lguest: documentation VII: FIXMEs

Documentation: The FIXMEs

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

6 files changed, 55 insertions, 0 deletions

diff --git a/drivers/lguest/interrupts_and_traps.c b/drivers/lguest/interrupts_and_traps.c
index 3d98303..bd0091b 100644
--- a/drivers/lguest/interrupts_and_traps.c
+++ b/drivers/lguest/interrupts_and_traps.c
@@ -231,6 +231,20 @@ static int direct_trap(const struct lguest *lg,
 	 * 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
diff --git a/drivers/lguest/io.c b/drivers/lguest/io.c
index da28812..ea68613 100644
--- a/drivers/lguest/io.c
+++ b/drivers/lguest/io.c
@@ -553,6 +553,16 @@ void release_all_dma(struct lguest *lg)
 	up_read(&lg->mm->mmap_sem);
 }
 
+/*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,
diff --git a/drivers/lguest/lguest.c b/drivers/lguest/lguest.c
index 7e7e9fb..6dfe568 100644
--- a/drivers/lguest/lguest.c
+++ b/drivers/lguest/lguest.c
@@ -250,6 +250,14 @@ static void irq_enable(void)
 {
 	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).
diff --git a/drivers/lguest/lguest_asm.S b/drivers/lguest/lguest_asm.S
index 3126ae9..f182c6a 100644
--- a/drivers/lguest/lguest_asm.S
+++ b/drivers/lguest/lguest_asm.S
@@ -39,6 +39,20 @@ LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax)
 .global lguest_noirq_start
 .global lguest_noirq_end
 
+/*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*!
  *
diff --git a/drivers/lguest/page_tables.c b/drivers/lguest/page_tables.c
index cd047e8..b7a924a 100644
--- a/drivers/lguest/page_tables.c
+++ b/drivers/lguest/page_tables.c
@@ -15,6 +15,11 @@
 #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
  *
diff --git a/drivers/lguest/segments.c b/drivers/lguest/segments.c
index 4d4e5a4..f675a41 100644
--- a/drivers/lguest/segments.c
+++ b/drivers/lguest/segments.c
@@ -94,6 +94,10 @@ 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