xen-devel.lists.xenproject.org archive mirror
 help / color / mirror / Atom feed
From: Shuai Ruan <shuai.ruan@linux.intel.com>
To: xen-devel@lists.xen.org
Cc: andrew.cooper3@citrix.com, keir@xen.org, jbeulich@suse.com
Subject: [PATCH V6 2/2] x86/xsaves: fix overwriting between non-lazy/lazy xsaves
Date: Thu, 24 Mar 2016 16:29:33 +0800	[thread overview]
Message-ID: <1458808173-23279-3-git-send-email-shuai.ruan@linux.intel.com> (raw)
In-Reply-To: <1458808173-23279-1-git-send-email-shuai.ruan@linux.intel.com>

The offset at which components xsaved by xsave[sc] are not fixed.
So when when a save with v->fpu_dirtied set is followed by one
with v->fpu_dirtied clear, non-lazy xsave[sc] may overwriting data
written by the lazy one.

The solution is when using_xsave_compact is enabled and taking xcr0_accum into
consideration, if guest has ever used XSTATE_LAZY & ~XSTATE_FP_SSE
(XSTATE_FP_SSE will be excluded beacause xsave will write XSTATE_FP_SSE
part in legacy region of xsave area which is fixed, saving XSTATE_FS_SSE
will not cause overwriting problem), vcpu_xsave_mask will return XSTATE_ALL.
Otherwise vcpu_xsave_mask will return XSTATE_NONLAZY.

This may cause overhead save on lazy states which will cause performance
impact. After doing some performance tests on xsavec and xsaveopt
(suggested by jan), the results show xsaveopt performs better than xsavec.
So hypervisor will not use xsavec anymore.

xsaves will be used until supervised state is instroduced in hypervisor.
And XSTATE_XSAVES_ONLY (indicates supervised state is understood in xen)
is instroduced, the use of xsaves depend on whether XSTATE_XSAVES_ONLY
is set in xcr0_accum.

Signed-off-by: Shuai Ruan <shuai.ruan@linux.intel.com>
Reported-by: Jan Beulich <jbeulich@suse.com>
---
v6: Address comments from Jan
1. Do all changes in XSAVE()/XRSTOR() marco definition.
2. Add code in arch_set_info_guest()/hvm_load_cpu_ctxt()
   hvm_vcpu_reset_state() to guarante xcomp_bv is zero.

v5: Address comments from Jan
1. Add XSTATE_XSAVES_ONLY and using xsaves depend on whether this bits are
   set in xcr0_accum
2. Change compress logic in compress_xsave_states() depend on
   !(v->arch.xcr0_accum & XSTATE_XSAVES_ONLY) && !xsave_area_compressed(src)).
3. XSTATE_COMPACTION_ENABLED only set in xrstor().
4. Rebase the code on
   [V4] x86/xsaves: calculate the xstate_comp_offsets base on xstate_bv
   (already sent out) For they both change same code.
   (I am not sure whether this rebase is ok or not).

v4: Address comments from Jan
1. Add synthetic CPU feature X86_FEATURE_XSAVE_COMPACT and use this feature
   indicate whether hypervisor use compact xsave area or not.
2. Fix type/grammer errors of the comment in vcpu_xsave_mask.

v3: Address comments from Jan
1. Add xsavc clean up code and disable xsaves.
2. Add comment on why certain mask should be return in vcpu_xsave_mask.

v2: Address comments from Jan
1. add performance impact and next step to do in the description.

 xen/arch/x86/domain.c        |  8 +-----
 xen/arch/x86/domctl.c        |  6 +----
 xen/arch/x86/hvm/hvm.c       |  8 +++---
 xen/arch/x86/i387.c          | 23 ++++++++++++++----
 xen/arch/x86/xstate.c        | 58 +++++++++++++++++++++++++++-----------------
 xen/include/asm-x86/xstate.h |  1 +
 6 files changed, 60 insertions(+), 44 deletions(-)

diff --git a/xen/arch/x86/domain.c b/xen/arch/x86/domain.c
index a33f975..e456e62 100644
--- a/xen/arch/x86/domain.c
+++ b/xen/arch/x86/domain.c
@@ -947,13 +947,7 @@ int arch_set_info_guest(
         fpu_sse->fcw = FCW_DEFAULT;
         fpu_sse->mxcsr = MXCSR_DEFAULT;
     }
-    if ( cpu_has_xsaves )
-    {
-        ASSERT(v->arch.xsave_area);
-        v->arch.xsave_area->xsave_hdr.xcomp_bv = XSTATE_COMPACTION_ENABLED |
-            v->arch.xsave_area->xsave_hdr.xstate_bv;
-    }
-    else if ( v->arch.xsave_area )
+    if ( v->arch.xsave_area )
         v->arch.xsave_area->xsave_hdr.xcomp_bv = 0;
 
     if ( !compat )
diff --git a/xen/arch/x86/domctl.c b/xen/arch/x86/domctl.c
index b34a295..1a36a36 100644
--- a/xen/arch/x86/domctl.c
+++ b/xen/arch/x86/domctl.c
@@ -922,7 +922,7 @@ long arch_do_domctl(
                 ret = -EFAULT;
 
             offset += sizeof(v->arch.xcr0_accum);
-            if ( !ret && (cpu_has_xsaves || cpu_has_xsavec) )
+            if ( !ret )
             {
                 void *xsave_area;
 
@@ -942,10 +942,6 @@ long arch_do_domctl(
                      ret = -EFAULT;
                 xfree(xsave_area);
            }
-           else if ( !ret && copy_to_guest_offset(evc->buffer, offset,
-                                                  (void *)v->arch.xsave_area,
-                                                  size - 2 * sizeof(uint64_t)) )
-                ret = -EFAULT;
 
             vcpu_unpause(v);
         }
diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
index 80d59ff..a99821a 100644
--- a/xen/arch/x86/hvm/hvm.c
+++ b/xen/arch/x86/hvm/hvm.c
@@ -2202,9 +2202,8 @@ static int hvm_load_cpu_ctxt(struct domain *d, hvm_domain_context_t *h)
         xsave_area->xsave_hdr.xstate_bv = 0;
         xsave_area->fpu_sse.mxcsr = MXCSR_DEFAULT;
     }
-    if ( cpu_has_xsaves && xsave_area )
-        xsave_area->xsave_hdr.xcomp_bv = XSTATE_COMPACTION_ENABLED |
-            xsave_area->xsave_hdr.xstate_bv;
+    if ( xsave_area )
+        xsave_area->xsave_hdr.xcomp_bv = 0;
 
     v->arch.user_regs.eax = ctxt.rax;
     v->arch.user_regs.ebx = ctxt.rbx;
@@ -5581,8 +5580,7 @@ void hvm_vcpu_reset_state(struct vcpu *v, uint16_t cs, uint16_t ip)
     if ( v->arch.xsave_area )
     {
         v->arch.xsave_area->xsave_hdr.xstate_bv = XSTATE_FP;
-        v->arch.xsave_area->xsave_hdr.xcomp_bv = cpu_has_xsaves
-            ? XSTATE_COMPACTION_ENABLED | XSTATE_FP : 0;
+        v->arch.xsave_area->xsave_hdr.xcomp_bv = 0;
     }
 
     v->arch.vgc_flags = VGCF_online;
diff --git a/xen/arch/x86/i387.c b/xen/arch/x86/i387.c
index c29d0fa..2afa762 100644
--- a/xen/arch/x86/i387.c
+++ b/xen/arch/x86/i387.c
@@ -118,7 +118,24 @@ static inline uint64_t vcpu_xsave_mask(const struct vcpu *v)
     if ( v->fpu_dirtied )
         return v->arch.nonlazy_xstate_used ? XSTATE_ALL : XSTATE_LAZY;
 
-    return v->arch.nonlazy_xstate_used ? XSTATE_NONLAZY : 0;
+    ASSERT(v->arch.nonlazy_xstate_used);
+
+    /*
+     * The offsets of components which live in the extended region of
+     * compact xsave area are not fixed. Xsave area may be overwritten
+     * when a xsave with v->fpu_dirtied set is followed by one with
+     * v->fpu_dirtied clear.
+     * In such case, if hypervisor uses compact xsave area and guest
+     * has ever used lazy states (checking xcr0_accum excluding
+     * XSTATE_FP_SSE), vcpu_xsave_mask will return XSTATE_ALL. Otherwise
+     * return XSTATE_NONLAZY.
+     * XSTATE_FP_SSE may be excluded, because the offsets of XSTATE_FP_SSE
+     * (in the legacy region of xsave area) are fixed, so saving
+     * XSTATE_FP_SSE will not cause overwriting problem.
+     */
+    return (v->arch.xsave_area->xsave_hdr.xcomp_bv & XSTATE_COMPACTION_ENABLED)
+           && (v->arch.xcr0_accum & XSTATE_LAZY & ~XSTATE_FP_SSE)
+           ? XSTATE_ALL : XSTATE_NONLAZY;
 }
 
 /* Save x87 extended state */
@@ -275,11 +292,7 @@ int vcpu_init_fpu(struct vcpu *v)
         return rc;
 
     if ( v->arch.xsave_area )
-    {
         v->arch.fpu_ctxt = &v->arch.xsave_area->fpu_sse;
-        if ( cpu_has_xsaves )
-            v->arch.xsave_area->xsave_hdr.xcomp_bv = XSTATE_COMPACTION_ENABLED;
-    }
     else
     {
         BUILD_BUG_ON(__alignof(v->arch.xsave_area->fpu_sse) < 16);
diff --git a/xen/arch/x86/xstate.c b/xen/arch/x86/xstate.c
index a7b6a04..c75292c 100644
--- a/xen/arch/x86/xstate.c
+++ b/xen/arch/x86/xstate.c
@@ -179,7 +179,7 @@ void expand_xsave_states(struct vcpu *v, void *dest, unsigned int size)
     u64 xstate_bv = xsave->xsave_hdr.xstate_bv;
     u64 valid;
 
-    if ( !cpu_has_xsaves && !cpu_has_xsavec )
+    if ( !(xsave->xsave_hdr.xcomp_bv & XSTATE_COMPACTION_ENABLED) )
     {
         memcpy(dest, xsave, size);
         return;
@@ -223,13 +223,15 @@ void compress_xsave_states(struct vcpu *v, const void *src, unsigned int size)
     u64 xstate_bv = ((const struct xsave_struct *)src)->xsave_hdr.xstate_bv;
     u64 valid;
 
-    if ( !cpu_has_xsaves && !cpu_has_xsavec )
+    ASSERT(!xsave_area_compressed(src));
+
+    if ( !(v->arch.xcr0_accum & XSTATE_XSAVES_ONLY) &&
+         !xsave_area_compressed(src) )
     {
         memcpy(xsave, src, size);
         return;
     }
 
-    ASSERT(!xsave_area_compressed(src));
     /*
      * Copy legacy XSAVE area, to avoid complications with CPUID
      * leaves 0 and 1 in the loop below.
@@ -269,15 +271,16 @@ void xsave(struct vcpu *v, uint64_t mask)
     uint32_t lmask = mask;
     unsigned int fip_width = v->domain->arch.x87_fip_width;
 #define XSAVE(pfx) \
-        alternative_io_3(".byte " pfx "0x0f,0xae,0x27\n", /* xsave */ \
-                         ".byte " pfx "0x0f,0xae,0x37\n", /* xsaveopt */ \
-                         X86_FEATURE_XSAVEOPT, \
-                         ".byte " pfx "0x0f,0xc7,0x27\n", /* xsavec */ \
-                         X86_FEATURE_XSAVEC, \
-                         ".byte " pfx "0x0f,0xc7,0x2f\n", /* xsaves */ \
-                         X86_FEATURE_XSAVES, \
-                         "=m" (*ptr), \
-                         "a" (lmask), "d" (hmask), "D" (ptr))
+        if ( v->arch.xcr0_accum & XSTATE_XSAVES_ONLY ) \
+            asm volatile ( ".byte " pfx "0x0f,0xc7,0x2f\n" /* xsaves */ \
+                           : "=m" (*ptr) \
+                           : "a" (lmask), "d" (hmask), "D" (ptr) ); \
+        else \
+            alternative_io(".byte " pfx "0x0f,0xae,0x27\n", /* xsave */ \
+                           ".byte " pfx "0x0f,0xae,0x37\n", /* xsaveopt */ \
+                           X86_FEATURE_XSAVEOPT, \
+                           "=m" (*ptr), \
+                           "a" (lmask), "d" (hmask), "D" (ptr))
 
     if ( fip_width == 8 || !(mask & XSTATE_FP) )
     {
@@ -368,19 +371,29 @@ void xrstor(struct vcpu *v, uint64_t mask)
         switch ( __builtin_expect(ptr->fpu_sse.x[FPU_WORD_SIZE_OFFSET], 8) )
         {
             BUILD_BUG_ON(sizeof(faults) != 4); /* Clang doesn't support %z in asm. */
-#define XRSTOR(pfx) \
-        alternative_io("1: .byte " pfx "0x0f,0xae,0x2f\n" \
+#define _xrstor(xrstor_ins) \
+        asm volatile ( "1: .byte "xrstor_ins"\n" \
                        "3:\n" \
                        "   .section .fixup,\"ax\"\n" \
                        "2: incl %[faults]\n" \
                        "   jmp 3b\n" \
                        "   .previous\n" \
-                       _ASM_EXTABLE(1b, 2b), \
-                       ".byte " pfx "0x0f,0xc7,0x1f\n", \
-                       X86_FEATURE_XSAVES, \
-                       ASM_OUTPUT2([mem] "+m" (*ptr), [faults] "+g" (faults)), \
-                       [lmask] "a" (lmask), [hmask] "d" (hmask), \
-                       [ptr] "D" (ptr))
+                       _ASM_EXTABLE(1b, 2b) \
+                       : [mem] "+m" (*ptr), [faults] "+g" (faults) \
+                       : [lmask] "a" (lmask), [hmask] "d" (hmask), \
+                         [ptr] "D" (ptr) )
+
+#define XRSTOR(pfx) \
+        if ( v->arch.xcr0_accum & XSTATE_XSAVES_ONLY ) \
+        { \
+            if ( unlikely(!(ptr->xsave_hdr.xcomp_bv \
+                            & XSTATE_COMPACTION_ENABLED)) ) \
+                ptr->xsave_hdr.xcomp_bv |= ptr->xsave_hdr.xstate_bv \
+                                           | XSTATE_COMPACTION_ENABLED; \
+                _xrstor(pfx "0x0f,0xc7,0x1f"); /* xrstors */ \
+        } \
+        else \
+            _xrstor(pfx "0x0f,0xae,0x2f") /* xrstor */
 
         default:
             XRSTOR("0x48,");
@@ -389,6 +402,7 @@ void xrstor(struct vcpu *v, uint64_t mask)
             XRSTOR("");
             break;
 #undef XRSTOR
+#undef _xrstor
         }
         if ( likely(faults == prev_faults) )
             break;
@@ -416,7 +430,7 @@ void xrstor(struct vcpu *v, uint64_t mask)
                   ((mask & XSTATE_YMM) &&
                    !(ptr->xsave_hdr.xcomp_bv & XSTATE_COMPACTION_ENABLED))) )
                 ptr->fpu_sse.mxcsr &= mxcsr_mask;
-            if ( cpu_has_xsaves || cpu_has_xsavec )
+            if ( v->arch.xcr0_accum & XSTATE_XSAVES_ONLY )
             {
                 ptr->xsave_hdr.xcomp_bv &= this_cpu(xcr0) | this_cpu(xss);
                 ptr->xsave_hdr.xstate_bv &= ptr->xsave_hdr.xcomp_bv;
@@ -433,7 +447,7 @@ void xrstor(struct vcpu *v, uint64_t mask)
         case 2: /* Stage 2: Reset all state. */
             ptr->fpu_sse.mxcsr = MXCSR_DEFAULT;
             ptr->xsave_hdr.xstate_bv = 0;
-            ptr->xsave_hdr.xcomp_bv = cpu_has_xsaves
+            ptr->xsave_hdr.xcomp_bv = v->arch.xcr0_accum & XSTATE_XSAVES_ONLY
                                       ? XSTATE_COMPACTION_ENABLED : 0;
             continue;
         }
diff --git a/xen/include/asm-x86/xstate.h b/xen/include/asm-x86/xstate.h
index a488688..91d1c39 100644
--- a/xen/include/asm-x86/xstate.h
+++ b/xen/include/asm-x86/xstate.h
@@ -44,6 +44,7 @@
 #define XSTATE_NONLAZY (XSTATE_LWP | XSTATE_BNDREGS | XSTATE_BNDCSR | \
                         XSTATE_PKRU)
 #define XSTATE_LAZY    (XSTATE_ALL & ~XSTATE_NONLAZY)
+#define XSTATE_XSAVES_ONLY         0
 #define XSTATE_COMPACTION_ENABLED  (1ULL << 63)
 
 #define XSTATE_ALIGN64 (1U << 1)
-- 
1.9.1


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel

  parent reply	other threads:[~2016-03-24  8:29 UTC|newest]

Thread overview: 7+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2016-03-24  8:29 [PATCH V6 0/2] xsaves bug fix Shuai Ruan
2016-03-24  8:29 ` [PATCH V6 1/2] x86/xsaves: calculate the comp_offsets base on xcomp_bv Shuai Ruan
2016-03-29 14:35   ` Jan Beulich
2016-03-30  5:48     ` Shuai Ruan
2016-03-24  8:29 ` Shuai Ruan [this message]
2016-03-29 15:00   ` [PATCH V6 2/2] x86/xsaves: fix overwriting between non-lazy/lazy xsaves Jan Beulich
2016-03-30  5:46     ` Shuai Ruan

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1458808173-23279-3-git-send-email-shuai.ruan@linux.intel.com \
    --to=shuai.ruan@linux.intel.com \
    --cc=andrew.cooper3@citrix.com \
    --cc=jbeulich@suse.com \
    --cc=keir@xen.org \
    --cc=xen-devel@lists.xen.org \
    --subject='Re: [PATCH V6 2/2] x86/xsaves: fix overwriting between non-lazy/lazy xsaves' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).