Skip to contentSkip to conversation

KVM: X86: Avoid explictly fetch instruction in x86_decode_insn()

PATCH linmiaohe 4 participants 6 messages Mar 4, 2020 KVM

lore
Share & tools
lore

linmiaohe

author
#0 lore raw

Hi:

Quoted 12 lines · Peter Xu

insn_fetch() will always implicitly refill instruction buffer properly when the buffer is empty, so we don't need to explicitly fetch it even if insn_len==0 for x86_decode_insn().

---
arch/x86/kvm/emulate.c | 5 -----

1 file changed, 5 deletions(-)
arch/x86/kvm/emulate.c +0 -0
@@ -5175,11 +5175,6 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len)
 	ctxt->opcode_len = 1;

arch/x86/kvm/emulate.c int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len) · line 5180

Quoted 10 lines
arch/x86/kvm/emulate.c +0 -5
 	if (insn_len > 0)
 		memcpy(ctxt->fetch.data, insn, insn_len);
-	else {
-		rc = __do_insn_fetch_bytes(ctxt, 1);
-		if (rc != X86EMUL_CONTINUE)
-			goto done;
-	}
 
 	switch (mode) {
 	case X86EMUL_MODE_REAL:

Looks good, thanks. But it seems we should also take care of the comment in __do_insn_fetch_bytes(), as we do not load instruction at the beginning of x86_decode_insn() now, which may be misleading:
/*
* One instruction can only straddle two pages,
* and one has been loaded at the beginning of
* x86_decode_insn. So, if not enough bytes
* still, we must have hit the 15-byte boundary.
*/
if (unlikely(size < op_size))
return emulate_gp(ctxt, 0);

maintainer
#1 lore raw
Quoted 22 lines · linmiaohe

Hi:

Quoted 20 lines · Peter Xu

insn_fetch() will always implicitly refill instruction buffer properly when the buffer is empty, so we don't need to explicitly fetch it even if insn_len==0 for x86_decode_insn().

---
arch/x86/kvm/emulate.c | 5 -----

1 file changed, 5 deletions(-)
arch/x86/kvm/emulate.c +0 -5
@@ -5175,11 +5175,6 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len)
 	ctxt->opcode_len = 1;
 	if (insn_len > 0)
 		memcpy(ctxt->fetch.data, insn, insn_len);
-	else {
-		rc = __do_insn_fetch_bytes(ctxt, 1);
-		if (rc != X86EMUL_CONTINUE)
-			goto done;
-	}
 
Quoted 2 lines

arch/x86/kvm/emulate.c int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len) · line 5180

Quoted 2 lines
arch/x86/kvm/emulate.c +0 -0
 	switch (mode) {
 	case X86EMUL_MODE_REAL:

This is a a small (but measurable) optimization; going through __do_insn_fetch_bytes instead of do_insn_fetch_bytes is a little bit faster because it lets you mark the branch in do_insn_fetch_bytes as unlikely, and in general it allows the branch predictor to do a better job.

Paolo

Quoted 11 lines

Looks good, thanks. But it seems we should also take care of the comment in __do_insn_fetch_bytes(), as we do not load instruction at the beginning of x86_decode_insn() now, which may be misleading:
/*
* One instruction can only straddle two pages,
* and one has been loaded at the beginning of
* x86_decode_insn. So, if not enough bytes
* still, we must have hit the 15-byte boundary.
*/
if (unlikely(size < op_size))
return emulate_gp(ctxt, 0);

#2 lore raw
Quoted 23 lines · Paolo Bonzini
Quoted 22 lines · linmiaohe

Hi:

Quoted 20 lines · Peter Xu

insn_fetch() will always implicitly refill instruction buffer properly when the buffer is empty, so we don't need to explicitly fetch it even if insn_len==0 for x86_decode_insn().

---
arch/x86/kvm/emulate.c | 5 -----

1 file changed, 5 deletions(-)
arch/x86/kvm/emulate.c +0 -5
@@ -5175,11 +5175,6 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len)
 	ctxt->opcode_len = 1;
 	if (insn_len > 0)
 		memcpy(ctxt->fetch.data, insn, insn_len);
-	else {
-		rc = __do_insn_fetch_bytes(ctxt, 1);
-		if (rc != X86EMUL_CONTINUE)
-			goto done;
-	}
 
Quoted 7 lines
Quoted 2 lines

arch/x86/kvm/emulate.c int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len) · line 5180

Quoted 2 lines
arch/x86/kvm/emulate.c +0 -0
 	switch (mode) {
 	case X86EMUL_MODE_REAL:

This is a a small (but measurable) optimization; going through __do_insn_fetch_bytes instead of do_insn_fetch_bytes is a little bit faster because it lets you mark the branch in do_insn_fetch_bytes as unlikely, and in general it allows the branch predictor to do a better job.

Ah I see, that makes sense. Thanks!

Signature · 1 line
Peter Xu
#3 lore raw
In reply to linmiaohe
Quoted 22 lines · linmiaohe

Hi:

Quoted 20 lines · Peter Xu

insn_fetch() will always implicitly refill instruction buffer properly when the buffer is empty, so we don't need to explicitly fetch it even if insn_len==0 for x86_decode_insn().

---
arch/x86/kvm/emulate.c | 5 -----

1 file changed, 5 deletions(-)
arch/x86/kvm/emulate.c +0 -5
@@ -5175,11 +5175,6 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len)
 	ctxt->opcode_len = 1;
 	if (insn_len > 0)
 		memcpy(ctxt->fetch.data, insn, insn_len);
-	else {
-		rc = __do_insn_fetch_bytes(ctxt, 1);
-		if (rc != X86EMUL_CONTINUE)
-			goto done;
-	}
 
Quoted 13 lines

arch/x86/kvm/emulate.c int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len) · line 5180

Quoted 2 lines
arch/x86/kvm/emulate.c +0 -0
 	switch (mode) {
 	case X86EMUL_MODE_REAL:

Looks good, thanks. But it seems we should also take care of the comment in __do_insn_fetch_bytes(), as we do not load instruction at the beginning of x86_decode_insn() now, which may be misleading:
/*
* One instruction can only straddle two pages,
* and one has been loaded at the beginning of
* x86_decode_insn. So, if not enough bytes
* still, we must have hit the 15-byte boundary.
*/
if (unlikely(size < op_size))
return emulate_gp(ctxt, 0);

Right, thanks for spotting that (even if the patch to be dropped :).

I guess not only the comment, but the check might even fail if we apply the patch. Because when the fetch is the 1st attempt and unluckily that acrosses one page boundary (because we'll only fetch until either 15 bytes or the page boundary), so that single fetch could be smaller than op_size provided.

Thanks,

Signature · 1 line
Peter Xu
maintainer
#4 lore raw
Quoted 10 lines · Peter Xu
Quoted 10 lines

Looks good, thanks. But it seems we should also take care of the comment in __do_insn_fetch_bytes(), as we do not load instruction at the beginning of x86_decode_insn() now, which may be misleading:
/*
* One instruction can only straddle two pages,
* and one has been loaded at the beginning of
* x86_decode_insn. So, if not enough bytes
* still, we must have hit the 15-byte boundary.
*/
if (unlikely(size < op_size))
return emulate_gp(ctxt, 0);

Quoted 7 lines

Right, thanks for spotting that (even if the patch to be dropped :).

I guess not only the comment, but the check might even fail if we apply the patch. Because when the fetch is the 1st attempt and unluckily that acrosses one page boundary (because we'll only fetch until either 15 bytes or the page boundary), so that single fetch could be smaller than op_size provided.

Right, priming the decode cache with one byte from the current page cannot fail, and then we know that the next call must be at the beginning of the next page.

Paolo

#5 lore raw
Quoted 21 lines · Paolo Bonzini
Quoted 10 lines · Peter Xu
Quoted 10 lines

Looks good, thanks. But it seems we should also take care of the comment in __do_insn_fetch_bytes(), as we do not load instruction at the beginning of x86_decode_insn() now, which may be misleading:
/*
* One instruction can only straddle two pages,
* and one has been loaded at the beginning of
* x86_decode_insn. So, if not enough bytes
* still, we must have hit the 15-byte boundary.
*/
if (unlikely(size < op_size))
return emulate_gp(ctxt, 0);

Quoted 7 lines

Right, thanks for spotting that (even if the patch to be dropped :).

I guess not only the comment, but the check might even fail if we apply the patch. Because when the fetch is the 1st attempt and unluckily that acrosses one page boundary (because we'll only fetch until either 15 bytes or the page boundary), so that single fetch could be smaller than op_size provided.

Quoted 3 lines

Right, priming the decode cache with one byte from the current page cannot fail, and then we know that the next call must be at the beginning of the next page.

IIRC I encountered (and fixed) a similar KVM bug in the past. It is a shame I never wrote a unit test (and I don’t have time now), but it would be nice to have one.