%fs PDA & KVM

PATCH In review Ingo Molnar 5 participants 6 messages Mar 11, 2007 2 messages not in the archive KVM

Share & tools

Ingo Molnar

author

2007-03-11 12:24 UTC #0 lore raw

hm, i just noticed that in vmx_vcpu_run() KVM restores %fs quite late.

This is a bit awkward under v2.6.21 x86 because there we use %fs for the KERNEL_PDA area. Things like 'current' rely on it. So maybe we should just do this unconditionally in assembly:

movl $(__KERNEL_PDA), %edx;
movl %edx, %fs

it's also a problem with -rt too, which has that whole codepath up to the load_fs() call preemptible.

i'm also wondering about this bit:

        kvm_run->exit_type = 0;
        if (fail) {
                kvm_run->exit_type = KVM_EXIT_TYPE_FAIL_ENTRY;
                kvm_run->exit_reason = vmcs_read32(VM_INSTRUCTION_ERROR);
                kvm_cr3_cache_sync(vcpu);
                r = 0;
        } else {
                if (fs_gs_ldt_reload_needed) {
                        load_ldt(ldt_sel);
                        load_fs(fs_sel);

are you sure we dont need an FS reload in the 'fail' case?

Ingo

-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT & business topics through brief surveys-and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV

A message that is not in the archive (sent off-list or not yet received): [email protected]

Avi Kivity

2007-03-11 12:38 UTC #2 lore raw

Quoted 9 lines · Ingo Molnar

hm, i just noticed that in vmx_vcpu_run() KVM restores %fs quite late.
This is a bit awkward under v2.6.21 x86 because there we use %fs for the KERNEL_PDA area. Things like 'current' rely on it. So maybe we should just do this unconditionally in assembly:
movl $(__KERNEL_PDA), %edx;
movl %edx, %fs

Sure, but with %dx instead of %edx, please.

Quoted 16 lines

it's also a problem with -rt too, which has that whole codepath up to the load_fs() call preemptible.

i'm also wondering about this bit:

        kvm_run->exit_type = 0;
        if (fail) {
                kvm_run->exit_type = KVM_EXIT_TYPE_FAIL_ENTRY;
                kvm_run->exit_reason = vmcs_read32(VM_INSTRUCTION_ERROR);
                kvm_cr3_cache_sync(vcpu);
                r = 0;
        } else {
                if (fs_gs_ldt_reload_needed) {
                        load_ldt(ldt_sel);
                        load_fs(fs_sel);

Quoted 2 lines

are you sure we dont need an FS reload in the 'fail' case?

The manual is not explicit about it, but I think that an entry failure during loading of guest state _can_ cause fs to be loaded, so yes, that path is missing the segment reloads.

Non -rt kvm cannot preempt there even with CONFIG_PREEMPT=y. Can interrupts reference current? If so, that fix is needed badly for .21.

Signature · 8 lines

error compiling committee.c: too many arguments to function


-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your
opinions on IT & business topics through brief surveys-and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV

A message that is not in the archive (sent off-list or not yet received): 45F3F85D.7000509-atKUWr5tajBWk0Htik3J/[email protected]

Ingo Molnar

author

2007-03-11 12:41 UTC #4 lore raw

Quoted 13 lines · Avi Kivity

Quoted 13 lines

i'm also wondering about this bit:

       kvm_run->exit_type = 0;
       if (fail) {
               kvm_run->exit_type = KVM_EXIT_TYPE_FAIL_ENTRY;
               kvm_run->exit_reason = vmcs_read32(VM_INSTRUCTION_ERROR);
               kvm_cr3_cache_sync(vcpu);
               r = 0;
       } else {
               if (fs_gs_ldt_reload_needed) {
                       load_ldt(ldt_sel);
                       load_fs(fs_sel);

Quoted 6 lines

Quoted 2 lines
are you sure we dont need an FS reload in the 'fail' case?
The manual is not explicit about it, but I think that an entry failure during loading of guest state _can_ cause fs to be loaded, so yes, that path is missing the segment reloads.

ok, i'll send a patch.

Quoted 3 lines

Non -rt kvm cannot preempt there even with CONFIG_PREEMPT=y. Can interrupts reference current? If so, that fix is needed badly for .21.

interrupts can reference 'current' (although it's generally not encouraged) - for example scheduler_tick(), but the IRQ entry code loads %fs anyway via SAVE_ALL, so it's not an issue.

Ingo

Ingo Molnar

2007-03-11 12:52 UTC #5 lore raw

PATCH 1/1 KVM: always reload segment selectors

Subject: [patch] KVM: always reload segment selectors
From: Ingo Molnar <[email protected]>

failed VM entry on VMX might still change %fs or %gs, thus make sure that KVM always reloads the segment selectors. This is crutial on both x86 and x86_64: x86 has __KERNEL_PDA in %fs on which things like 'current' depends and x86_64 has 0 there and needs MSR_GS_BASE to work.

Signed-off-by Ingo Molnar <[email protected]>

---
 drivers/kvm/vmx.c |   37 +++++++++++++++++++++----------------
 1 file changed, 21 insertions(+), 16 deletions(-)

drivers/kvm/vmx.c +21 ~~-16~~

Index: linux/drivers/kvm/vmx.c
===================================================================
@@ -1896,6 +1896,27 @@ again:
 		[cr2]"i"(offsetof(struct kvm_vcpu, cr2))
 	      : "cc", "memory" );
 
+	/*
+	 * Reload segment selectors ASAP. (it's needed for a functional
+	 * kernel: x86 relies on having __KERNEL_PDA in %fs and x86_64
+	 * relies on having 0 in %gs for the CPU PDA to work.)
+	 */
+	if (fs_gs_ldt_reload_needed) {
+		load_ldt(ldt_sel);
+		load_fs(fs_sel);
+		/*
+		 * If we have to reload gs, we must take care to
+		 * preserve our gs base.
+		 */
+		local_irq_disable();
+		load_gs(gs_sel);
+#ifdef CONFIG_X86_64
+		wrmsrl(MSR_GS_BASE, vmcs_readl(HOST_GS_BASE));
+#endif
+		local_irq_enable();
+
+		reload_tss();
+	}
 	++kvm_stat.exits;
 
 	save_msrs(vcpu->guest_msrs, NR_BAD_MSRS);

@@ -1913,22 +1934,6 @@ again:
 		kvm_run->exit_reason = vmcs_read32(VM_INSTRUCTION_ERROR);
 		r = 0;
 	} else {
-		if (fs_gs_ldt_reload_needed) {
-			load_ldt(ldt_sel);
-			load_fs(fs_sel);
-			/*
-			 * If we have to reload gs, we must take care to
-			 * preserve our gs base.
-			 */
-			local_irq_disable();
-			load_gs(gs_sel);
-#ifdef CONFIG_X86_64
-			wrmsrl(MSR_GS_BASE, vmcs_readl(HOST_GS_BASE));
-#endif
-			local_irq_enable();
-
-			reload_tss();
-		}
 		/*
 		 * Profile KVM exit RIPs:
 		 */

Avi Kivity

2007-03-11 13:04 UTC #6 lore raw

Re: [patch] KVM: always reload segment selectors

Quoted 35 lines · Ingo Molnar

Subject: [patch] KVM: always reload segment selectors
From: Ingo Molnar <[email protected]>
failed VM entry on VMX might still change %fs or %gs, thus make sure that KVM always reloads the segment selectors. This is crutial on both x86 and x86_64: x86 has __KERNEL_PDA in %fs on which things like 'current' depends and x86_64 has 0 there and needs MSR_GS_BASE to work.
Signed-off-by Ingo Molnar <[email protected]>
---
 drivers/kvm/vmx.c |   37 +++++++++++++++++++++----------------
 1 file changed, 21 insertions(+), 16 deletions(-)
Index: linux/drivers/kvm/vmx.c
===================================================================
drivers/kvm/vmx.c +14 -0
@@ -1896,6 +1896,27 @@ again:
 		[cr2]"i"(offsetof(struct kvm_vcpu, cr2))
 	      : "cc", "memory" );
 
+	/*
+	 * Reload segment selectors ASAP. (it's needed for a functional
+	 * kernel: x86 relies on having __KERNEL_PDA in %fs and x86_64
+	 * relies on having 0 in %gs for the CPU PDA to work.)
+	 */
+	if (fs_gs_ldt_reload_needed) {
+		load_ldt(ldt_sel);
+		load_fs(fs_sel);
+		/*
+		 * If we have to reload gs, we must take care to
+		 * preserve our gs base.
+		 */
+		local_irq_disable();
+		load_gs(gs_sel);

drivers/kvm/vmx.c again: · line 1922

Quoted 10 lines

drivers/kvm/vmx.c +7 -0

+#ifdef CONFIG_X86_64
+		wrmsrl(MSR_GS_BASE, vmcs_readl(HOST_GS_BASE));
+#endif
+		local_irq_enable();
+
+		reload_tss();
+	}
 	++kvm_stat.exits;
 
 	save_msrs(vcpu->guest_msrs, NR_BAD_MSRS);

btw, looking at the code, we could just remove fs from the fs_gs_reload_needed and make in unconditional. VT knows how to reload segments, except if they're user segments (groan). In the case of fs, if it's used for the pda, it's obviously a kernel segment.

gs is different: since only the segment base is loaded (via swapgs), the selector part could well be a userspace selector, and thus the irq-protected reload is needed.

Anyway, I'm applying the patch as the above discourse is irrelevant to the fix.

Signature · 8 lines

error compiling committee.c: too many arguments to function


-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your
opinions on IT & business topics through brief surveys-and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV

Rusty Russell

2007-03-11 22:19 UTC #7 lore raw

In reply to a message not in the archive

Quoted 8 lines · Ingo Molnar

hm, i just noticed that in vmx_vcpu_run() KVM restores %fs quite late.
This is a bit awkward under v2.6.21 x86 because there we use %fs for the KERNEL_PDA area. Things like 'current' rely on it. So maybe we should just do this unconditionally in assembly:
movl $(__KERNEL_PDA), %edx;
movl %edx, %fs

Note that at some stage as Andi merges the pda->percpu conversion, this will need to be replaced by __KERNEL_PERCPU. Trivial change.

Cheers,
Rusty.