From e7a5cd063c7b4c58417f674821d63f5eb6747e37 Mon Sep 17 00:00:00 2001 From: Konrad Rzeszutek Wilk Date: Fri, 5 Apr 2013 16:42:21 -0400 Subject: x86-64, gdt: Store/load GDT for ACPI S3 or hibernate/resume path is not needed. During the ACPI S3 resume path the trampoline code handles it already. During the ACPI S3 suspend phase (acpi_suspend_lowlevel) we set: early_gdt_descr.address = (..)get_cpu_gdt_table(smp_processor_id()); which is then used during the resume path and has the same exact value as what the store/load_gdt do with the saved_context (which is saved/restored via save/restore_processor_state()). The flow during resume is complex and for 64-bit kernels we use three GDTs - one early bootstrap GDT (wakeup_igdt) that we load to workaround broken BIOSes, an early Protected Mode to Long Mode transition one (tr_gdt), and the final one - early_gdt_descr (which points to the real GDT). The early ('wakeup_gdt') is loaded in 'trampoline_start' for working around broken BIOSes, and then when we end up in Protected Mode in the startup_32 (in trampoline_64.s, not head_32.s) we use the 'tr_gdt' (still in trampoline_64.s). This 'tr_gdt' has a a 32-bit code segment, 64-bit code segment with L=1, and a 32-bit data segment. Once we have transitioned from Protected Mode to Long Mode we then set the GDT to 'early_gdt_desc' and then via an iretq emerge in wakeup_long64 (set via 'initial_code' variable in acpi_suspend_lowlevel). In the wakeup_long64 we end up restoring the %rip (which is set to 'resume_point') and jump there. In 'resume_point' we call 'restore_processor_state' which does the load_gdt on the saved context. This load_gdt is redundant as the GDT loaded via early_gdt_desc is the same. Here is the call-chain: wakeup_start |- lgdtl wakeup_gdt [the work-around broken BIOSes] | \-- trampoline_start (trampoline_64.S) |- lgdtl tr_gdt | \-- startup_32 (trampoline_64.S) | \-- startup_64 (trampoline_64.S) | \-- secondary_startup_64 |- lgdtl early_gdt_desc | ... |- movq initial_code(%rip), %eax |-.. lretq \-- wakeup_64 |-- other registers are reloaded |-- call restore_processor_state The hibernate path is much simpler. During the saving of the hibernation image we call save_processor_state() and save the contents of that along with the rest of the kernel in the hibernation image destination. We save the EIP of 'restore_registers' (restore_jump_address) and cr3 (restore_cr3). During hibernate resume, the 'restore_registers' (via the 'restore_jump_address) in hibernate_asm_64.S is invoked which restores the contents of most registers. Naturally the resume path benefits from already being in 64-bit mode, so it does not have to load the GDT. It only reloads the cr3 (from restore_cr3) and continues on. Note that the restoration of the restore image page-tables is done prior to this. After the 'restore_registers' it returns and we end up called restore_processor_state() - where we reload the GDT. The reload of the GDT is not needed as bootup kernel has already loaded the GDT which is at the same physical location as the the restored kernel. Note that the hibernation path assumes the GDT is correct during its 'restore_registers'. The assumption in the code is that the restored image is the same as saved - meaning we are not trying to restore an different kernel in the virtual address space of a new kernel. Signed-off-by: Konrad Rzeszutek Wilk Link: http://lkml.kernel.org/r/1365194544-14648-2-git-send-email-konrad.wilk@oracle.com Cc: Rafael J. Wysocki Signed-off-by: H. Peter Anvin --- arch/x86/power/cpu.c | 2 -- 1 file changed, 2 deletions(-) (limited to 'arch/x86/power/cpu.c') diff --git a/arch/x86/power/cpu.c b/arch/x86/power/cpu.c index 120cee1..6bd9423 100644 --- a/arch/x86/power/cpu.c +++ b/arch/x86/power/cpu.c @@ -65,7 +65,6 @@ static void __save_processor_state(struct saved_context *ctxt) store_idt(&ctxt->idt); #else /* CONFIG_X86_64 */ - store_gdt((struct desc_ptr *)&ctxt->gdt_limit); store_idt((struct desc_ptr *)&ctxt->idt_limit); #endif store_tr(ctxt->tr); @@ -186,7 +185,6 @@ static void __restore_processor_state(struct saved_context *ctxt) load_idt(&ctxt->idt); #else /* CONFIG_X86_64 */ - load_gdt((const struct desc_ptr *)&ctxt->gdt_limit); load_idt((const struct desc_ptr *)&ctxt->idt_limit); #endif -- cgit v1.1 From 84e70971e67d97bc2db18a4e76d42846272a54bd Mon Sep 17 00:00:00 2001 From: Konrad Rzeszutek Wilk Date: Fri, 5 Apr 2013 16:42:22 -0400 Subject: x86-32, gdt: Store/load GDT for ACPI S3 or hibernation/resume path is not needed During the ACPI S3 suspend, we store the GDT in the wakup_header (see wakeup_asm.s) field called 'pmode_gdt'. Which is then used during the resume path and has the same exact value as what the store/load_gdt do with the saved_context (which is saved/restored via save/restore_processor_state()). The flow during resume from ACPI S3 is simpler than the 64-bit counterpart. We only use the early bootstrap once (wakeup_gdt) and do various checks in real mode. After the checks are completed, we load the saved GDT ('pmode_gdt') and continue on with the resume (by heading to startup_32 in trampoline_32.S) - which quickly jumps to what was saved in 'pmode_entry' aka 'wakeup_pmode_return'. The 'wakeup_pmode_return' restores the GDT (saved_gdt) again (which was saved in do_suspend_lowlevel initially). After that it ends up calling the 'ret_point' which calls 'restore_processor_state()'. We have two opportunities to remove code where we restore the same GDT twice. Here is the call chain: wakeup_start |- lgdtl wakeup_gdt [the work-around broken BIOSes] | | - lgdtl pmode_gdt [the real one] | \-- startup_32 (in trampoline_32.S) \-- wakeup_pmode_return (in wakeup_32.S) |- lgdtl saved_gdt [the real one] \-- ret_point |.. |- call restore_processor_state The hibernate path is much simpler. During the saving of the hibernation image we call save_processor_state() and save the contents of that along with the rest of the kernel in the hibernation image destination. We save the EIP of 'restore_registers' (restore_jump_address) and cr3 (restore_cr3). During hibernate resume, the 'restore_registers' (via the 'restore_jump_address) in hibernate_asm_32.S is invoked which restores the contents of most registers. Naturally the resume path benefits from already being in 32-bit mode, so it does not have to reload the GDT. It only reloads the cr3 (from restore_cr3) and continues on. Note that the restoration of the restore image page-tables is done prior to this. After the 'restore_registers' it returns and we end up called restore_processor_state() - where we reload the GDT. The reload of the GDT is not needed as bootup kernel has already loaded the GDT which is at the same physical location as the the restored kernel. Note that the hibernation path assumes the GDT is correct during its 'restore_registers'. The assumption in the code is that the restored image is the same as saved - meaning we are not trying to restore an different kernel in the virtual address space of a new kernel. Signed-off-by: Konrad Rzeszutek Wilk Link: http://lkml.kernel.org/r/1365194544-14648-3-git-send-email-konrad.wilk@oracle.com Cc: Rafael J. Wysocki Signed-off-by: H. Peter Anvin --- arch/x86/power/cpu.c | 2 -- 1 file changed, 2 deletions(-) (limited to 'arch/x86/power/cpu.c') diff --git a/arch/x86/power/cpu.c b/arch/x86/power/cpu.c index 6bd9423..82c39c5 100644 --- a/arch/x86/power/cpu.c +++ b/arch/x86/power/cpu.c @@ -61,7 +61,6 @@ static void __save_processor_state(struct saved_context *ctxt) * descriptor tables */ #ifdef CONFIG_X86_32 - store_gdt(&ctxt->gdt); store_idt(&ctxt->idt); #else /* CONFIG_X86_64 */ @@ -181,7 +180,6 @@ static void __restore_processor_state(struct saved_context *ctxt) * ltr is done i fix_processor_context(). */ #ifdef CONFIG_X86_32 - load_gdt(&ctxt->gdt); load_idt(&ctxt->idt); #else /* CONFIG_X86_64 */ -- cgit v1.1 From 4d681be3c33dd74efffbe2a8f70634f7128602ec Mon Sep 17 00:00:00 2001 From: "konrad@kernel.org" Date: Fri, 5 Apr 2013 16:42:24 -0400 Subject: x86, wakeup, sleep: Use pvops functions for changing GDT entries We check the TSS descriptor before we try to dereference it. Also we document what the value '9' actually means using the AMD64 Architecture Programmer's Manual Volume 2, pg 90: "Hex value 9: Available 64-bit TSS" and pg 91: "The available 32-bit TSS (09h), which is redefined as the available 64-bit TSS." Without this, on Xen, where the GDT is available as R/O (to protect the hypervisor from the guest modifying it), we end up with a pagetable fault. Signed-off-by: Konrad Rzeszutek Wilk Link: http://lkml.kernel.org/r/1365194544-14648-5-git-send-email-konrad.wilk@oracle.com Cc: Rafael J. Wysocki Signed-off-by: H. Peter Anvin --- arch/x86/power/cpu.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) (limited to 'arch/x86/power/cpu.c') diff --git a/arch/x86/power/cpu.c b/arch/x86/power/cpu.c index 82c39c5..168da84 100644 --- a/arch/x86/power/cpu.c +++ b/arch/x86/power/cpu.c @@ -132,7 +132,10 @@ static void fix_processor_context(void) { int cpu = smp_processor_id(); struct tss_struct *t = &per_cpu(init_tss, cpu); - +#ifdef CONFIG_X86_64 + struct desc_struct *desc = get_cpu_gdt_table(cpu); + tss_desc tss; +#endif set_tss_desc(cpu, t); /* * This just modifies memory; should not be * necessary. But... This is necessary, because @@ -141,7 +144,9 @@ static void fix_processor_context(void) */ #ifdef CONFIG_X86_64 - get_cpu_gdt_table(cpu)[GDT_ENTRY_TSS].type = 9; + memcpy(&tss, &desc[GDT_ENTRY_TSS], sizeof(tss_desc)); + tss.type = 0x9; /* The available 64-bit TSS (see AMD vol 2, pg 91 */ + write_gdt_entry(desc, GDT_ENTRY_TSS, &tss, DESC_TSS); syscall_init(); /* This sets MSR_*STAR and related */ #endif -- cgit v1.1