :*/ + +/* * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation. * * This program is free software; you can redistribute it and/or modify @@ -40,6 +66,12 @@ #include <asm/mce.h> #include <asm/io.h> +/*G:010 Welcome to the Guest! + * + * The Guest in our tale is a simple creature: identical to the Host but + * behaving in simplified but equivalent ways. In particular, the Guest is the + * same kernel as the Host (or at least, built from the same source code). :*/ + /* Declarations for definitions in lguest_guest.S */ extern char lguest_noirq_start[], lguest_noirq_end[]; extern const char lgstart_cli[], lgend_cli[]; @@ -58,7 +90,26 @@ struct lguest_data lguest_data = { struct lguest_device_desc *lguest_devices; static cycle_t clock_base; -static enum paravirt_lazy_mode lazy_mode; +/*G:035 Notice the lazy_hcall() above, rather than hcall(). This is our first + * real optimization trick! + * + * When lazy_mode is set, it means we're allowed to defer all hypercalls and do + * them as a batch when lazy_mode is eventually turned off. Because hypercalls + * are reasonably expensive, batching them up makes sense. For example, a + * large mmap might update dozens of page table entries: that code calls + * lguest_lazy_mode(PARAVIRT_LAZY_MMU), does the dozen updates, then calls + * lguest_lazy_mode(PARAVIRT_LAZY_NONE). + * + * So, when we're in lazy mode, we call async_hypercall() to store the call for + * future processing. When lazy mode is turned off we issue a hypercall to + * flush the stored calls. + * + * There's also a hack where "mode" is set to "PARAVIRT_LAZY_FLUSH" which + * indicates we're to flush any outstanding calls immediately. This is used + * when an interrupt handler does a kmap_atomic(): the page table changes must + * happen immediately even if we're in the middle of a batch. Usually we're + * not, though, so there's nothing to do. */ +static enum paravirt_lazy_mode lazy_mode; /* Note: not SMP-safe! */ static void lguest_lazy_mode(enum paravirt_lazy_mode mode) { if (mode == PARAVIRT_LAZY_FLUSH) { @@ -82,6 +133,16 @@ static void lazy_hcall(unsigned long call, async_hcall(call, arg1, arg2, arg3); } +/* async_hcall() is pretty simple: I'm quite proud of it really. We have a + * ring buffer of stored hypercalls which the Host will run though next time we + * do a normal hypercall. Each entry in the ring has 4 slots for the hypercall + * arguments, and a "hcall_status" word which is 0 if the call is ready to go, + * and 255 once the Host has finished with it. + * + * If we come around to a slot which hasn't been finished, then the table is + * full and we just make the hypercall directly. This has the nice side + * effect of causing the Host to run all the stored calls in the ring buffer + * which empties it for next time! */ void async_hcall(unsigned long call, unsigned long arg1, unsigned long arg2, unsigned long arg3) { @@ -89,6 +150,9 @@ void async_hcall(unsigned long call, static unsigned int next_call; unsigned long flags; + /* Disable interrupts if not already disabled: we don't want an + * interrupt handler making a hypercall while we're already doing + * one! */ local_irq_save(flags); if (lguest_data.hcall_status[next_call] != 0xFF) { /* Table full, so do normal hcall which will flush table. */ @@ -98,7 +162,7 @@ void async_hcall(unsigned long call, lguest_data.hcalls[next_call].edx = arg1; lguest_data.hcalls[next_call].ebx = arg2; lguest_data.hcalls[next_call].ecx = arg3; - /* Make sure host sees arguments before "valid" flag. */ + /* Arguments must all be written before we mark it to go */ wmb(); lguest_data.hcall_status[next_call] = 0; if (++next_call == LHCALL_RING_SIZE) @@ -106,9 +170,14 @@ void async_hcall(unsigned long call, } local_irq_restore(flags); } +/*:*/ +/* Wrappers for the SEND_DMA and BIND_DMA hypercalls. This is mainly because + * Jeff Garzik complained that __pa() should never appear in drivers, and this + * helps remove most of them. But also, it wraps some ugliness. */ void lguest_send_dma(unsigned long key, struct lguest_dma *dma) { + /* The hcall might not write this if something goes wrong */ dma->used_len = 0; hcall(LHCALL_SEND_DMA, key, __pa(dma), 0); } @@ -116,11 +185,16 @@ void lguest_send_dma(unsigned long key, struct lguest_dma *dma) int lguest_bind_dma(unsigned long key, struct lguest_dma *dmas, unsigned int num, u8 irq) { + /* This is the only hypercall which actually wants 5 arguments, and we + * only support 4. Fortunately the interrupt number is always less + * than 256, so we can pack it with the number of dmas in the final + * argument. */ if (!hcall(LHCALL_BIND_DMA, key, __pa(dmas), (num << 8) | irq)) return -ENOMEM; return 0; } +/* Unbinding is the same hypercall as binding, but with 0 num & irq. */ void lguest_unbind_dma(unsigned long key, struct lguest_dma *dmas) { hcall(LHCALL_BIND_DMA, key, __pa(dmas), 0); @@ -138,35 +212,73 @@ void lguest_unmap(void *addr) iounmap((__force void __iomem *)addr); } +/*G:033 + * Here are our first native-instruction replacements: four functions for + * interrupt control. + * + * The simplest way of implementing these would be to have "turn interrupts + * off" and "turn interrupts on" hypercalls. Unfortunately, this is too slow: + * these are by far the most commonly called functions of those we override. + * + * So instead we keep an "irq_enabled" field inside our "struct lguest_data", + * which the Guest can update with a single instruction. The Host knows to + * check there when it wants to deliver an interrupt. + */ + +/* save_flags() is expected to return the processor state (ie. "eflags"). The + * eflags word contains all kind of stuff, but in practice Linux only cares + * about the interrupt flag. Our "save_flags()" just returns that. */ static unsigned long save_fl(void) { return lguest_data.irq_enabled; } +/* "restore_flags" just sets the flags back to the value given. */ static void restore_fl(unsigned long flags) { - /* FIXME: Check if interrupt pending... */ lguest_data.irq_enabled = flags; } +/* Interrupts go off... */ static void irq_disable(void) { lguest_data.irq_enabled = 0; } +/* Interrupts go on... */ static void irq_enable(void) { - /* FIXME: Check if interrupt pending... */ lguest_data.irq_enabled = X86_EFLAGS_IF; } - +/*:*/ +/*M:003 Note that we don't check for outstanding interrupts when we re-enable + * them (or when we unmask an interrupt). This seems to work for the moment, + * since interrupts are rare and we'll just get the interrupt on the next timer + * tick, but when we turn on CONFIG_NO_HZ, we should revisit this. One way + * would be to put the "irq_enabled" field in a page by itself, and have the + * Host write-protect it when an interrupt comes in when irqs are disabled. + * There will then be a page fault as soon as interrupts are re-enabled. :*/ + +/*G:034 + * The Interrupt Descriptor Table (IDT). + * + * The IDT tells the processor what to do when an interrupt comes in. Each + * entry in the table is a 64-bit descriptor: this holds the privilege level, + * address of the handler, and... well, who cares? The Guest just asks the + * Host to make the change anyway, because the Host controls the real IDT. + */ static void lguest_write_idt_entry(struct desc_struct *dt, int entrynum, u32 low, u32 high) { + /* Keep the local copy up to date. */ write_dt_entry(dt, entrynum, low, high); + /* Tell Host about this new entry. */ hcall(LHCALL_LOAD_IDT_ENTRY, entrynum, low, high); } +/* Changing to a different IDT is very rare: we keep the IDT up-to-date every + * time it is written, so we can simply loop through all entries and tell the + * Host about them. */ static void lguest_load_idt(const struct Xgt_desc_struct *desc) { unsigned int i; @@ -176,12 +288,29 @@ static void lguest_load_idt(const struct Xgt_desc_struct *desc) hcall(LHCALL_LOAD_IDT_ENTRY, i, idt[i].a, idt[i].b); } +/* + * The Global Descriptor Table. + * + * The Intel architecture defines another table, called the Global Descriptor + * Table (GDT). You tell the CPU where it is (and its size) using the "lgdt" + * instruction, and then several other instructions refer to entries in the + * table. There are three entries which the Switcher needs, so the Host simply + * controls the entire thing and the Guest asks it to make changes using the + * LOAD_GDT hypercall. + * + * This is the opposite of the IDT code where we have a LOAD_IDT_ENTRY + * hypercall and use that repeatedly to load a new IDT. I don't think it + * really matters, but wouldn't it be nice if they were the same? + */ static void lguest_load_gdt(const struct Xgt_desc_struct *desc) { BUG_ON((desc->size+1)/8 != GDT_ENTRIES); hcall(LHCALL_LOAD_GDT, __pa(desc->address), GDT_ENTRIES, 0); } +/* For a single GDT entry which changes, we do the lazy thing: alter our GDT, + * then tell the Host to reload the entire thing. This operation is so rare + * that this naive implementation is reasonable. */ static void lguest_write_gdt_entry(struct desc_struct *dt, int entrynum, u32 low, u32 high) { @@ -189,19 +318,58 @@ static void lguest_write_gdt_entry(struct desc_struct *dt, hcall(LHCALL_LOAD_GDT, __pa(dt), GDT_ENTRIES, 0); } +/* OK, I lied. There are three "thread local storage" GDT entries which change + * on every context switch (these three entries are how glibc implements + * __thread variables). So we have a hypercall specifically for this case. */ static void lguest_load_tls(struct thread_struct *t, unsigned int cpu) { lazy_hcall(LHCALL_LOAD_TLS, __pa(&t->tls_array), cpu, 0); } +/*:*/ +/*G:038 That's enough excitement for now, back to ploughing through each of + * the paravirt_ops (we're about 1/3 of the way through). + * + * This is the Local Descriptor Table, another weird Intel thingy. Linux only + * uses this for some strange applications like Wine. We don't do anything + * here, so they'll get an informative and friendly Segmentation Fault. */ static void lguest_set_ldt(const void *addr, unsigned entries) { } +/* This loads a GDT entry into the "Task Register": that entry points to a + * structure called the Task State Segment. Some comments scattered though the + * kernel code indicate that this used for task switching in ages past, along + * with blood sacrifice and astrology. + * + * Now there's nothing interesting in here that we don't get told elsewhere. + * But the native version uses the "ltr" instruction, which makes the Host + * complain to the Guest about a Segmentation Fault and it'll oops. So we + * override the native version with a do-nothing version. */ static void lguest_load_tr_desc(void) { } +/* The "cpuid" instruction is a way of querying both the CPU identity + * (manufacturer, model, etc) and its features. It was introduced before the + * Pentium in 1993 and keeps getting extended by both Intel and AMD. As you + * might imagine, after a decade and a half this treatment, it is now a giant + * ball of hair. Its entry in the current Intel manual runs to 28 pages. + * + * This instruction even it has its own Wikipedia entry. The Wikipedia entry + * has been translated into 4 languages. I am not making this up! + * + * We could get funky here and identify ourselves as "GenuineLguest", but + * instead we just use the real "cpuid" instruction. Then I pretty much turned + * off feature bits until the Guest booted. (Don't say that: you'll damage + * lguest sales!) Shut up, inner voice! (Hey, just pointing out that this is + * hardly future proof.) Noone's listening! They don't like you anyway, + * parenthetic weirdo! + * + * Replacing the cpuid so we can turn features off is great for the kernel, but + * anyone (including userspace) can just use the raw "cpuid" instruction and + * the Host won't even notice since it isn't privileged. So we try not to get + * too worked up about it. */ static void lguest_cpuid(unsigned int *eax, unsigned int *ebx, unsigned int *ecx, unsigned int *edx) { @@ -214,21 +382,43 @@ static void lguest_cpuid(unsigned int *eax, unsigned int *ebx, *ecx &= 0x00002201; /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, FPU. */ *edx &= 0x07808101; - /* Host wants to know when we flush kernel pages: set PGE. */ + /* The Host can do a nice optimization if it knows that the + * kernel mappings (addresses above 0xC0000000 or whatever + * PAGE_OFFSET is set to) haven't changed. But Linux calls + * flush_tlb_user() for both user and kernel mappings unless + * the Page Global Enable (PGE) feature bit is set. */ *edx |= 0x00002000; break; case 0x80000000: /* Futureproof this a little: if they ask how much extended - * processor information, limit it to known fields. */ + * processor information there is, limit it to known fields. */ if (*eax > 0x80000008) *eax = 0x80000008; break; } } +/* Intel has four control registers, imaginatively named cr0, cr2, cr3 and cr4. + * I assume there's a cr1, but it hasn't bothered us yet, so we'll not bother + * it. The Host needs to know when the Guest wants to change them, so we have + * a whole series of functions like read_cr0() and write_cr0(). + * + * We start with CR0. CR0 allows you to turn on and off all kinds of basic + * features, but Linux only really cares about one: the horrifically-named Task + * Switched (TS) bit at bit 3 (ie. 8) + * + * What does the TS bit do? Well, it causes the CPU to trap (interrupt 7) if + * the floating point unit is used. Which allows us to restore FPU state + * lazily after a task switch, and Linux uses that gratefully, but wouldn't a + * name like "FPUTRAP bit" be a little less cryptic? + * + * We store cr0 (and cr3) locally, because the Host never changes it. The + * Guest sometimes wants to read it and we'd prefer not to bother the Host + * unnecessarily. */ static unsigned long current_cr0, current_cr3; static void lguest_write_cr0(unsigned long val) { + /* 8 == TS bit. */ lazy_hcall(LHCALL_TS, val & 8, 0, 0); current_cr0 = val; } @@ -238,17 +428,25 @@ static unsigned long lguest_read_cr0(void) return current_cr0; } +/* Intel provided a special instruction to clear the TS bit for people too cool + * to use write_cr0() to do it. This "clts" instruction is faster, because all + * the vowels have been optimized out. */ static void lguest_clts(void) { lazy_hcall(LHCALL_TS, 0, 0, 0); current_cr0 &= ~8U; } +/* CR2 is the virtual address of the last page fault, which the Guest only ever + * reads. The Host kindly writes this into our "struct lguest_data", so we + * just read it out of there. */ static unsigned long lguest_read_cr2(void) { return lguest_data.cr2; } +/* CR3 is the current toplevel pagetable page: the principle is the same as + * cr0. Keep a local copy, and tell the Host when it changes. */ static void lguest_write_cr3(unsigned long cr3) { lazy_hcall(LHCALL_NEW_PGTABLE, cr3, 0, 0); @@ -260,7 +458,7 @@ static unsigned long lguest_read_cr3(void) return current_cr3; } -/* Used to enable/disable PGE, but we don't care. */ +/* CR4 is used to enable and disable PGE, but we don't care. */ static unsigned long lguest_read_cr4(void) { return 0; @@ -270,6 +468,59 @@ static void lguest_write_cr4(unsigned long val) { } +/* + * Page Table Handling. + * + * Now would be a good time to take a rest and grab a coffee or similarly + * relaxing stimulant. The easy parts are behind us, and the trek gradually + * winds uphill from here. + * + * Quick refresher: memory is divided into "pages" of 4096 bytes each. The CPU + * maps virtual addresses to physical addresses using "page tables". We could + * use one huge index of 1 million entries: each address is 4 bytes, so that's + * 1024 pages just to hold the page tables. But since most virtual addresses + * are unused, we use a two level index which saves space. The CR3 register + * contains the physical address of the top level "page directory" page, which + * contains physical addresses of up to 1024 second-level pages. Each of these + * second level pages contains up to 1024 physical addresses of actual pages, + * or Page Table Entries (PTEs). + * + * Here's a diagram, where arrows indicate physical addresses: + * + * CR3 ---> +---------+ + * | --------->+---------+ + * | | | PADDR1 | + * Top-level | | PADDR2 | + * (PMD) page | | | + * | | Lower-level | + * | | (PTE) page | + * | | | | + * .... .... + * + * So to convert a virtual address to a physical address, we look up the top + * level, which points us to the second level, which gives us the physical + * address of that page. If the top level entry was not present, or the second + * level entry was not present, then the virtual address is invalid (we + * say "the page was not mapped"). + * + * Put another way, a 32-bit virtual address is divided up like so: + * + * 1 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 + * |<---- 10 bits ---->|<---- 10 bits ---->|<------ 12 bits ------>| + * Index into top Index into second Offset within page + * page directory page pagetable page + * + * The kernel spends a lot of time changing both the top-level page directory + * and lower-level pagetable pages. The Guest doesn't know physical addresses, + * so while it maintains these page tables exactly like normal, it also needs + * to keep the Host informed whenever it makes a change: the Host will create + * the real page tables based on the Guests'. + */ + +/* The Guest calls this to set a second-level entry (pte), ie. to map a page + * into a process' address space. We set the entry then tell the Host the + * toplevel and address this corresponds to. The Guest uses one pagetable per + * process, so we need to tell the Host which one we're changing (mm->pgd). */ static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pteval) { @@ -277,7 +528,9 @@ static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr, lazy_hcall(LHCALL_SET_PTE, __pa(mm->pgd), addr, pteval.pte_low); } -/* We only support two-level pagetables at the moment. */ +/* The Guest calls this to set a top-level entry. Again, we set the entry then + * tell the Host which top-level page we changed, and the index of the entry we + * changed. */ static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) { *pmdp = pmdval; @@ -285,7 +538,15 @@ static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval) (__pa(pmdp)&(PAGE_SIZE-1))/4, 0); } -/* FIXME: Eliminate all callers of this. */ +/* There are a couple of legacy places where the kernel sets a PTE, but we + * don't know the top level any more. This is useless for us, since we don't + * know which pagetable is changing or what address, so we just tell the Host + * to forget all of them. Fortunately, this is very rare. + * + * ... except in early boot when the kernel sets up the initial pagetables, + * which makes booting astonishingly slow. So we don't even tell the Host + * anything changed until we've done the first page table switch. + */ static void lguest_set_pte(pte_t *ptep, pte_t pteval) { *ptep = pteval; @@ -294,22 +555,51 @@ static void lguest_set_pte(pte_t *ptep, pte_t pteval) lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0); } +/* Unfortunately for Lguest, the paravirt_ops for page tables were based on + * native page table operations. On native hardware you can set a new page + * table entry whenever you want, but if you want to remove one you have to do + * a TLB flush (a TLB is a little cache of page table entries kept by the CPU). + * + * So the lguest_set_pte_at() and lguest_set_pmd() functions above are only + * called when a valid entry is written, not when it's removed (ie. marked not + * present). Instead, this is where we come when the Guest wants to remove a + * page table entry: we tell the Host to set that entry to 0 (ie. the present + * bit is zero). */ static void lguest_flush_tlb_single(unsigned long addr) { - /* Simply set it to zero, and it will fault back in. */ + /* Simply set it to zero: if it was not, it will fault back in. */ lazy_hcall(LHCALL_SET_PTE, current_cr3, addr, 0); } +/* This is what happens after the Guest has removed a large number of entries. + * This tells the Host that any of the page table entries for userspace might + * have changed, ie. virtual addresses below PAGE_OFFSET. */ static void lguest_flush_tlb_user(void) { lazy_hcall(LHCALL_FLUSH_TLB, 0, 0, 0); } +/* This is called when the kernel page tables have changed. That's not very + * common (unless the Guest is using highmem, which makes the Guest extremely + * slow), so it's worth separating this from the user flushing above. */ static void lguest_flush_tlb_kernel(void) { lazy_hcall(LHCALL_FLUSH_TLB, 1, 0, 0); } +/* + * The Unadvanced Programmable Interrupt Controller. + * + * This is an attempt to implement the simplest possible interrupt controller. + * I spent some time looking though routines like set_irq_chip_and_handler, + * set_irq_chip_and_handler_name, set_irq_chip_data and set_phasers_to_stun and + * I *think* this is as simple as it gets. + * + * We can tell the Host what interrupts we want blocked ready for using the + * lguest_data.interrupts bitmap, so disabling (aka "masking") them is as + * simple as setting a bit. We don't actually "ack" interrupts as such, we + * just mask and unmask them. I wonder if we should be cleverer? + */ static void disable_lguest_irq(unsigned int irq) { set_bit(irq, lguest_data.blocked_interrupts); @@ -318,9 +608,9 @@ static void disable_lguest_irq(unsigned int irq) static void enable_lguest_irq(unsigned int irq) { clear_bit(irq, lguest_data.blocked_interrupts); - /* FIXME: If it's pending? */ } +/* This structure describes the lguest IRQ controller. */ static struct irq_chip lguest_irq_controller = { .name = "lguest", .mask = disable_lguest_irq, @@ -328,6 +618,10 @@ static struct irq_chip lguest_irq_controller = { .unmask = enable_lguest_irq, }; +/* This sets up the Interrupt Descriptor Table (IDT) entry for each hardware + * interrupt (except 128, which is used for system calls), and then tells the + * Linux infrastructure that each interrupt is controlled by our level-based + * lguest interrupt controller. */ static void __init lguest_init_IRQ(void) { unsigned int i; @@ -340,20 +634,51 @@ static void __init lguest_init_IRQ(void) handle_level_irq); } } + /* This call is required to set up for 4k stacks, where we have + * separate stacks for hard and soft interrupts. */ irq_ctx_init(smp_processor_id()); } +/* + * Time. + * + * It would be far better for everyone if the Guest had its own clock, but + * until then the Host gives us the time on every interrupt. + */ static unsigned long lguest_get_wallclock(void) { - return hcall(LHCALL_GET_WALLCLOCK, 0, 0, 0); + return lguest_data.time.tv_sec; } static cycle_t lguest_clock_read(void) { + unsigned long sec, nsec; + + /* If the Host tells the TSC speed, we can trust that. */ if (lguest_data.tsc_khz) return native_read_tsc(); - else - return jiffies; + + /* If we can't use the TSC, we read the time value written by the Host. + * Since it's in two parts (seconds and nanoseconds), we risk reading + * it just as it's changing from 99 & 0.999999999 to 100 and 0, and + * getting 99 and 0. As Linux tends to come apart under the stress of + * time travel, we must be careful: */ + do { + /* First we read the seconds part. */ + sec = lguest_data.time.tv_sec; + /* This read memory barrier tells the compiler and the CPU that + * this can't be reordered: we have to complete the above + * before going on. */ + rmb(); + /* Now we read the nanoseconds part. */ + nsec = lguest_data.time.tv_nsec; + /* Make sure we've done that. */ + rmb(); + /* Now if the seconds part has changed, try again. */ + } while (unlikely(lguest_data.time.tv_sec != sec)); + + /* Our non-TSC clock is in real nanoseconds. */ + return sec*1000000000ULL + nsec; } /* This is what we tell the kernel is our clocksource. */ @@ -361,8 +686,11 @@ static struct clocksource lguest_clock = { .name = "lguest", .rating = 400, .read = lguest_clock_read, + .mask = CLOCKSOURCE_MASK(64), + .mult = 1, }; +/* The "scheduler clock" is just our real clock, adjusted to start at zero */ static unsigned long long lguest_sched_clock(void) { return cyc2ns(&lguest_clock, lguest_clock_read() - clock_base); @@ -428,34 +756,55 @@ static void lguest_time_irq(unsigned int irq, struct irq_desc *desc) local_irq_restore(flags); } +/* At some point in the boot process, we get asked to set up our timing + * infrastructure. The kernel doesn't expect timer interrupts before this, but + * we cleverly initialized the "blocked_interrupts" field of "struct + * lguest_data" so that timer interrupts were blocked until now. */ static void lguest_time_init(void) { + /* Set up the timer interrupt (0) to go to our simple timer routine */ set_irq_handler(0, lguest_time_irq); - /* We use the TSC if the Host tells us we can, otherwise a dumb - * jiffies-based clock. */ + /* Our clock structure look like arch/i386/kernel/tsc.c if we can use + * the TSC, otherwise it's a dumb nanosecond-resolution clock. Either + * way, the "rating" is initialized so high that it's always chosen + * over any other clocksource. */ if (lguest_data.tsc_khz) { lguest_clock.shift = 22; lguest_clock.mult = clocksource_khz2mult(lguest_data.tsc_khz, lguest_clock.shift); - lguest_clock.mask = CLOCKSOURCE_MASK(64); lguest_clock.flags = CLOCK_SOURCE_IS_CONTINUOUS; - } else { - /* To understand this, start at kernel/time/jiffies.c... */ - lguest_clock.shift = 8; - lguest_clock.mult = (((u64)NSEC_PER_SEC<<8)/ACTHZ) << 8; - lguest_clock.mask = CLOCKSOURCE_MASK(32); } clock_base = lguest_clock_read(); clocksource_register(&lguest_clock); - /* We can't set cpumask in the initializer: damn C limitations! */ + /* Now we've set up our clock, we can use it as the scheduler clock */ + paravirt_ops.sched_clock = lguest_sched_clock; + + /* We can't set cpumask in the initializer: damn C limitations! Set it + * here and register our timer device. */ lguest_clockevent.cpumask = cpumask_of_cpu(0); clockevents_register_device(&lguest_clockevent); + /* Finally, we unblock the timer interrupt. */ enable_lguest_irq(0); } +/* + * Miscellaneous bits and pieces. + * + * Here is an oddball collection of functions which the Guest needs for things + * to work. They're pretty simple. + */ + +/* The Guest needs to tell the host what stack it expects traps to use. For + * native hardware, this is part of the Task State Segment mentioned above in + * lguest_load_tr_desc(), but to help hypervisors there's this special call. + * + * We tell the Host the segment we want to use (__KERNEL_DS is the kernel data + * segment), the privilege level (we're privilege level 1, the Host is 0 and + * will not tolerate us trying to use that), the stack pointer, and the number + * of pages in the stack. */ static void lguest_load_esp0(struct tss_struct *tss, struct thread_struct *thread) { @@ -463,15 +812,31 @@ static void lguest_load_esp0(struct tss_struct *tss, THREAD_SIZE/PAGE_SIZE); } +/* Let's just say, I wouldn't do debugging under a Guest. */ static void lguest_set_debugreg(int regno, unsigned long value) { /* FIXME: Implement */ } +/* There are times when the kernel wants to make sure that no memory writes are + * caught in the cache (that they've all reached real hardware devices). This + * doesn't matter for the Guest which has virtual hardware. + * + * On the Pentium 4 and above, cpuid() indicates that the Cache Line Flush + * (clflush) instruction is available and the kernel uses that. Otherwise, it + * uses the older "Write Back and Invalidate Cache" (wbinvd) instruction. + * Unlike clflush, wbinvd can only be run at privilege level 0. So we can + * ignore clflush, but replace wbinvd. + */ static void lguest_wbinvd(void) { } +/* If the Guest expects to have an Advanced Programmable Interrupt Controller, + * we play dumb by ignoring writes and returning 0 for reads. So it's no + * longer Programmable nor Controlling anything, and I don't think 8 lines of + * code qualifies for Advanced. It will also never interrupt anything. It + * does, however, allow us to get through the Linux boot code. */ #ifdef CONFIG_X86_LOCAL_APIC static void lguest_apic_write(unsigned long reg, unsigned long v) { @@ -483,19 +848,32 @@ static unsigned long lguest_apic_read(unsigned long reg) } #endif +/* STOP! Until an interrupt comes in. */ static void lguest_safe_halt(void) { hcall(LHCALL_HALT, 0, 0, 0); } +/* Perhaps CRASH isn't the best name for this hypercall, but we use it to get a + * message out when we're crashing as well as elegant termination like powering + * off. + * + * Note that the Host always prefers that the Guest speak in physical addresses + * rather than virtual addresses, so we use __pa() here. */ static void lguest_power_off(void) { hcall(LHCALL_CRASH, __pa("Power down"), 0, 0); } +/* + * Panicing. + * + * Don't. But if you did, this is what happens. + */ static int lguest_panic(struct notifier_block *nb, unsigned long l, void *p) { hcall(LHCALL_CRASH, __pa(p), 0, 0); + /* The hcall won't return, but to keep gcc happy, we're "done". */ return NOTIFY_DONE; } @@ -503,15 +881,45 @@ static struct notifier_block paniced = { .notifier_call = lguest_panic }; +/* Setting up memory is fairly easy. */ static __init char *lguest_memory_setup(void) { - /* We do this here because lockcheck barfs if before start_kernel */ + /* We do this here and not earlier because lockcheck barfs if we do it + * before start_kernel() */ atomic_notifier_chain_register(&panic_notifier_list, &paniced); + /* The Linux bootloader header contains an "e820" memory map: the + * Launcher populated the first entry with our memory limit. */ add_memory_region(E820_MAP->addr, E820_MAP->size, E820_MAP->type); + + /* This string is for the boot messages. */ return "LGUEST"; } +/*G:050 + * Patching (Powerfully Placating Performance Pedants) + * + * We have already seen that "struct paravirt_ops" lets us replace simple + * native instructions with calls to the appropriate back end all throughout + * the kernel. This allows the same kernel to run as a Guest and as a native + * kernel, but it's slow because of all the indirect branches. + * + * Remember that David Wheeler quote about "Any problem in computer science can + * be solved with another layer of indirection"? The rest of that quote is + * "... But that usually will create another problem." This is the first of + * those problems. + * + * Our current solution is to allow the paravirt back end to optionally patch + * over the indirect calls to replace them with something more efficient. We + * patch the four most commonly called functions: disable interrupts, enable + * interrupts, restore interrupts and save interrupts. We usually have 10 + * bytes to patch into: the Guest versions of these operations are small enough + * that we can fit comfortably. + * + * First we need assembly templates of each of the patchable Guest operations, + * and these are in lguest_asm.S. */ + +/*G:060 We construct a table from the assembler templates: */ static const struct lguest_insns { const char *start, *end; @@ -521,35 +929,52 @@ static const struct lguest_insns [PARAVIRT_PATCH(restore_fl)] = { lgstart_popf, lgend_popf }, [PARAVIRT_PATCH(save_fl)] = { lgstart_pushf, lgend_pushf }, }; + +/* Now our patch routine is fairly simple (based on the native one in + * paravirt.c). If we have a replacement, we copy it in and return how much of + * the available space we used. */ static unsigned lguest_patch(u8 type, u16 clobber, void *insns, unsigned len) { unsigned int insn_len; - /* Don't touch it if we don't have a replacement */ + /* Don't do anything special if we don't have a replacement */ if (type >= ARRAY_SIZE(lguest_insns) || !lguest_insns[type].start) return paravirt_patch_default(type, clobber, insns, len); insn_len = lguest_insns[type].end - lguest_insns[type].start; - /* Similarly if we can't fit replacement. */ + /* Similarly if we can't fit replacement (shouldn't happen, but let's + * be thorough). */ if (len < insn_len) return paravirt_patch_default(type, clobber, insns, len); + /* Copy in our instructions. */ memcpy(insns, lguest_insns[type].start, insn_len); return insn_len; } +/*G:030 Once we get to lguest_init(), we know we're a Guest. The paravirt_ops + * structure in the kernel provides a single point for (almost) every routine + * we have to override to avoid privileged instructions. */ __init void lguest_init(void *boot) { - /* Copy boot parameters first. */ + /* Copy boot parameters first: the Launcher put the physical location + * in %esi, and head.S converted that to a virtual address and handed + * it to us. */ memcpy(&boot_params, boot, PARAM_SIZE); + /* The boot parameters also tell us where the command-line is: save + * that, too. */ memcpy(boot_command_line, __va(boot_params.hdr.cmd_line_ptr), COMMAND_LINE_SIZE); + /* We're under lguest, paravirt is enabled, and we're running at + * privilege level 1, not 0 as normal. */ paravirt_ops.name = "lguest"; paravirt_ops.paravirt_enabled = 1; paravirt_ops.kernel_rpl = 1; + /* We set up all the lguest overrides for sensitive operations. These + * are detailed with the operations themselves. */ paravirt_ops.save_fl = save_fl; paravirt_ops.restore_fl = restore_fl; paravirt_ops.irq_disable = irq_disable; @@ -592,21 +1017,50 @@ __init void lguest_init(void *boot) paravirt_ops.time_init = lguest_time_init; paravirt_ops.set_lazy_mode = lguest_lazy_mode; paravirt_ops.wbinvd = lguest_wbinvd; - paravirt_ops.sched_clock = lguest_sched_clock; - + /* Now is a good time to look at the implementations of these functions + * before returning to the rest of lguest_init(). */ + + /*G:070 Now we've seen all the paravirt_ops, we return to + * lguest_init() where the rest of the fairly chaotic boot setup + * occurs. + * + * The Host expects our first hypercall to tell it where our "struct + * lguest_data" is, so we do that first. */ hcall(LHCALL_LGUEST_INIT, __pa(&lguest_data), 0, 0); - /* We use top of mem for initial pagetables. */ + /* The native boot code sets up initial page tables immediately after + * the kernel itself, and sets init_pg_tables_end so they're not + * clobbered. The Launcher places our initial pagetables somewhere at + * the top of our physical memory, so we don't need extra space: set + * init_pg_tables_end to the end of the kernel. */ init_pg_tables_end = __pa(pg0); + /* Load the %fs segment register (the per-cpu segment register) with + * the normal data segment to get through booting. */ asm volatile ("mov %0, %%fs" : : "r" (__KERNEL_DS) : "memory"); + /* Clear the part of the kernel data which is expected to be zero. + * Normally it will be anyway, but if we're loading from a bzImage with + * CONFIG_RELOCATALE=y, the relocations will be sitting here. */ + memset(__bss_start, 0, __bss_stop - __bss_start); + + /* The Host uses the top of the Guest's virtual address space for the + * Host<->Guest Switcher, and it tells us how much it needs in + * lguest_data.reserve_mem, set up on the LGUEST_INIT hypercall. */ reserve_top_address(lguest_data.reserve_mem); + /* If we don't initialize the lock dependency checker now, it crashes + * paravirt_disable_iospace. */ lockdep_init(); + /* The IDE code spends about 3 seconds probing for disks: if we reserve + * all the I/O ports up front it can't get them and so doesn't probe. + * Other device drivers are similar (but less severe). This cuts the + * kernel boot time on my machine from 4.1 seconds to 0.45 seconds. */ paravirt_disable_iospace(); + /* This is messy CPU setup stuff which the native boot code does before + * start_kernel, so we have to do, too: */ cpu_detect(&new_cpu_data); /* head.S usually sets up the first capability word, so do it here. */ new_cpu_data.x86_capability[0] = cpuid_edx(1); @@ -617,14 +1071,27 @@ __init void lguest_init(void *boot) #ifdef CONFIG_X86_MCE mce_disabled = 1; #endif - #ifdef CONFIG_ACPI acpi_disabled = 1; acpi_ht = 0; #endif + /* We set the perferred console to "hvc". This is the "hypervisor + * virtual console" driver written by the PowerPC people, which we also + * adapted for lguest's use. */ add_preferred_console("hvc", 0, NULL); + /* Last of all, we set the power management poweroff hook to point to + * the Guest routine to power off. */ pm_power_off = lguest_power_off; + + /* Now we're set up, call start_kernel() in init/main.c and we proceed + * to boot as normal. It never returns. */ start_kernel(); } +/* + * This marks the end of stage II of our journey, The Guest. + * + * It is now time for us to explore the nooks and crannies of the three Guest + * devices and complete our understanding of the Guest in "make Drivers". + */ diff --git a/drivers/lguest/lguest_asm.S b/drivers/lguest/lguest_asm.S index a3dbf22..f182c6a 100644 --- a/drivers/lguest/lguest_asm.S +++ b/drivers/lguest/lguest_asm.S @@ -4,15 +4,15 @@ #include <asm/thread_info.h> #include <asm/processor-flags.h> -/* - * This is where we begin: we have a magic signature which the launcher looks - * for. The plan is that the Linux boot protocol will be extended with a +/*G:020 This is where we begin: we have a magic signature which the launcher + * looks for. The plan is that the Linux boot protocol will be extended with a * "platform type" field which will guide us here from the normal entry point, - * but for the moment this suffices. We pass the virtual address of the boot - * info to lguest_init(). + * but for the moment this suffices. The normal boot code uses %esi for the + * boot header, so we do too. We convert it to a virtual address by adding + * PAGE_OFFSET, and hand it to lguest_init() as its argument (ie. %eax). * - * We put it in .init.text will be discarded after boot. - */ + * The .section line puts this code in .init.text so it will be discarded after + * boot. */ .section .init.text, "ax", @progbits .ascii "GenuineLguest" /* Set up initial stack. */ @@ -21,7 +21,9 @@ addl $__PAGE_OFFSET, %eax jmp lguest_init -/* The templates for inline patching. */ +/*G:055 We create a macro which puts the assembler code between lgstart_ and + * lgend_ markers. These templates end up in the .init.text section, so they + * are discarded after boot. */ #define LGUEST_PATCH(name, insns...) \ lgstart_##name: insns; lgend_##name:; \ .globl lgstart_##name; .globl lgend_##name @@ -30,24 +32,61 @@ LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled) LGUEST_PATCH(sti, movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled) LGUEST_PATCH(popf, movl %eax, lguest_data+LGUEST_DATA_irq_enabled) LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax) +/*:*/ .text /* These demark the EIP range where host should never deliver interrupts. */ .global lguest_noirq_start .global lguest_noirq_end -/* - * We move eflags word to lguest_data.irq_enabled to restore interrupt state. - * For page faults, gpfs and virtual interrupts, the hypervisor has saved - * eflags manually, otherwise it was delivered directly and so eflags reflects - * the real machine IF state, ie. interrupts on. Since the kernel always dies - * if it takes such a trap with interrupts disabled anyway, turning interrupts - * back on unconditionally here is OK. - */ +/*M:004 When the Host reflects a trap or injects an interrupt into the Guest, + * it sets the eflags interrupt bit on the stack based on + * lguest_data.irq_enabled, so the Guest iret logic does the right thing when + * restoring it. However, when the Host sets the Guest up for direct traps, + * such as system calls, the processor is the one to push eflags onto the + * stack, and the interrupt bit will be 1 (in reality, interrupts are always + * enabled in the Guest). + * + * This turns out to be harmless: the only trap which should happen under Linux + * with interrupts disabled is Page Fault (due to our lazy mapping of vmalloc + * regions), which has to be reflected through the Host anyway. If another + * trap *does* go off when interrupts are disabled, the Guest will panic, and + * we'll never get to this iret! :*/ + +/*G:045 There is one final paravirt_op that the Guest implements, and glancing + * at it you can see why I left it to last. It's *cool*! It's in *assembler*! + * + * The "iret" instruction is used to return from an interrupt or trap. The + * stack looks like this: + * old address + * old code segment & privilege level + * old processor flags ("eflags") + * + * The "iret" instruction pops those values off the stack and restores them all + * at once. The only problem is that eflags includes the Interrupt Flag which + * the Guest can't change: the CPU will simply ignore it when we do an "iret". + * So we have to copy eflags from the stack to lguest_data.irq_enabled before + * we do the "iret". + * + * There are two problems with this: firstly, we need to use a register to do + * the copy and secondly, the whole thing needs to be atomic. The first + * problem is easy to solve: push %eax on the stack so we can use it, and then + * restore it at the end just before the real "iret". + * + * The second is harder: copying eflags to lguest_data.irq_enabled will turn + * interrupts on before we're finished, so we could be interrupted before we + * return to userspace or wherever. Our solution to this is to surround the + * code with lguest_noirq_start: and lguest_noirq_end: labels. We tell the + * Host that it is *never* to interrupt us there, even if interrupts seem to be + * enabled. */ ENTRY(lguest_iret) pushl %eax movl 12(%esp), %eax lguest_noirq_start: + /* Note the %ss: segment prefix here. Normal data accesses use the + * "ds" segment, but that will have already been restored for whatever + * we're returning to (such as userspace): we can't trust it. The %ss: + * prefix makes sure we use the stack segment, which is still valid. */ movl %eax,%ss:lguest_data+LGUEST_DATA_irq_enabled popl %eax iret diff --git a/drivers/lguest/lguest_bus.c b/drivers/lguest/lguest_bus.c index 18d6ab2..55a7940 100644 --- a/drivers/lguest/lguest_bus.c +++ b/drivers/lguest/lguest_bus.c @@ -1,3 +1,6 @@ +/*P:050 Lguest guests use a very simple bus for devices. It's a simple array + * of device descriptors contained just above the top of normal memory. The + * lguest bus is 80% tedious boilerplate code. :*/ #include <linux/init.h> #include <linux/bootmem.h> #include <linux/lguest_bus.h> @@ -43,6 +46,10 @@ static struct device_attribute lguest_dev_attrs[] = { __ATTR_NULL }; +/*D:130 The generic bus infrastructure requires a function which says whether a + * device matches a driver. For us, it is simple: "struct lguest_driver" + * contains a "device_type" field which indicates what type of device it can + * handle, so we just cast the args and compare: */ static int lguest_dev_match(struct device *_dev, struct device_driver *_drv) { struct lguest_device *dev = container_of(_dev,struct lguest_device,dev); @@ -50,6 +57,7 @@ static int lguest_dev_match(struct device *_dev, struct device_driver *_drv) return (drv->device_type == lguest_devices[dev->index].type); } +/*:*/ struct lguest_bus { struct bus_type bus; @@ -68,11 +76,24 @@ static struct lguest_bus lguest_bus = { } }; +/*D:140 This is the callback which occurs once the bus infrastructure matches + * up a device and driver, ie. in response to add_lguest_device() calling + * device_register(), or register_lguest_driver() calling driver_register(). + * + * At the moment it's always the latter: the devices are added first, since + * scan_devices() is called from a "core_initcall", and the drivers themselves + * called later as a normal "initcall". But it would work the other way too. + * + * So now we have the happy couple, we add the status bit to indicate that we + * found a driver. If the driver truly loves the device, it will return + * happiness from its probe function (ok, perhaps this wasn't my greatest + * analogy), and we set the final "driver ok" bit so the Host sees it's all + * green. */ static int lguest_dev_probe(struct device *_dev) { int ret; - struct lguest_device *dev = container_of(_dev,struct lguest_device,dev); - struct lguest_driver *drv = container_of(dev->dev.driver, + struct lguest_device*dev = container_of(_dev,struct lguest_device,dev); + struct lguest_driver*drv = container_of(dev->dev.driver, struct lguest_driver, drv); lguest_devices[dev->index].status |= LGUEST_DEVICE_S_DRIVER; @@ -82,6 +103,10 @@ static int lguest_dev_probe(struct device *_dev) return ret; } +/* The last part of the bus infrastructure is the function lguest drivers use + * to register themselves. Firstly, we do nothing if there's no lguest bus + * (ie. this is not a Guest), otherwise we fill in the embedded generic "struct + * driver" fields and call the generic driver_register(). */ int register_lguest_driver(struct lguest_driver *drv) { if (!lguest_devices) @@ -94,12 +119,36 @@ int register_lguest_driver(struct lguest_driver *drv) return driver_register(&drv->drv); } + +/* At the moment we build all the drivers into the kernel because they're so + * simple: 8144 bytes for all three of them as I type this. And as the console + * really needs to be built in, it's actually only 3527 bytes for the network + * and block drivers. + * + * If they get complex it will make sense for them to be modularized, so we + * need to explicitly export the symbol. + * + * I don't think non-GPL modules make sense, so it's a GPL-only export. + */ EXPORT_SYMBOL_GPL(register_lguest_driver); +/*D:120 This is the core of the lguest bus: actually adding a new device. + * It's a separate function because it's neater that way, and because an + * earlier version of the code supported hotplug and unplug. They were removed + * early on because they were never used. + * + * As Andrew Tridgell says, "Untested code is buggy code". + * + * It's worth reading this carefully: we start with an index into the array of + * "struct lguest_device_desc"s indicating the device which is new: */ static void add_lguest_device(unsigned int index) { struct lguest_device *new; + /* Each "struct lguest_device_desc" has a "status" field, which the + * Guest updates as the device is probed. In the worst case, the Host + * can look at these bits to tell what part of device setup failed, + * even if the console isn't available. */ lguest_devices[index].status |= LGUEST_DEVICE_S_ACKNOWLEDGE; new = kmalloc(sizeof(struct lguest_device), GFP_KERNEL); if (!new) { @@ -108,12 +157,17 @@ static void add_lguest_device(unsigned int index) return; } + /* The "struct lguest_device" setup is pretty straight-forward example + * code. */ new->index = index; new->private = NULL; memset(&new->dev, 0, sizeof(new->dev)); new->dev.parent = &lguest_bus.dev; new->dev.bus = &lguest_bus.bus; sprintf(new->dev.bus_id, "%u", index); + + /* device_register() causes the bus infrastructure to look for a + * matching driver. */ if (device_register(&new->dev) != 0) { printk(KERN_EMERG "Cannot register lguest device %u\n", index); lguest_devices[index].status |= LGUEST_DEVICE_S_FAILED; @@ -121,6 +175,9 @@ static void add_lguest_device(unsigned int index) } } +/*D:110 scan_devices() simply iterates through the device array. The type 0 + * is reserved to mean "no device", and anything else means we have found a + * device: add it. */ static void scan_devices(void) { unsigned int i; @@ -130,12 +187,23 @@ static void scan_devices(void) add_lguest_device(i); } +/*D:100 Fairly early in boot, lguest_bus_init() is called to set up the lguest + * bus. We check that we are a Guest by checking paravirt_ops.name: there are + * other ways of checking, but this seems most obvious to me. + * + * So we can access the array of "struct lguest_device_desc"s easily, we map + * that memory and store the pointer in the global "lguest_devices". Then we + * register the bus with the core. Doing two registrations seems clunky to me, + * but it seems to be the correct sysfs incantation. + * + * Finally we call scan_devices() which adds all the devices found in the + * "struct lguest_device_desc" array. */ static int __init lguest_bus_init(void) { if (strcmp(paravirt_ops.name, "lguest") != 0) return 0; - /* Devices are in page above top of "normal" mem. */ + /* Devices are in a single page above top of "normal" mem */ lguest_devices = lguest_map(max_pfn<<PAGE_SHIFT, 1); if (bus_register(&lguest_bus.bus) != 0 @@ -145,4 +213,5 @@ static int __init lguest_bus_init(void) scan_devices(); return 0; } +/* Do this after core stuff, before devices. */ postcore_initcall(lguest_bus_init); diff --git a/drivers/lguest/lguest_user.c b/drivers/lguest/lguest_user.c index e90d7a7..80d1b58 100644 --- a/drivers/lguest/lguest_user.c +++ b/drivers/lguest/lguest_user.c @@ -1,36 +1,70 @@ -/* Userspace control of the guest, via /dev/lguest. */ +/*P:200 This contains all the /dev/lguest code, whereby the userspace launcher + * controls and communicates with the Guest. For example, the first write will + * tell us the memory size, pagetable, entry point and kernel address offset. + * A read will run the Guest until a signal is pending (-EINTR), or the Guest + * does a DMA out to the Launcher. Writes are also used to get a DMA buffer + * registered by the Guest and to send the Guest an interrupt. :*/ #include <linux/uaccess.h> #include <linux/miscdevice.h> #include <linux/fs.h> #include "lg.h" +/*L:030 setup_regs() doesn't really belong in this file, but it gives us an + * early glimpse deeper into the Host so it's worth having here. + * + * Most of the Guest's registers are left alone: we used get_zeroed_page() to + * allocate the structure, so they will be 0. */ static void setup_regs(struct lguest_regs *regs, unsigned long start) { - /* Write out stack in format lguest expects, so we can switch to it. */ + /* There are four "segment" registers which the Guest needs to boot: + * The "code segment" register (cs) refers to the kernel code segment + * __KERNEL_CS, and the "data", "extra" and "stack" segment registers + * refer to the kernel data segment __KERNEL_DS. + * + * The privilege level is packed into the lower bits. The Guest runs + * at privilege level 1 (GUEST_PL).*/ regs->ds = regs->es = regs->ss = __KERNEL_DS|GUEST_PL; regs->cs = __KERNEL_CS|GUEST_PL; - regs->eflags = 0x202; /* Interrupts enabled. */ + + /* The "eflags" register contains miscellaneous flags. Bit 1 (0x002) + * is supposed to always be "1". Bit 9 (0x200) controls whether + * interrupts are enabled. We always leave interrupts enabled while + * running the Guest. */ + regs->eflags = 0x202; + + /* The "Extended Instruction Pointer" register says where the Guest is + * running. */ regs->eip = start; - /* esi points to our boot information (physical address 0) */ + + /* %esi points to our boot information, at physical address 0, so don't + * touch it. */ } -/* + addr */ +/*L:310 To send DMA into the Guest, the Launcher needs to be able to ask for a + * DMA buffer. This is done by writing LHREQ_GETDMA and the key to + * /dev/lguest. */ static long user_get_dma(struct lguest *lg, const u32 __user *input) { unsigned long key, udma, irq; + /* Fetch the key they wrote to us. */ if (get_user(key, input) != 0) return -EFAULT; + /* Look for a free Guest DMA buffer bound to that key. */ udma = get_dma_buffer(lg, key, &irq); if (!udma) return -ENOENT; - /* We put irq number in udma->used_len. */ + /* We need to tell the Launcher what interrupt the Guest expects after + * the buffer is filled. We stash it in udma->used_len. */ lgwrite_u32(lg, udma + offsetof(struct lguest_dma, used_len), irq); + + /* The (guest-physical) address of the DMA buffer is returned from + * the write(). */ return udma; } -/* To force the Guest to stop running and return to the Launcher, the +/*L:315 To force the Guest to stop running and return to the Launcher, the * Waker sets writes LHREQ_BREAK and the value "1" to /dev/lguest. The * Launcher then writes LHREQ_BREAK and "0" to release the Waker. */ static int break_guest_out(struct lguest *lg, const u32 __user *input) @@ -54,7 +88,8 @@ static int break_guest_out(struct lguest *lg, const u32 __user *input) } } -/* + irq */ +/*L:050 Sending an interrupt is done by writing LHREQ_IRQ and an interrupt + * number to /dev/lguest. */ static int user_send_irq(struct lguest *lg, const u32 __user *input) { u32 irq; @@ -63,14 +98,19 @@ static int user_send_irq(struct lguest *lg, const u32 __user *input) return -EFAULT; if (irq >= LGUEST_IRQS) return -EINVAL; + /* Next time the Guest runs, the core code will see if it can deliver + * this interrupt. */ set_bit(irq, lg->irqs_pending); return 0; } +/*L:040 Once our Guest is initialized, the Launcher makes it run by reading + * from /dev/lguest. */ static ssize_t read(struct file *file, char __user *user, size_t size,loff_t*o) { struct lguest *lg = file->private_data; + /* You must write LHREQ_INITIALIZE first! */ if (!lg) return -EINVAL; @@ -78,27 +118,52 @@ static ssize_t read(struct file *file, char __user *user, size_t size,loff_t*o) if (current != lg->tsk) return -EPERM; + /* If the guest is already dead, we indicate why */ if (lg->dead) { size_t len; + /* lg->dead either contains an error code, or a string. */ if (IS_ERR(lg->dead)) return PTR_ERR(lg->dead); + /* We can only return as much as the buffer they read with. */ len = min(size, strlen(lg->dead)+1); if (copy_to_user(user, lg->dead, len) != 0) return -EFAULT; return len; } + /* If we returned from read() last time because the Guest sent DMA, + * clear the flag. */ if (lg->dma_is_pending) lg->dma_is_pending = 0; + /* Run the Guest until something interesting happens. */ return run_guest(lg, (unsigned long __user *)user); } -/* Take: pfnlimit, pgdir, start, pageoffset. */ +/*L:020 The initialization write supplies 4 32-bit values (in addition to the + * 32-bit LHREQ_INITIALIZE value). These are: + * + * pfnlimit: The highest (Guest-physical) page number the Guest should be + * allowed to access. The Launcher has to live in Guest memory, so it sets + * this to ensure the Guest can't reach it. + * + * pgdir: The (Guest-physical) address of the top of the initial Guest + * pagetables (which are set up by the Launcher). + * + * start: The first instruction to execute ("eip" in x86-speak). + * + * page_offset: The PAGE_OFFSET constant in the Guest kernel. We should + * probably wean the code off this, but it's a very useful constant! Any + * address above this is within the Guest kernel, and any kernel address can + * quickly converted from physical to virtual by adding PAGE_OFFSET. It's + * 0xC0000000 (3G) by default, but it's configurable at kernel build time. + */ static int initialize(struct file *file, const u32 __user *input) { + /* "struct lguest" contains everything we (the Host) know about a + * Guest. */ struct lguest *lg; int err, i; u32 args[4]; @@ -106,7 +171,7 @@ static int initialize(struct file *file, const u32 __user *input) /* We grab the Big Lguest lock, which protects the global array * "lguests" and multiple simultaneous initializations. */ mutex_lock(&lguest_lock); - + /* You can't initialize twice! Close the device and start again... */ if (file->private_data) { err = -EBUSY; goto unlock; @@ -117,37 +182,70 @@ static int initialize(struct file *file, const u32 __user *input) goto unlock; } + /* Find an unused guest. */ i = find_free_guest(); if (i < 0) { err = -ENOSPC; goto unlock; } + /* OK, we have an index into the "lguest" array: "lg" is a convenient + * pointer. */ lg = &lguests[i]; + + /* Populate the easy fields of our "struct lguest" */ lg->guestid = i; lg->pfn_limit = args[0]; lg->page_offset = args[3]; + + /* We need a complete page for the Guest registers: they are accessible + * to the Guest and we can only grant it access to whole pages. */ lg->regs_page = get_zeroed_page(GFP_KERNEL); if (!lg->regs_page) { err = -ENOMEM; goto release_guest; } + /* We actually put the registers at the bottom of the page. */ lg->regs = (void *)lg->regs_page + PAGE_SIZE - sizeof(*lg->regs); + /* Initialize the Guest's shadow page tables, using the toplevel + * address the Launcher gave us. This allocates memory, so can + * fail. */ err = init_guest_pagetable(lg, args[1]); if (err) goto free_regs; + /* Now we initialize the Guest's registers, handing it the start + * address. */ setup_regs(lg->regs, args[2]); + + /* There are a couple of GDT entries the Guest expects when first + * booting. */ setup_guest_gdt(lg); + + /* The timer for lguest's clock needs initialization. */ init_clockdev(lg); + + /* We keep a pointer to the Launcher task (ie. current task) for when + * other Guests want to wake this one (inter-Guest I/O). */ lg->tsk = current; + /* We need to keep a pointer to the Launcher's memory map, because if + * the Launcher dies we need to clean it up. If we don't keep a + * reference, it is destroyed before close() is called. */ lg->mm = get_task_mm(lg->tsk); + + /* Initialize the queue for the waker to wait on */ init_waitqueue_head(&lg->break_wq); + + /* We remember which CPU's pages this Guest used last, for optimization + * when the same Guest runs on the same CPU twice. */ lg->last_pages = NULL; + + /* We keep our "struct lguest" in the file's private_data. */ file->private_data = lg; mutex_unlock(&lguest_lock); + /* And because this is a write() call, we return the length used. */ return sizeof(args); free_regs: @@ -159,9 +257,15 @@ unlock: return err; } +/*L:010 The first operation the Launcher does must be a write. All writes + * start with a 32 bit number: for the first write this must be + * LHREQ_INITIALIZE to set up the Guest. After that the Launcher can use + * writes of other values to get DMA buffers and send interrupts. */ static ssize_t write(struct file *file, const char __user *input, size_t size, loff_t *off) { + /* Once the guest is initialized, we hold the "struct lguest" in the + * file private data. */ struct lguest *lg = file->private_data; u32 req; @@ -169,8 +273,11 @@ static ssize_t write(struct file *file, const char __user *input, return -EFAULT; input += sizeof(req); + /* If you haven't initialized, you must do that first. */ if (req != LHREQ_INITIALIZE && !lg) return -EINVAL; + + /* Once the Guest is dead, all you can do is read() why it died. */ if (lg && lg->dead) return -ENOENT; @@ -192,33 +299,72 @@ static ssize_t write(struct file *file, const char __user *input, } } +/*L:060 The final piece of interface code is the close() routine. It reverses + * everything done in initialize(). This is usually called because the + * Launcher exited. + * + * Note that the close routine returns 0 or a negative error number: it can't + * really fail, but it can whine. I blame Sun for this wart, and K&R C for + * letting them do it. :*/ static int close(struct inode *inode, struct file *file) { struct lguest *lg = file->private_data; + /* If we never successfully initialized, there's nothing to clean up */ if (!lg) return 0; + /* We need the big lock, to protect from inter-guest I/O and other + * Launchers initializing guests. */ mutex_lock(&lguest_lock); /* Cancels the hrtimer set via LHCALL_SET_CLOCKEVENT. */ hrtimer_cancel(&lg->hrt); + /* Free any DMA buffers the Guest had bound. */ release_all_dma(lg); + /* Free up the shadow page tables for the Guest. */ free_guest_pagetable(lg); + /* Now all the memory cleanups are done, it's safe to release the + * Launcher's memory management structure. */ mmput(lg->mm); + /* If lg->dead doesn't contain an error code it will be NULL or a + * kmalloc()ed string, either of which is ok to hand to kfree(). */ if (!IS_ERR(lg->dead)) kfree(lg->dead); + /* We can free up the register page we allocated. */ free_page(lg->regs_page); + /* We clear the entire structure, which also marks it as free for the + * next user. */ memset(lg, 0, sizeof(*lg)); + /* Release lock and exit. */ mutex_unlock(&lguest_lock); + return 0; } +/*L:000 + * Welcome to our journey through the Launcher! + * + * The Launcher is the Host userspace program which sets up, runs and services + * the Guest. In fact, many comments in the Drivers which refer to "the Host" + * doing things are inaccurate: the Launcher does all the device handling for + * the Guest. The Guest can't tell what's done by the the Launcher and what by + * the Host. + * + * Just to confuse you: to the Host kernel, the Launcher *is* the Guest and we + * shall see more of that later. + * + * We begin our understanding with the Host kernel interface which the Launcher + * uses: reading and writing a character device called /dev/lguest. All the + * work happens in the read(), write() and close() routines: */ static struct file_operations lguest_fops = { .owner = THIS_MODULE, .release = close, .write = write, .read = read, }; + +/* This is a textbook example of a "misc" character device. Populate a "struct + * miscdevice" and register it with misc_register(). */ static struct miscdevice lguest_dev = { .minor = MISC_DYNAMIC_MINOR, .name = "lguest", diff --git a/drivers/lguest/page_tables.c b/drivers/lguest/page_tables.c index 1b0ba09..b7a924a 100644 --- a/drivers/lguest/page_tables.c +++ b/drivers/lguest/page_tables.c @@ -1,5 +1,11 @@ -/* Shadow page table operations. - * Copyright (C) Rusty Russell IBM Corporation 2006. +/*P:700 The pagetable code, on the other hand, still shows the scars of + * previous encounters. It's functional, and as neat as it can be in the + * circumstances, but be wary, for these things are subtle and break easily. + * The Guest provides a virtual to physical mapping, but we can neither trust + * it nor use it: we verify and convert it here to point the hardware to the + * actual Guest pages when running the Guest. :*/ + +/* Copyright (C) Rusty Russell IBM Corporation 2006. * GPL v2 and any later version */ #include <linux/mm.h> #include <linux/types.h> @@ -9,38 +15,96 @@ #include <asm/tlbflush.h> #include "lg.h" +/*M:008 We hold reference to pages, which prevents them from being swapped. + * It'd be nice to have a callback in the "struct mm_struct" when Linux wants + * to swap out. If we had this, and a shrinker callback to trim PTE pages, we + * could probably consider launching Guests as non-root. :*/ + +/*H:300 + * The Page Table Code + * + * We use two-level page tables for the Guest. If you're not entirely + * comfortable with virtual addresses, physical addresses and page tables then + * I recommend you review lguest.c's "Page Table Handling" (with diagrams!). + * + * The Guest keeps page tables, but we maintain the actual ones here: these are + * called "shadow" page tables. Which is a very Guest-centric name: these are + * the real page tables the CPU uses, although we keep them up to date to + * reflect the Guest's. (See what I mean about weird naming? Since when do + * shadows reflect anything?) + * + * Anyway, this is the most complicated part of the Host code. There are seven + * parts to this: + * (i) Setting up a page table entry for the Guest when it faults, + * (ii) Setting up the page table entry for the Guest stack, + * (iii) Setting up a page table entry when the Guest tells us it has changed, + * (iv) Switching page tables, + * (v) Flushing (thowing away) page tables, + * (vi) Mapping the Switcher when the Guest is about to run, + * (vii) Setting up the page tables initially. + :*/ + +/* Pages a 4k long, and each page table entry is 4 bytes long, giving us 1024 + * (or 2^10) entries per page. */ #define PTES_PER_PAGE_SHIFT 10 #define PTES_PER_PAGE (1 << PTES_PER_PAGE_SHIFT) + +/* 1024 entries in a page table page maps 1024 pages: 4MB. The Switcher is + * conveniently placed at the top 4MB, so it uses a separate, complete PTE + * page. */ #define SWITCHER_PGD_INDEX (PTES_PER_PAGE - 1) +/* We actually need a separate PTE page for each CPU. Remember that after the + * Switcher code itself comes two pages for each CPU, and we don't want this + * CPU's guest to see the pages of any other CPU. */ static DEFINE_PER_CPU(spte_t *, switcher_pte_pages); #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu) +/*H:320 With our shadow and Guest types established, we need to deal with + * them: the page table code is curly enough to need helper functions to keep + * it clear and clean. + * + * The first helper takes a virtual address, and says which entry in the top + * level page table deals with that address. Since each top level entry deals + * with 4M, this effectively divides by 4M. */ static unsigned vaddr_to_pgd_index(unsigned long vaddr) { return vaddr >> (PAGE_SHIFT + PTES_PER_PAGE_SHIFT); } -/* These access the shadow versions (ie. the ones used by the CPU). */ +/* There are two functions which return pointers to the shadow (aka "real") + * page tables. + * + * spgd_addr() takes the virtual address and returns a pointer to the top-level + * page directory entry for that address. Since we keep track of several page + * tables, the "i" argument tells us which one we're interested in (it's + * usually the current one). */ static spgd_t *spgd_addr(struct lguest *lg, u32 i, unsigned long vaddr) { unsigned int index = vaddr_to_pgd_index(vaddr); + /* We kill any Guest trying to touch the Switcher addresses. */ if (index >= SWITCHER_PGD_INDEX) { kill_guest(lg, "attempt to access switcher pages"); index = 0; } + /* Return a pointer index'th pgd entry for the i'th page table. */ return &lg->pgdirs[i].pgdir[index]; } +/* This routine then takes the PGD entry given above, which contains the + * address of the PTE page. It then returns a pointer to the PTE entry for the + * given address. */ static spte_t *spte_addr(struct lguest *lg, spgd_t spgd, unsigned long vaddr) { spte_t *page = __va(spgd.pfn << PAGE_SHIFT); + /* You should never call this if the PGD entry wasn't valid */ BUG_ON(!(spgd.flags & _PAGE_PRESENT)); return &page[(vaddr >> PAGE_SHIFT) % PTES_PER_PAGE]; } -/* These access the guest versions. */ +/* These two functions just like the above two, except they access the Guest + * page tables. Hence they return a Guest address. */ static unsigned long gpgd_addr(struct lguest *lg, unsigned long vaddr) { unsigned int index = vaddr >> (PAGE_SHIFT + PTES_PER_PAGE_SHIFT); @@ -55,12 +119,24 @@ static unsigned long gpte_addr(struct lguest *lg, return gpage + ((vaddr>>PAGE_SHIFT) % PTES_PER_PAGE) * sizeof(gpte_t); } -/* Do a virtual -> physical mapping on a user page. */ +/*H:350 This routine takes a page number given by the Guest and converts it to + * an actual, physical page number. It can fail for several reasons: the + * virtual address might not be mapped by the Launcher, the write flag is set + * and the page is read-only, or the write flag was set and the page was + * shared so had to be copied, but we ran out of memory. + * + * This holds a reference to the page, so release_pte() is careful to + * put that back. */ static unsigned long get_pfn(unsigned long virtpfn, int write) { struct page *page; + /* This value indicates failure. */ unsigned long ret = -1UL; + /* get_user_pages() is a complex interface: it gets the "struct + * vm_area_struct" and "struct page" assocated with a range of pages. + * It also needs the task's mmap_sem held, and is not very quick. + * It returns the number of pages it got. */ down_read(¤t->mm->mmap_sem); if (get_user_pages(current, current->mm, virtpfn << PAGE_SHIFT, 1, write, 1, &page, NULL) == 1) @@ -69,28 +145,47 @@ static unsigned long get_pfn(unsigned long virtpfn, int write) return ret; } +/*H:340 Converting a Guest page table entry to a shadow (ie. real) page table + * entry can be a little tricky. The flags are (almost) the same, but the + * Guest PTE contains a virtual page number: the CPU needs the real page + * number. */ static spte_t gpte_to_spte(struct lguest *lg, gpte_t gpte, int write) { spte_t spte; unsigned long pfn; - /* We ignore the global flag. */ + /* The Guest sets the global flag, because it thinks that it is using + * PGE. We only told it to use PGE so it would tell us whether it was + * flushing a kernel mapping or a userspace mapping. We don't actually + * use the global bit, so throw it away. */ spte.flags = (gpte.flags & ~_PAGE_GLOBAL); + + /* We need a temporary "unsigned long" variable to hold the answer from + * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't + * fit in spte.pfn. get_pfn() finds the real physical number of the + * page, given the virtual number. */ pfn = get_pfn(gpte.pfn, write); if (pfn == -1UL) { kill_guest(lg, "failed to get page %u", gpte.pfn); - /* Must not put_page() bogus page on cleanup. */ + /* When we destroy the Guest, we'll go through the shadow page + * tables and release_pte() them. Make sure we don't think + * this one is valid! */ spte.flags = 0; } + /* Now we assign the page number, and our shadow PTE is complete. */ spte.pfn = pfn; return spte; } +/*H:460 And to complete the chain, release_pte() looks like this: */ static void release_pte(spte_t pte) { + /* Remember that get_user_pages() took a reference to the page, in + * get_pfn()? We have to put it back now. */ if (pte.flags & _PAGE_PRESENT) put_page(pfn_to_page(pte.pfn)); } +/*:*/ static void check_gpte(struct lguest *lg, gpte_t gpte) { @@ -104,11 +199,16 @@ static void check_gpgd(struct lguest *lg, gpgd_t gpgd) kill_guest(lg, "bad page directory entry"); } -/* FIXME: We hold reference to pages, which prevents them from being - swapped. It'd be nice to have a callback when Linux wants to swap out. */ - -/* We fault pages in, which allows us to update accessed/dirty bits. - * Return true if we got page. */ +/*H:330 + * (i) Setting up a page table entry for the Guest when it faults + * + * We saw this call in run_guest(): when we see a page fault in the Guest, we + * come here. That's because we only set up the shadow page tables lazily as + * they're needed, so we get page faults all the time and quietly fix them up + * and return to the Guest without it knowing. + * + * If we fixed up the fault (ie. we mapped the address), this routine returns + * true. */ int demand_page(struct lguest *lg, unsigned long vaddr, int errcode) { gpgd_t gpgd; @@ -117,106 +217,161 @@ int demand_page(struct lguest *lg, unsigned long vaddr, int errcode) gpte_t gpte; spte_t *spte; + /* First step: get the top-level Guest page table entry. */ gpgd = mkgpgd(lgread_u32(lg, gpgd_addr(lg, vaddr))); + /* Toplevel not present? We can't map it in. */ if (!(gpgd.flags & _PAGE_PRESENT)) return 0; + /* Now look at the matching shadow entry. */ spgd = spgd_addr(lg, lg->pgdidx, vaddr); if (!(spgd->flags & _PAGE_PRESENT)) { - /* Get a page of PTEs for them. */ + /* No shadow entry: allocate a new shadow PTE page. */ unsigned long ptepage = get_zeroed_page(GFP_KERNEL); - /* FIXME: Steal from self in this case? */ + /* This is not really the Guest's fault, but killing it is + * simple for this corner case. */ if (!ptepage) { kill_guest(lg, "out of memory allocating pte page"); return 0; } + /* We check that the Guest pgd is OK. */ check_gpgd(lg, gpgd); + /* And we copy the flags to the shadow PGD entry. The page + * number in the shadow PGD is the page we just allocated. */ spgd->raw.val = (__pa(ptepage) | gpgd.flags); } + /* OK, now we look at the lower level in the Guest page table: keep its + * address, because we might update it later. */ gpte_ptr = gpte_addr(lg, gpgd, vaddr); gpte = mkgpte(lgread_u32(lg, gpte_ptr)); - /* No page? */ + /* If this page isn't in the Guest page tables, we can't page it in. */ if (!(gpte.flags & _PAGE_PRESENT)) return 0; - /* Write to read-only page? */ + /* Check they're not trying to write to a page the Guest wants + * read-only (bit 2 of errcode == write). */ if ((errcode & 2) && !(gpte.flags & _PAGE_RW)) return 0; - /* User access to a non-user page? */ + /* User access to a kernel page? (bit 3 == user access) */ if ((errcode & 4) && !(gpte.flags & _PAGE_USER)) return 0; + /* Check that the Guest PTE flags are OK, and the page number is below + * the pfn_limit (ie. not mapping the Launcher binary). */ check_gpte(lg, gpte); + /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */ gpte.flags |= _PAGE_ACCESSED; if (errcode & 2) gpte.flags |= _PAGE_DIRTY; - /* We're done with the old pte. */ + /* Get the pointer to the shadow PTE entry we're going to set. */ spte = spte_addr(lg, *spgd, vaddr); + /* If there was a valid shadow PTE entry here before, we release it. + * This can happen with a write to a previously read-only entry. */ release_pte(*spte); - /* We don't make it writable if this isn't a write: later - * write will fault so we can set dirty bit in guest. */ + /* If this is a write, we insist that the Guest page is writable (the + * final arg to gpte_to_spte()). */ if (gpte.flags & _PAGE_DIRTY) *spte = gpte_to_spte(lg, gpte, 1); else { + /* If this is a read, don't set the "writable" bit in the page + * table entry, even if the Guest says it's writable. That way + * we come back here when a write does actually ocur, so we can + * update the Guest's _PAGE_DIRTY flag. */ gpte_t ro_gpte = gpte; ro_gpte.flags &= ~_PAGE_RW; *spte = gpte_to_spte(lg, ro_gpte, 0); } - /* Now we update dirty/accessed on guest. */ + /* Finally, we write the Guest PTE entry back: we've set the + * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */ lgwrite_u32(lg, gpte_ptr, gpte.raw.val); + + /* We succeeded in mapping the page! */ return 1; } -/* This is much faster than the full demand_page logic. */ +/*H:360 (ii) Setting up the page table entry for the Guest stack. + * + * Remember pin_stack_pages() which makes sure the stack is mapped? It could + * simply call demand_page(), but as we've seen that logic is quite long, and + * usually the stack pages are already mapped anyway, so it's not required. + * + * This is a quick version which answers the question: is this virtual address + * mapped by the shadow page tables, and is it writable? */ static int page_writable(struct lguest *lg, unsigned long vaddr) { spgd_t *spgd; unsigned long flags; + /* Look at the top level entry: is it present? */ spgd = spgd_addr(lg, lg->pgdidx, vaddr); if (!(spgd->flags & _PAGE_PRESENT)) return 0; + /* Check the flags on the pte entry itself: it must be present and + * writable. */ flags = spte_addr(lg, *spgd, vaddr)->flags; return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW); } +/* So, when pin_stack_pages() asks us to pin a page, we check if it's already + * in the page tables, and if not, we call demand_page() with error code 2 + * (meaning "write"). */ void pin_page(struct lguest *lg, unsigned long vaddr) { if (!page_writable(lg, vaddr) && !demand_page(lg, vaddr, 2)) kill_guest(lg, "bad stack page %#lx", vaddr); } +/*H:450 If we chase down the release_pgd() code, it looks like this: */ static void release_pgd(struct lguest *lg, spgd_t *spgd) { + /* If the entry's not present, there's nothing to release. */ if (spgd->flags & _PAGE_PRESENT) { unsigned int i; + /* Converting the pfn to find the actual PTE page is easy: turn + * the page number into a physical address, then convert to a + * virtual address (easy for kernel pages like this one). */ spte_t *ptepage = __va(spgd->pfn << PAGE_SHIFT); + /* For each entry in the page, we might need to release it. */ for (i = 0; i < PTES_PER_PAGE; i++) release_pte(ptepage[i]); + /* Now we can free the page of PTEs */ free_page((long)ptepage); + /* And zero out the PGD entry we we never release it twice. */ spgd->raw.val = 0; } } +/*H:440 (v) Flushing (thowing away) page tables, + * + * We saw flush_user_mappings() called when we re-used a top-level pgdir page. + * It simply releases every PTE page from 0 up to the kernel address. */ static void flush_user_mappings(struct lguest *lg, int idx) { unsigned int i; + /* Release every pgd entry up to the kernel's address. */ for (i = 0; i < vaddr_to_pgd_index(lg->page_offset); i++) release_pgd(lg, lg->pgdirs[idx].pgdir + i); } +/* The Guest also has a hypercall to do this manually: it's used when a large + * number of mappings have been changed. */ void guest_pagetable_flush_user(struct lguest *lg) { + /* Drop the userspace part of the current page table. */ flush_user_mappings(lg, lg->pgdidx); } +/*:*/ +/* We keep several page tables. This is a simple routine to find the page + * table (if any) corresponding to this top-level address the Guest has given + * us. */ static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) { unsigned int i; @@ -226,21 +381,30 @@ static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) return i; } +/*H:435 And this is us, creating the new page directory. If we really do + * allocate a new one (and so the kernel parts are not there), we set + * blank_pgdir. */ static unsigned int new_pgdir(struct lguest *lg, unsigned long cr3, int *blank_pgdir) { unsigned int next; + /* We pick one entry at random to throw out. Choosing the Least + * Recently Used might be better, but this is easy. */ next = random32() % ARRAY_SIZE(lg->pgdirs); + /* If it's never been allocated at all before, try now. */ if (!lg->pgdirs[next].pgdir) { lg->pgdirs[next].pgdir = (spgd_t *)get_zeroed_page(GFP_KERNEL); + /* If the allocation fails, just keep using the one we have */ if (!lg->pgdirs[next].pgdir) next = lg->pgdidx; else - /* There are no mappings: you'll need to re-pin */ + /* This is a blank page, so there are no kernel + * mappings: caller must map the stack! */ *blank_pgdir = 1; } + /* Record which Guest toplevel this shadows. */ lg->pgdirs[next].cr3 = cr3; /* Release all the non-kernel mappings. */ flush_user_mappings(lg, next); @@ -248,82 +412,161 @@ static unsigned int new_pgdir(struct lguest *lg, return next; } +/*H:430 (iv) Switching page tables + * + * This is what happens when the Guest changes page tables (ie. changes the + * top-level pgdir). This happens on almost every context switch. */ void guest_new_pagetable(struct lguest *lg, unsigned long pgtable) { int newpgdir, repin = 0; + /* Look to see if we have this one already. */ newpgdir = find_pgdir(lg, pgtable); + /* If not, we allocate or mug an existing one: if it's a fresh one, + * repin gets set to 1. */ if (newpgdir == ARRAY_SIZE(lg->pgdirs)) newpgdir = new_pgdir(lg, pgtable, &repin); + /* Change the current pgd index to the new one. */ lg->pgdidx = newpgdir; + /* If it was completely blank, we map in the Guest kernel stack */ if (repin) pin_stack_pages(lg); } +/*H:470 Finally, a routine which throws away everything: all PGD entries in all + * the shadow page tables. This is used when we destroy the Guest. */ static void release_all_pagetables(struct lguest *lg) { unsigned int i, j; + /* Every shadow pagetable this Guest has */ for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) if (lg->pgdirs[i].pgdir) + /* Every PGD entry except the Switcher at the top */ for (j = 0; j < SWITCHER_PGD_INDEX; j++) release_pgd(lg, lg->pgdirs[i].pgdir + j); } +/* We also throw away everything when a Guest tells us it's changed a kernel + * mapping. Since kernel mappings are in every page table, it's easiest to + * throw them all away. This is amazingly slow, but thankfully rare. */ void guest_pagetable_clear_all(struct lguest *lg) { release_all_pagetables(lg); + /* We need the Guest kernel stack mapped again. */ pin_stack_pages(lg); } +/*H:420 This is the routine which actually sets the page table entry for then + * "idx"'th shadow page table. + * + * Normally, we can just throw out the old entry and replace it with 0: if they + * use it demand_page() will put the new entry in. We need to do this anyway: + * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page + * is read from, and _PAGE_DIRTY when it's written to. + * + * But Avi Kivity pointed out that most Operating Systems (Linux included) set + * these bits on PTEs immediately anyway. This is done to save the CPU from + * having to update them, but it helps us the same way: if they set + * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if + * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately. + */ static void do_set_pte(struct lguest *lg, int idx, unsigned long vaddr, gpte_t gpte) { + /* Look up the matching shadow page directot entry. */ spgd_t *spgd = spgd_addr(lg, idx, vaddr); + + /* If the top level isn't present, there's no entry to update. */ if (spgd->flags & _PAGE_PRESENT) { + /* Otherwise, we start by releasing the existing entry. */ spte_t *spte = spte_addr(lg, *spgd, vaddr); release_pte(*spte); + + /* If they're setting this entry as dirty or accessed, we might + * as well put that entry they've given us in now. This shaves + * 10% off a copy-on-write micro-benchmark. */ if (gpte.flags & (_PAGE_DIRTY | _PAGE_ACCESSED)) { check_gpte(lg, gpte); *spte = gpte_to_spte(lg, gpte, gpte.flags&_PAGE_DIRTY); } else + /* Otherwise we can demand_page() it in later. */ spte->raw.val = 0; } } +/*H:410 Updating a PTE entry is a little trickier. + * + * We keep track of several different page tables (the Guest uses one for each + * process, so it makes sense to cache at least a few). Each of these have + * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for + * all processes. So when the page table above that address changes, we update + * all the page tables, not just the current one. This is rare. + * + * The benefit is that when we have to track a new page table, we can copy keep + * all the kernel mappings. This speeds up context switch immensely. */ void guest_set_pte(struct lguest *lg, unsigned long cr3, unsigned long vaddr, gpte_t gpte) { - /* Kernel mappings must be changed on all top levels. */ + /* Kernel mappings must be changed on all top levels. Slow, but + * doesn't happen often. */ if (vaddr >= lg->page_offset) { unsigned int i; for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) if (lg->pgdirs[i].pgdir) do_set_pte(lg, i, vaddr, gpte); } else { + /* Is this page table one we have a shadow for? */ int pgdir = find_pgdir(lg, cr3); if (pgdir != ARRAY_SIZE(lg->pgdirs)) + /* If so, do the update. */ do_set_pte(lg, pgdir, vaddr, gpte); } } +/*H:400 + * (iii) Setting up a page table entry when the Guest tells us it has changed. + * + * Just like we did in interrupts_and_traps.c, it makes sense for us to deal + * with the other side of page tables while we're here: what happens when the + * Guest asks for a page table to be updated? + * + * We already saw that demand_page() will fill in the shadow page tables when + * needed, so we can simply remove shadow page table entries whenever the Guest + * tells us they've changed. When the Guest tries to use the new entry it will + * fault and demand_page() will fix it up. + * + * So with that in mind here's our code to to update a (top-level) PGD entry: + */ void guest_set_pmd(struct lguest *lg, unsigned long cr3, u32 idx) { int pgdir; + /* The kernel seems to try to initialize this early on: we ignore its + * attempts to map over the Switcher. */ if (idx >= SWITCHER_PGD_INDEX) return; + /* If they're talking about a page table we have a shadow for... */ pgdir = find_pgdir(lg, cr3); if (pgdir < ARRAY_SIZE(lg->pgdirs)) + /* ... throw it away. */ release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx); } +/*H:500 (vii) Setting up the page tables initially. + * + * When a Guest is first created, the Launcher tells us where the toplevel of + * its first page table is. We set some things up here: */ int init_guest_pagetable(struct lguest *lg, unsigned long pgtable) { - /* We assume this in flush_user_mappings, so check now */ + /* In flush_user_mappings() we loop from 0 to + * "vaddr_to_pgd_index(lg->page_offset)". This assumes it won't hit + * the Switcher mappings, so check that now. */ if (vaddr_to_pgd_index(lg->page_offset) >= SWITCHER_PGD_INDEX) return -EINVAL; + /* We start on the first shadow page table, and give it a blank PGD + * page. */ lg->pgdidx = 0; lg->pgdirs[lg->pgdidx].cr3 = pgtable; lg->pgdirs[lg->pgdidx].pgdir = (spgd_t*)get_zeroed_page(GFP_KERNEL); @@ -332,33 +575,48 @@ int init_guest_pagetable(struct lguest *lg, unsigned long pgtable) return 0; } +/* When a Guest dies, our cleanup is fairly simple. */ void free_guest_pagetable(struct lguest *lg) { unsigned int i; + /* Throw away all page table pages. */ release_all_pagetables(lg); + /* Now free the top levels: free_page() can handle 0 just fine. */ for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) free_page((long)lg->pgdirs[i].pgdir); } -/* Caller must be preempt-safe */ +/*H:480 (vi) Mapping the Switcher when the Guest is about to run. + * + * The Switcher and the two pages for this CPU need to be available to the + * Guest (and not the pages for other CPUs). We have the appropriate PTE pages + * for each CPU already set up, we just need to hook them in. */ void map_switcher_in_guest(struct lguest *lg, struct lguest_pages *pages) { spte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages); spgd_t switcher_pgd; spte_t regs_pte; - /* Since switcher less that 4MB, we simply mug top pte page. */ + /* Make the last PGD entry for this Guest point to the Switcher's PTE + * page for this CPU (with appropriate flags). */ switcher_pgd.pfn = __pa(switcher_pte_page) >> PAGE_SHIFT; switcher_pgd.flags = _PAGE_KERNEL; lg->pgdirs[lg->pgdidx].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd; - /* Map our regs page over stack page. */ + /* We also change the Switcher PTE page. When we're running the Guest, + * we want the Guest's "regs" page to appear where the first Switcher + * page for this CPU is. This is an optimization: when the Switcher + * saves the Guest registers, it saves them into the first page of this + * CPU's "struct lguest_pages": if we make sure the Guest's register + * page is already mapped there, we don't have to copy them out + * again. */ regs_pte.pfn = __pa(lg->regs_page) >> PAGE_SHIFT; regs_pte.flags = _PAGE_KERNEL; switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTES_PER_PAGE] = regs_pte; } +/*:*/ static void free_switcher_pte_pages(void) { @@ -368,6 +626,10 @@ static void free_switcher_pte_pages(void) free_page((long)switcher_pte_page(i)); } +/*H:520 Setting up the Switcher PTE page for given CPU is fairly easy, given + * the CPU number and the "struct page"s for the Switcher code itself. + * + * Currently the Switcher is less than a page long, so "pages" is always 1. */ static __init void populate_switcher_pte_page(unsigned int cpu, struct page *switcher_page[], unsigned int pages) @@ -375,21 +637,26 @@ static __init void populate_switcher_pte_page(unsigned int cpu, unsigned int i; spte_t *pte = switcher_pte_page(cpu); + /* The first entries are easy: they map the Switcher code. */ for (i = 0; i < pages; i++) { pte[i].pfn = page_to_pfn(switcher_page[i]); pte[i].flags = _PAGE_PRESENT|_PAGE_ACCESSED; } - /* We only map this CPU's pages, so guest can't see others. */ + /* The only other thing we map is this CPU's pair of pages. */ i = pages + cpu*2; - /* First page (regs) is rw, second (state) is ro. */ + /* First page (Guest registers) is writable from the Guest */ pte[i].pfn = page_to_pfn(switcher_page[i]); pte[i].flags = _PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW; + /* The second page contains the "struct lguest_ro_state", and is + * read-only. */ pte[i+1].pfn = page_to_pfn(switcher_page[i+1]); pte[i+1].flags = _PAGE_PRESENT|_PAGE_ACCESSED; } +/*H:510 At boot or module load time, init_pagetables() allocates and populates + * the Switcher PTE page for each CPU. */ __init int init_pagetables(struct page **switcher_page, unsigned int pages) { unsigned int i; @@ -404,7 +671,9 @@ __init int init_pagetables(struct page **switcher_page, unsigned int pages) } return 0; } +/*:*/ +/* Cleaning up simply involves freeing the PTE page for each CPU. */ void free_pagetables(void) { free_switcher_pte_pages(); diff --git a/drivers/lguest/segments.c b/drivers/lguest/segments.c index 1b2cfe8..f675a41 100644 --- a/drivers/lguest/segments.c +++ b/drivers/lguest/segments.c @@ -1,16 +1,68 @@ +/*P:600 The x86 architecture has segments, which involve a table of descriptors + * which can be used to do funky things with virtual address interpretation. + * We originally used to use segments so the Guest couldn't alter the + * Guest<->Host Switcher, and then we had to trim Guest segments, and restore + * for userspace per-thread segments, but trim again for on userspace->kernel + * transitions... This nightmarish creation was contained within this file, + * where we knew not to tread without heavy armament and a change of underwear. + * + * In these modern times, the segment handling code consists of simple sanity + * checks, and the worst you'll experience reading this code is butterfly-rash + * from frolicking through its parklike serenity. :*/ #include "lg.h" +/*H:600 + * We've almost completed the Host; there's just one file to go! + * + * Segments & The Global Descriptor Table + * + * (That title sounds like a bad Nerdcore group. Not to suggest that there are + * any good Nerdcore groups, but in high school a friend of mine had a band + * called Joe Fish and the Chips, so there are definitely worse band names). + * + * To refresh: the GDT is a table of 8-byte values describing segments. Once + * set up, these segments can be loaded into one of the 6 "segment registers". + * + * GDT entries are passed around as "struct desc_struct"s, which like IDT + * entries are split into two 32-bit members, "a" and "b". One day, someone + * will clean that up, and be declared a Hero. (No pressure, I'm just saying). + * + * Anyway, the GDT entry contains a base (the start address of the segment), a + * limit (the size of the segment - 1), and some flags. Sounds simple, and it + * would be, except those zany Intel engineers decided that it was too boring + * to put the base at one end, the limit at the other, and the flags in + * between. They decided to shotgun the bits at random throughout the 8 bytes, + * like so: + * + * 0 16 40 48 52 56 63 + * [ limit part 1 ][ base part 1 ][ flags ][li][fl][base ] + * mit ags part 2 + * part 2 + * + * As a result, this file contains a certain amount of magic numeracy. Let's + * begin. + */ + +/* Is the descriptor the Guest wants us to put in OK? + * + * The flag which Intel says must be zero: must be zero. The descriptor must + * be present, (this is actually checked earlier but is here for thorougness), + * and the descriptor type must be 1 (a memory segment). */ static int desc_ok(const struct desc_struct *gdt) { - /* MBZ=0, P=1, DT=1 */ return ((gdt->b & 0x00209000) == 0x00009000); } +/* Is the segment present? (Otherwise it can't be used by the Guest). */ static int segment_present(const struct desc_struct *gdt) { return gdt->b & 0x8000; } +/* There are several entries we don't let the Guest set. The TSS entry is the + * "Task State Segment" which controls all kinds of delicate things. The + * LGUEST_CS and LGUEST_DS entries are reserved for the Switcher, and the + * the Guest can't be trusted to deal with double faults. */ static int ignored_gdt(unsigned int num) { return (num == GDT_ENTRY_TSS @@ -19,9 +71,18 @@ static int ignored_gdt(unsigned int num) || num == GDT_ENTRY_DOUBLEFAULT_TSS); } -/* We don't allow removal of CS, DS or SS; it doesn't make sense. */ +/* If the Guest asks us to remove an entry from the GDT, we have to be careful. + * If one of the segment registers is pointing at that entry the Switcher will + * crash when it tries to reload the segment registers for the Guest. + * + * It doesn't make much sense for the Guest to try to remove its own code, data + * or stack segments while they're in use: assume that's a Guest bug. If it's + * one of the lesser segment registers using the removed entry, we simply set + * that register to 0 (unusable). */ static void check_segment_use(struct lguest *lg, unsigned int desc) { + /* GDT entries are 8 bytes long, so we divide to get the index and + * ignore the bottom bits. */ if (lg->regs->gs / 8 == desc) lg->regs->gs = 0; if (lg->regs->fs / 8 == desc) @@ -33,13 +94,21 @@ static void check_segment_use(struct lguest *lg, unsigned int desc) || lg->regs->ss / 8 == desc) kill_guest(lg, "Removed live GDT entry %u", desc); } - +/*:*/ +/*M:009 We wouldn't need to check for removal of in-use segments if we handled + * faults in the Switcher. However, it's probably not a worthwhile + * optimization. :*/ + +/*H:610 Once the GDT has been changed, we look through the changed entries and + * see if they're OK. If not, we'll call kill_guest() and the Guest will never + * get to use the invalid entries. */ static void fixup_gdt_table(struct lguest *lg, unsigned start, unsigned end) { unsigned int i; for (i = start; i < end; i++) { - /* We never copy these ones to real gdt */ + /* We never copy these ones to real GDT, so we don't care what + * they say */ if (ignored_gdt(i)) continue; @@ -53,41 +122,57 @@ static void fixup_gdt_table(struct lguest *lg, unsigned start, unsigned end) if (!desc_ok(&lg->gdt[i])) kill_guest(lg, "Bad GDT descriptor %i", i); - /* DPL 0 presumably means "for use by guest". */ + /* Segment descriptors contain a privilege level: the Guest is + * sometimes careless and leaves this as 0, even though it's + * running at privilege level 1. If so, we fix it here. */ if ((lg->gdt[i].b & 0x00006000) == 0) lg->gdt[i].b |= (GUEST_PL << 13); - /* Set accessed bit, since gdt isn't writable. */ + /* Each descriptor has an "accessed" bit. If we don't set it + * now, the CPU will try to set it when the Guest first loads + * that entry into a segment register. But the GDT isn't + * writable by the Guest, so bad things can happen. */ lg->gdt[i].b |= 0x00000100; } } +/* This routine is called at boot or modprobe time for each CPU to set up the + * "constant" GDT entries for Guests running on that CPU. */ void setup_default_gdt_entries(struct lguest_ro_state *state) { struct desc_struct *gdt = state->guest_gdt; unsigned long tss = (unsigned long)&state->guest_tss; - /* Hypervisor segments. */ + /* The hypervisor segments are full 0-4G segments, privilege level 0 */ gdt[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT; gdt[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT; - /* This is the one which we *cannot* copy from guest, since tss - is depended on this lguest_ro_state, ie. this cpu. */ + /* The TSS segment refers to the TSS entry for this CPU, so we cannot + * copy it from the Guest. Forgive the magic flags */ gdt[GDT_ENTRY_TSS].a = 0x00000067 | (tss << 16); gdt[GDT_ENTRY_TSS].b = 0x00008900 | (tss & 0xFF000000) | ((tss >> 16) & 0x000000FF); } +/* This routine is called before the Guest is run for the first time. */ void setup_guest_gdt(struct lguest *lg) { + /* Start with full 0-4G segments... */ lg->gdt[GDT_ENTRY_KERNEL_CS] = FULL_EXEC_SEGMENT; lg->gdt[GDT_ENTRY_KERNEL_DS] = FULL_SEGMENT; + /* ...except the Guest is allowed to use them, so set the privilege + * level appropriately in the flags. */ lg->gdt[GDT_ENTRY_KERNEL_CS].b |= (GUEST_PL << 13); lg->gdt[GDT_ENTRY_KERNEL_DS].b |= (GUEST_PL << 13); } -/* This is a fast version for the common case where only the three TLS entries - * have changed. */ +/* Like the IDT, we never simply use the GDT the Guest gives us. We set up the + * GDTs for each CPU, then we copy across the entries each time we want to run + * a different Guest on that CPU. */ + +/* A partial GDT load, for the three "thead-local storage" entries. Otherwise + * it's just like load_guest_gdt(). So much, in fact, it would probably be + * neater to have a single hypercall to cover both. */ void copy_gdt_tls(const struct lguest *lg, struct desc_struct *gdt) { unsigned int i; @@ -96,22 +181,31 @@ void copy_gdt_tls(const struct lguest *lg, struct desc_struct *gdt) gdt[i] = lg->gdt[i]; } +/* This is the full version */ void copy_gdt(const struct lguest *lg, struct desc_struct *gdt) { unsigned int i; + /* The default entries from setup_default_gdt_entries() are not + * replaced. See ignored_gdt() above. */ for (i = 0; i < GDT_ENTRIES; i++) if (!ignored_gdt(i)) gdt[i] = lg->gdt[i]; } +/* This is where the Guest asks us to load a new GDT (LHCALL_LOAD_GDT). */ void load_guest_gdt(struct lguest *lg, unsigned long table, u32 num) { + /* We assume the Guest has the same number of GDT entries as the + * Host, otherwise we'd have to dynamically allocate the Guest GDT. */ if (num > ARRAY_SIZE(lg->gdt)) kill_guest(lg, "too many gdt entries %i", num); + /* We read the whole thing in, then fix it up. */ lgread(lg, lg->gdt, table, num * sizeof(lg->gdt[0])); fixup_gdt_table(lg, 0, ARRAY_SIZE(lg->gdt)); + /* Mark that the GDT changed so the core knows it has to copy it again, + * even if the Guest is run on the same CPU. */ lg->changed |= CHANGED_GDT; } @@ -123,3 +217,13 @@ void guest_load_tls(struct lguest *lg, unsigned long gtls) fixup_gdt_table(lg, GDT_ENTRY_TLS_MIN, GDT_ENTRY_TLS_MAX+1); lg->changed |= CHANGED_GDT_TLS; } + +/* + * With this, we have finished the Host. + * + * Five of the seven parts of our task are complete. You have made it through + * the Bit of Despair (I think that's somewhere in the page table code, + * myself). + * + * Next, we examine "make Switcher". It's short, but intense. + */ diff --git a/drivers/lguest/switcher.S b/drivers/lguest/switcher.S index eadd4cc..d418179 100644 --- a/drivers/lguest/switcher.S +++ b/drivers/lguest/switcher.S @@ -1,45 +1,136 @@ -/* This code sits at 0xFFC00000 to do the low-level guest<->host switch. +/*P:900 This is the Switcher: code which sits at 0xFFC00000 to do the low-level + * Guest<->Host switch. It is as simple as it can be made, but it's naturally + * very specific to x86. + * + * You have now completed Preparation. If this has whet your appetite; if you + * are feeling invigorated and refreshed then the next, more challenging stage + * can be found in "make Guest". :*/ - There is are two pages above us for this CPU (struct lguest_pages). - The second page (struct lguest_ro_state) becomes read-only after the - context switch. The first page (the stack for traps) remains writable, - but while we're in here, the guest cannot be running. -*/ +/*S:100 + * Welcome to the Switcher itself! + * + * This file contains the low-level code which changes the CPU to run the Guest + * code, and returns to the Host when something happens. Understand this, and + * you understand the heart of our journey. + * + * Because this is in assembler rather than C, our tale switches from prose to + * verse. First I tried limericks: + * + * There once was an eax reg, + * To which our pointer was fed, + * It needed an add, + * Which asm-offsets.h had + * But this limerick is hurting my head. + * + * Next I tried haikus, but fitting the required reference to the seasons in + * every stanza was quickly becoming tiresome: + * + * The %eax reg + * Holds "struct lguest_pages" now: + * Cherry blossoms fall. + * + * Then I started with Heroic Verse, but the rhyming requirement leeched away + * the content density and led to some uniquely awful oblique rhymes: + * + * These constants are coming from struct offsets + * For use within the asm switcher text. + * + * Finally, I settled for something between heroic hexameter, and normal prose + * with inappropriate linebreaks. Anyway, it aint no Shakespeare. + */ + +// Not all kernel headers work from assembler +// But these ones are needed: the ENTRY() define +// And constants extracted from struct offsets +// To avoid magic numbers and breakage: +// Should they change the compiler can't save us +// Down here in the depths of assembler code. #include <linux/linkage.h> #include <asm/asm-offsets.h> #include "lg.h" +// We mark the start of the code to copy +// It's placed in .text tho it's never run here +// You'll see the trick macro at the end +// Which interleaves data and text to effect. .text ENTRY(start_switcher_text) -/* %eax points to lguest pages for this CPU. %ebx contains cr3 value. - All normal registers can be clobbered! */ +// When we reach switch_to_guest we have just left +// The safe and comforting shores of C code +// %eax has the "struct lguest_pages" to use +// Where we save state and still see it from the Guest +// And %ebx holds the Guest shadow pagetable: +// Once set we have truly left Host behind. ENTRY(switch_to_guest) - /* Save host segments on host stack. */ + // We told gcc all its regs could fade, + // Clobbered by our journey into the Guest + // We could have saved them, if we tried + // But time is our master and cycles count. + + // Segment registers must be saved for the Host + // We push them on the Host stack for later pushl %es pushl %ds pushl %gs pushl %fs - /* With CONFIG_FRAME_POINTER, gcc doesn't let us clobber this! */ + // But the compiler is fickle, and heeds + // No warning of %ebp clobbers + // When frame pointers are used. That register + // Must be saved and restored or chaos strikes. pushl %ebp - /* Save host stack. */ + // The Host's stack is done, now save it away + // In our "struct lguest_pages" at offset + // Distilled into asm-offsets.h movl %esp, LGUEST_PAGES_host_sp(%eax) - /* Switch to guest stack: if we get NMI we expect to be there. */ + + // All saved and there's now five steps before us: + // Stack, GDT, IDT, TSS + // And last of all the page tables are flipped. + + // Yet beware that our stack pointer must be + // Always valid lest an NMI hits + // %edx does the duty here as we juggle + // %eax is lguest_pages: our stack lies within. movl %eax, %edx addl $LGUEST_PAGES_regs, %edx movl %edx, %esp - /* Switch to guest's GDT, IDT. */ + + // The Guest's GDT we so carefully + // Placed in the "struct lguest_pages" before lgdt LGUEST_PAGES_guest_gdt_desc(%eax) + + // The Guest's IDT we did partially + // Move to the "struct lguest_pages" as well. lidt LGUEST_PAGES_guest_idt_desc(%eax) - /* Switch to guest's TSS while GDT still writable. */ + + // The TSS entry which controls traps + // Must be loaded up with "ltr" now: + // For after we switch over our page tables + // It (as the rest) will be writable no more. + // (The GDT entry TSS needs + // Changes type when we load it: damn Intel!) movl $(GDT_ENTRY_TSS*8), %edx ltr %dx - /* Set host's TSS GDT entry to available (clear byte 5 bit 2). */ + + // Look back now, before we take this last step! + // The Host's TSS entry was also marked used; + // Let's clear it again, ere we return. + // The GDT descriptor of the Host + // Points to the table after two "size" bytes movl (LGUEST_PAGES_host_gdt_desc+2)(%eax), %edx + // Clear the type field of "used" (byte 5, bit 2) andb $0xFD, (GDT_ENTRY_TSS*8 + 5)(%edx) - /* Switch to guest page tables: lguest_pages->state now read-only. */ + + // Once our page table's switched, the Guest is live! + // The Host fades as we run this final step. + // Our "struct lguest_pages" is now read-only. movl %ebx, %cr3 - /* Restore guest regs */ + + // The page table change did one tricky thing: + // The Guest's register page has been mapped + // Writable onto our %esp (stack) -- + // We can simply pop off all Guest regs. popl %ebx popl %ecx popl %edx @@ -51,12 +142,27 @@ ENTRY(switch_to_guest) popl %fs popl %ds popl %es - /* Skip error code and trap number */ + + // Near the base of the stack lurk two strange fields + // Which we fill as we exit the Guest + // These are the trap number and its error + // We can simply step past them on our way. addl $8, %esp + + // The last five stack slots hold return address + // And everything needed to change privilege + // Into the Guest privilege level of 1, + // And the stack where the Guest had last left it. + // Interrupts are turned back on: we are Guest. iret +// There are two paths where we switch to the Host +// So we put the routine in a macro. +// We are on our way home, back to the Host +// Interrupted out of the Guest, we come here. #define SWITCH_TO_HOST \ - /* Save guest state */ \ + /* We save the Guest state: all registers first \ + * Laid out just as "struct lguest_regs" defines */ \ pushl %es; \ pushl %ds; \ pushl %fs; \ @@ -68,58 +174,119 @@ ENTRY(switch_to_guest) pushl %edx; \ pushl %ecx; \ pushl %ebx; \ - /* Load lguest ds segment for convenience. */ \ + /* Our stack and our code are using segments \ + * Set in the TSS and IDT \ + * Yet if we were to touch data we'd use \ + * Whatever data segment the Guest had. \ + * Load the lguest ds segment for now. */ \ movl $(LGUEST_DS), %eax; \ movl %eax, %ds; \ - /* Figure out where we are, based on stack (at top of regs). */ \ + /* So where are we? Which CPU, which struct? \ + * The stack is our clue: our TSS sets \ + * It at the end of "struct lguest_pages" \ + * And we then pushed and pushed and pushed Guest regs: \ + * Now stack points atop the "struct lguest_regs". \ + * Subtract that offset, and we find our struct. */ \ movl %esp, %eax; \ subl $LGUEST_PAGES_regs, %eax; \ - /* Put trap number in %ebx before we switch cr3 and lose it. */ \ + /* Save our trap number: the switch will obscure it \ + * (The Guest regs are not mapped here in the Host) \ + * %ebx holds it safe for deliver_to_host */ \ movl LGUEST_PAGES_regs_trapnum(%eax), %ebx; \ - /* Switch to host page tables (host GDT, IDT and stack are in host \ - mem, so need this first) */ \ + /* The Host GDT, IDT and stack! \ + * All these lie safely hidden from the Guest: \ + * We must return to the Host page tables \ + * (Hence that was saved in struct lguest_pages) */ \ movl LGUEST_PAGES_host_cr3(%eax), %edx; \ movl %edx, %cr3; \ - /* Set guest's TSS to available (clear byte 5 bit 2). */ \ + /* As before, when we looked back at the Host \ + * As we left and marked TSS unused \ + * So must we now for the Guest left behind. */ \ andb $0xFD, (LGUEST_PAGES_guest_gdt+GDT_ENTRY_TSS*8+5)(%eax); \ - /* Switch to host's GDT & IDT. */ \ + /* Switch to Host's GDT, IDT. */ \ lgdt LGUEST_PAGES_host_gdt_desc(%eax); \ lidt LGUEST_PAGES_host_idt_desc(%eax); \ - /* Switch to host's stack. */ \ + /* Restore the Host's stack where it's saved regs lie */ \ movl LGUEST_PAGES_host_sp(%eax), %esp; \ - /* Switch to host's TSS */ \ + /* Last the TSS: our Host is complete */ \ movl $(GDT_ENTRY_TSS*8), %edx; \ ltr %dx; \ + /* Restore now the regs saved right at the first. */ \ popl %ebp; \ popl %fs; \ popl %gs; \ popl %ds; \ popl %es -/* Return to run_guest_once. */ +// Here's where we come when the Guest has just trapped: +// (Which trap we'll see has been pushed on the stack). +// We need only switch back, and the Host will decode +// Why we came home, and what needs to be done. return_to_host: SWITCH_TO_HOST iret +// An interrupt, with some cause external +// Has ajerked us rudely from the Guest's code +// Again we must return home to the Host deliver_to_host: SWITCH_TO_HOST - /* Decode IDT and jump to hosts' irq handler. When that does iret, it - * will return to run_guest_once. This is a feature. */ + // But now we must go home via that place + // Where that interrupt was supposed to go + // Had we not been ensconced, running the Guest. + // Here we see the cleverness of our stack: + // The Host stack is formed like an interrupt + // With EIP, CS and EFLAGS layered. + // Interrupt handlers end with "iret" + // And that will take us home at long long last. + + // But first we must find the handler to call! + // The IDT descriptor for the Host + // Has two bytes for size, and four for address: + // %edx will hold it for us for now. movl (LGUEST_PAGES_host_idt_desc+2)(%eax), %edx + // We now know the table address we need, + // And saved the trap's number inside %ebx. + // Yet the pointer to the handler is smeared + // Across the bits of the table entry. + // What oracle can tell us how to extract + // From such a convoluted encoding? + // I consulted gcc, and it gave + // These instructions, which I gladly credit: leal (%edx,%ebx,8), %eax movzwl (%eax),%edx movl 4(%eax), %eax xorw %ax, %ax orl %eax, %edx + // Now the address of the handler's in %edx + // We call it now: its "iret" takes us home. jmp *%edx -/* Real hardware interrupts are delivered straight to the host. Others - cause us to return to run_guest_once so it can decide what to do. Note - that some of these are overridden by the guest to deliver directly, and - never enter here (see load_guest_idt_entry). */ +// Every interrupt can come to us here +// But we must truly tell each apart. +// They number two hundred and fifty six +// And each must land in a different spot, +// Push its number on stack, and join the stream. + +// And worse, a mere six of the traps stand apart +// And push on their stack an addition: +// An error number, thirty two bits long +// So we punish the other two fifty +// And make them push a zero so they match. + +// Yet two fifty six entries is long +// And all will look most the same as the last +// So we create a macro which can make +// As many entries as we need to fill. + +// Note the change to .data then .text: +// We plant the address of each entry +// Into a (data) table for the Host +// To know where each Guest interrupt should go. .macro IRQ_STUB N TARGET .data; .long 1f; .text; 1: - /* Make an error number for most traps, which don't have one. */ + // Trap eight, ten through fourteen and seventeen + // Supply an error number. Else zero. .if (\N <> 8) && (\N < 10 || \N > 14) && (\N <> 17) pushl $0 .endif @@ -128,6 +295,8 @@ deliver_to_host: ALIGN .endm +// This macro creates numerous entries +// Using GAS macros which out-power C's. .macro IRQ_STUBS FIRST LAST TARGET irq=\FIRST .rept \LAST-\FIRST+1 @@ -136,24 +305,43 @@ deliver_to_host: .endr .endm -/* We intercept every interrupt, because we may need to switch back to - * host. Unfortunately we can't tell them apart except by entry - * point, so we need 256 entry points. - */ +// Here's the marker for our pointer table +// Laid in the data section just before +// Each macro places the address of code +// Forming an array: each one points to text +// Which handles interrupt in its turn. .data .global default_idt_entries default_idt_entries: .text - IRQ_STUBS 0 1 return_to_host /* First two traps */ - IRQ_STUB 2 handle_nmi /* NMI */ - IRQ_STUBS 3 31 return_to_host /* Rest of traps */ - IRQ_STUBS 32 127 deliver_to_host /* Real interrupts */ - IRQ_STUB 128 return_to_host /* System call (overridden) */ - IRQ_STUBS 129 255 deliver_to_host /* Other real interrupts */ - -/* We ignore NMI and return. */ + // The first two traps go straight back to the Host + IRQ_STUBS 0 1 return_to_host + // We'll say nothing, yet, about NMI + IRQ_STUB 2 handle_nmi + // Other traps also return to the Host + IRQ_STUBS 3 31 return_to_host + // All interrupts go via their handlers + IRQ_STUBS 32 127 deliver_to_host + // 'Cept system calls coming from userspace + // Are to go to the Guest, never the Host. + IRQ_STUB 128 return_to_host + IRQ_STUBS 129 255 deliver_to_host + +// The NMI, what a fabulous beast +// Which swoops in and stops us no matter that +// We're suspended between heaven and hell, +// (Or more likely between the Host and Guest) +// When in it comes! We are dazed and confused +// So we do the simplest thing which one can. +// Though we've pushed the trap number and zero +// We discard them, return, and hope we live. handle_nmi: addl $8, %esp iret +// We are done; all that's left is Mastery +// And "make Mastery" is a journey long +// Designed to make your fingers itch to code. + +// Here ends the text, the file and poem. ENTRY(end_switcher_text) |
