diff mbox

[v7] x86/hvm: Implement hvmemul_write() using real mappings

Message ID 1506588748-14558-1-git-send-email-aisaila@bitdefender.com (mailing list archive)
State New, archived
Headers show

Commit Message

Alexandru Stefan ISAILA Sept. 28, 2017, 8:52 a.m. UTC
From: Andrew Cooper <andrew.cooper3@citrix.com>

An access which crosses a page boundary is performed atomically by x86
hardware, albeit with a severe performance penalty.  An important corner case
is when a straddled access hits two pages which differ in whether a
translation exists, or in net access rights.

The use of hvm_copy*() in hvmemul_write() is problematic, because it performs
a translation then completes the partial write, before moving onto the next
translation.

If an individual emulated write straddles two pages, the first of which is
writable, and the second of which is not, the first half of the write will
complete before #PF is raised from the second half.

This results in guest state corruption as a side effect of emulation, which
has been observed to cause windows to crash while under introspection.

Introduce the hvmemul_{,un}map_linear_addr() helpers, which translate an
entire contents of a linear access, and vmap() the underlying frames to
provide a contiguous virtual mapping for the emulator to use.  This is the
same mechanism as used by the shadow emulation code.

This will catch any translation issues and abort the emulation before any
modifications occur.

Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Signed-off-by: Alexandru Isaila <aisaila@bitdefender.com>

---
Changes since V6:
	- Added a direct way to calculate the nr_frames
	- The while loop uses a i counter
	- The frame address is calculated in the loop
---
 xen/arch/x86/hvm/emulate.c        | 185 ++++++++++++++++++++++++++++++++++----
 xen/include/asm-x86/hvm/emulate.h |   7 ++
 2 files changed, 176 insertions(+), 16 deletions(-)

Comments

Paul Durrant Sept. 28, 2017, 9:07 a.m. UTC | #1
> -----Original Message-----
> From: Alexandru Isaila [mailto:aisaila@bitdefender.com]
> Sent: 28 September 2017 09:52
> To: xen-devel@lists.xen.org
> Cc: jbeulich@suse.com; Andrew Cooper <Andrew.Cooper3@citrix.com>;
> Paul Durrant <Paul.Durrant@citrix.com>; Alexandru Isaila
> <aisaila@bitdefender.com>
> Subject: [PATCH v7] x86/hvm: Implement hvmemul_write() using real
> mappings
> 
> From: Andrew Cooper <andrew.cooper3@citrix.com>
> 
> An access which crosses a page boundary is performed atomically by x86
> hardware, albeit with a severe performance penalty.  An important corner
> case
> is when a straddled access hits two pages which differ in whether a
> translation exists, or in net access rights.
> 
> The use of hvm_copy*() in hvmemul_write() is problematic, because it
> performs
> a translation then completes the partial write, before moving onto the next
> translation.
> 
> If an individual emulated write straddles two pages, the first of which is
> writable, and the second of which is not, the first half of the write will
> complete before #PF is raised from the second half.
> 
> This results in guest state corruption as a side effect of emulation, which
> has been observed to cause windows to crash while under introspection.
> 
> Introduce the hvmemul_{,un}map_linear_addr() helpers, which translate an
> entire contents of a linear access, and vmap() the underlying frames to
> provide a contiguous virtual mapping for the emulator to use.  This is the
> same mechanism as used by the shadow emulation code.
> 
> This will catch any translation issues and abort the emulation before any
> modifications occur.
> 
> Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
> Signed-off-by: Alexandru Isaila <aisaila@bitdefender.com>
> 
> ---
> Changes since V6:
> 	- Added a direct way to calculate the nr_frames
> 	- The while loop uses a i counter
> 	- The frame address is calculated in the loop
> ---
>  xen/arch/x86/hvm/emulate.c        | 185
> ++++++++++++++++++++++++++++++++++----
>  xen/include/asm-x86/hvm/emulate.h |   7 ++
>  2 files changed, 176 insertions(+), 16 deletions(-)
> 
> diff --git a/xen/arch/x86/hvm/emulate.c b/xen/arch/x86/hvm/emulate.c
> index cc874ce..7d18bf5 100644
> --- a/xen/arch/x86/hvm/emulate.c
> +++ b/xen/arch/x86/hvm/emulate.c
> @@ -498,6 +498,166 @@ static int hvmemul_do_mmio_addr(paddr_t
> mmio_gpa,
>  }
> 
>  /*
> + * Map the frame(s) covering an individual linear access, for writeable
> + * access.  May return NULL for MMIO, or ERR_PTR(~X86EMUL_*) for other
> errors
> + * including ERR_PTR(~X86EMUL_OKAY) for write-discard mappings.
> + *
> + * In debug builds, map() checks that each slot in hvmemul_ctxt->mfn[] is
> + * clean before use, and poisions unused slots with INVALID_MFN.
> + */
> +static void *hvmemul_map_linear_addr(
> +    unsigned long linear, unsigned int bytes, uint32_t pfec,
> +    struct hvm_emulate_ctxt *hvmemul_ctxt)
> +{
> +    struct vcpu *curr = current;
> +    void *err, *mapping;
> +
> +    /* First and final gfns which need mapping. */
> +    unsigned long frame = linear >> PAGE_SHIFT, first = frame;
> +    unsigned long final = (linear + bytes - !!bytes) >> PAGE_SHIFT;
> +    unsigned long nr_frames = ((linear + bytes) >> PAGE_SHIFT) - (linear >>
> PAGE_SHIFT) + 1;

I'm unclear as to why you don't calculate nr_frames first and then calculate final as first + nr_frames - 1? Having completely independent calculations like this seems prone to error.

> +    int i = 0;
> +
> +    /*
> +     * mfn points to the next free slot.  All used slots have a page reference
> +     * held on them.
> +     */
> +    mfn_t *mfn = &hvmemul_ctxt->mfn[0];
> +

Extraneous blank line.

> +
> +    /*
> +     * The caller has no legitimate reason for trying a zero-byte write, but
> +     * final is calculate to fail safe in release builds.
> +     *
> +     * The maximum write size depends on the number of adjacent mfns[]
> which
> +     * can be vmap()'d, accouting for possible misalignment within the region.
> +     * The higher level emulation callers are responsible for ensuring that
> +     * mfns[] is large enough for the requested write size.
> +     */
> +    if ( bytes == 0 ||
> +         nr_frames > ARRAY_SIZE(hvmemul_ctxt->mfn) )
> +    {
> +        ASSERT_UNREACHABLE();
> +        goto unhandleable;
> +    }
> +
> +    do {
> +        enum hvm_translation_result res;
> +        struct page_info *page;
> +        pagefault_info_t pfinfo;
> +        p2m_type_t p2mt;
> +

Can't frame now be made local to this loop and calculated up-front using linear and i? It looks like the loop could become a simpler for-loop too.

> +        /* Error checking.  Confirm that the current slot is clean. */
> +        ASSERT(mfn_x(*mfn) == 0);
> +
> +        res = hvm_translate_get_page(curr, frame, true, pfec,
> +                                     &pfinfo, &page, NULL, &p2mt);
> +
> +        switch ( res )
> +        {
> +        case HVMTRANS_okay:
> +            break;
> +
> +        case HVMTRANS_bad_linear_to_gfn:
> +            x86_emul_pagefault(pfinfo.ec, pfinfo.linear, &hvmemul_ctxt->ctxt);
> +            err = ERR_PTR(~X86EMUL_EXCEPTION);
> +            goto out;
> +
> +        case HVMTRANS_bad_gfn_to_mfn:
> +            err = NULL;
> +            goto out;
> +
> +        case HVMTRANS_gfn_paged_out:
> +        case HVMTRANS_gfn_shared:
> +            err = ERR_PTR(~X86EMUL_RETRY);
> +            goto out;
> +
> +        default:
> +            goto unhandleable;
> +        }
> +
> +        *mfn++ = _mfn(page_to_mfn(page));
> +        i++;
> +
> +        if ( p2m_is_discard_write(p2mt) )
> +        {
> +            err = ERR_PTR(~X86EMUL_OKAY);
> +            goto out;
> +        }
> +
> +        frame = (linear + i);
> +        if ( hvmemul_ctxt->ctxt.addr_size < 64 )
> +            frame = (uint32_t)frame;
> +        frame = frame << PAGE_SHIFT;
> +
> +    } while ( i < nr_frames );
> +
> +    /* Entire access within a single frame? */
> +    if ( first == final )
> +        mapping = map_domain_page(hvmemul_ctxt->mfn[0]);
> +    /* Multiple frames? Need to vmap(). */
> +    else if ( (mapping = vmap(hvmemul_ctxt->mfn,
> +                              nr_frames)) == NULL )
> +        goto unhandleable;
> +
> +#ifndef NDEBUG /* Poision unused mfn[]s with INVALID_MFN. */
> +    while ( mfn < hvmemul_ctxt->mfn + ARRAY_SIZE(hvmemul_ctxt->mfn) )
> +    {
> +        ASSERT(mfn_x(*mfn) == 0);
> +        *mfn++ = INVALID_MFN;
> +    }
> +#endif
> +
> +    return mapping + (linear & ~PAGE_MASK);
> +
> + unhandleable:
> +    err = ERR_PTR(~X86EMUL_UNHANDLEABLE);
> +
> + out:
> +    /* Drop all held references. */
> +    while ( mfn-- > hvmemul_ctxt->mfn )
> +        put_page(mfn_to_page(mfn_x(*mfn)));
> +
> +    return err;
> +}
> +
> +static void hvmemul_unmap_linear_addr(
> +    void *mapping, unsigned long linear, unsigned int bytes,
> +    struct hvm_emulate_ctxt *hvmemul_ctxt)
> +{
> +    struct domain *currd = current->domain;
> +    unsigned long frame = linear >> PAGE_SHIFT;
> +    unsigned long final = (linear + bytes - !!bytes) >> PAGE_SHIFT;
> +    mfn_t *mfn = &hvmemul_ctxt->mfn[0];
> +

Extraneous blank line.

> +
> +    ASSERT(bytes > 0);
> +
> +    if ( frame == final )
> +        unmap_domain_page(mapping);
> +    else
> +        vunmap(mapping);
> +
> +    do
> +    {
> +        ASSERT(mfn_valid(*mfn));
> +        paging_mark_dirty(currd, *mfn);
> +        put_page(mfn_to_page(mfn_x(*mfn)));
> +
> +        *mfn++ = _mfn(0); /* Clean slot for map()'s error checking. */
> +    } while ( ++frame < final );

Do you not have the same overflow issues here? I would have though the looping construct here should be near identical to that in hvmemul_map_linear_addr().

  Paul

> +
> +
> +#ifndef NDEBUG /* Check (and clean) all unused mfns. */
> +    while ( mfn < hvmemul_ctxt->mfn + ARRAY_SIZE(hvmemul_ctxt->mfn) )
> +    {
> +        ASSERT(mfn_eq(*mfn, INVALID_MFN));
> +        *mfn++ = _mfn(0);
> +    }
> +#endif
> +}
> +
> +/*
>   * Convert addr from linear to physical form, valid over the range
>   * [addr, addr + *reps * bytes_per_rep]. *reps is adjusted according to
>   * the valid computed range. It is always >0 when X86EMUL_OKAY is
> returned.
> @@ -988,11 +1148,11 @@ static int hvmemul_write(
>      struct hvm_emulate_ctxt *hvmemul_ctxt =
>          container_of(ctxt, struct hvm_emulate_ctxt, ctxt);
>      struct vcpu *curr = current;
> -    pagefault_info_t pfinfo;
>      unsigned long addr, reps = 1;
>      uint32_t pfec = PFEC_page_present | PFEC_write_access;
>      struct hvm_vcpu_io *vio = &curr->arch.hvm_vcpu.hvm_io;
>      int rc;
> +    void *mapping;
> 
>      if ( is_x86_system_segment(seg) )
>          pfec |= PFEC_implicit;
> @@ -1008,23 +1168,16 @@ static int hvmemul_write(
>           (vio->mmio_gla == (addr & PAGE_MASK)) )
>          return hvmemul_linear_mmio_write(addr, bytes, p_data, pfec,
> hvmemul_ctxt, 1);
> 
> -    rc = hvm_copy_to_guest_linear(addr, p_data, bytes, pfec, &pfinfo);
> +    mapping = hvmemul_map_linear_addr(addr, bytes, pfec,
> hvmemul_ctxt);
> +    if ( IS_ERR(mapping) )
> +        return ~PTR_ERR(mapping);
> 
> -    switch ( rc )
> -    {
> -    case HVMTRANS_okay:
> -        break;
> -    case HVMTRANS_bad_linear_to_gfn:
> -        x86_emul_pagefault(pfinfo.ec, pfinfo.linear, &hvmemul_ctxt->ctxt);
> -        return X86EMUL_EXCEPTION;
> -    case HVMTRANS_bad_gfn_to_mfn:
> +    if ( !mapping )
>          return hvmemul_linear_mmio_write(addr, bytes, p_data, pfec,
> hvmemul_ctxt, 0);
> -    case HVMTRANS_gfn_paged_out:
> -    case HVMTRANS_gfn_shared:
> -        return X86EMUL_RETRY;
> -    default:
> -        return X86EMUL_UNHANDLEABLE;
> -    }
> +
> +    memcpy(mapping, p_data, bytes);
> +
> +    hvmemul_unmap_linear_addr(mapping, addr, bytes, hvmemul_ctxt);
> 
>      return X86EMUL_OKAY;
>  }
> diff --git a/xen/include/asm-x86/hvm/emulate.h b/xen/include/asm-
> x86/hvm/emulate.h
> index 8864775..d379a4a 100644
> --- a/xen/include/asm-x86/hvm/emulate.h
> +++ b/xen/include/asm-x86/hvm/emulate.h
> @@ -37,6 +37,13 @@ struct hvm_emulate_ctxt {
>      unsigned long seg_reg_accessed;
>      unsigned long seg_reg_dirty;
> 
> +    /*
> +     * MFNs behind temporary mappings in the write callback.  The length is
> +     * arbitrary, and can be increased if writes longer than PAGE_SIZE+1 are
> +     * needed.
> +     */
> +    mfn_t mfn[2];
> +
>      uint32_t intr_shadow;
> 
>      bool_t set_context;
> --
> 2.7.4
Jan Beulich Sept. 28, 2017, 9:30 a.m. UTC | #2
>>> On 28.09.17 at 11:07, <Paul.Durrant@citrix.com> wrote:
>> From: Alexandru Isaila [mailto:aisaila@bitdefender.com]
>> Sent: 28 September 2017 09:52
>> +static void *hvmemul_map_linear_addr(
>> +    unsigned long linear, unsigned int bytes, uint32_t pfec,
>> +    struct hvm_emulate_ctxt *hvmemul_ctxt)
>> +{
>> +    struct vcpu *curr = current;
>> +    void *err, *mapping;
>> +
>> +    /* First and final gfns which need mapping. */
>> +    unsigned long frame = linear >> PAGE_SHIFT, first = frame;
>> +    unsigned long final = (linear + bytes - !!bytes) >> PAGE_SHIFT;
>> +    unsigned long nr_frames = ((linear + bytes) >> PAGE_SHIFT) - (linear >>
>> PAGE_SHIFT) + 1;
> 
> I'm unclear as to why you don't calculate nr_frames first and then calculate 
> final as first + nr_frames - 1? Having completely independent calculations 
> like this seems prone to error.

Nor is the calculation any better than that in v6 for the 64-bit
wrapping case. Easiest here may be to utilize truncation by
making nr_frames unsigned int (but you'll need to check carefully
that this doesn't cause issues elsewhere).

Plus the use of "linear + bytes" (without the "- !!bytes") causes the
result to be one too high when "linear" is exactly "bytes" away from
a page boundary, and "bytes" is not zero.

>> +    int i = 0;

unsigned int

>> +    /*
>> +     * mfn points to the next free slot.  All used slots have a page 
> reference
>> +     * held on them.
>> +     */
>> +    mfn_t *mfn = &hvmemul_ctxt->mfn[0];
>> +
> 
> Extraneous blank line.
> 
>> +
>> +    /*
>> +     * The caller has no legitimate reason for trying a zero-byte write, but
>> +     * final is calculate to fail safe in release builds.
>> +     *
>> +     * The maximum write size depends on the number of adjacent mfns[]
>> which
>> +     * can be vmap()'d, accouting for possible misalignment within the region.
>> +     * The higher level emulation callers are responsible for ensuring that
>> +     * mfns[] is large enough for the requested write size.
>> +     */
>> +    if ( bytes == 0 ||
>> +         nr_frames > ARRAY_SIZE(hvmemul_ctxt->mfn) )
>> +    {
>> +        ASSERT_UNREACHABLE();
>> +        goto unhandleable;
>> +    }
>> +
>> +    do {
>> +        enum hvm_translation_result res;
>> +        struct page_info *page;
>> +        pagefault_info_t pfinfo;
>> +        p2m_type_t p2mt;
>> +
> 
> Can't frame now be made local to this loop and calculated up-front using 
> linear and i? It looks like the loop could become a simpler for-loop too.
> 
>> +        /* Error checking.  Confirm that the current slot is clean. */
>> +        ASSERT(mfn_x(*mfn) == 0);
>> +
>> +        res = hvm_translate_get_page(curr, frame, true, pfec,
>> +                                     &pfinfo, &page, NULL, &p2mt);
>> +
>> +        switch ( res )
>> +        {
>> +        case HVMTRANS_okay:
>> +            break;
>> +
>> +        case HVMTRANS_bad_linear_to_gfn:
>> +            x86_emul_pagefault(pfinfo.ec, pfinfo.linear, &hvmemul_ctxt->ctxt);
>> +            err = ERR_PTR(~X86EMUL_EXCEPTION);
>> +            goto out;
>> +
>> +        case HVMTRANS_bad_gfn_to_mfn:
>> +            err = NULL;
>> +            goto out;
>> +
>> +        case HVMTRANS_gfn_paged_out:
>> +        case HVMTRANS_gfn_shared:
>> +            err = ERR_PTR(~X86EMUL_RETRY);
>> +            goto out;
>> +
>> +        default:
>> +            goto unhandleable;
>> +        }
>> +
>> +        *mfn++ = _mfn(page_to_mfn(page));
>> +        i++;
>> +
>> +        if ( p2m_is_discard_write(p2mt) )
>> +        {
>> +            err = ERR_PTR(~X86EMUL_OKAY);
>> +            goto out;
>> +        }
>> +
>> +        frame = (linear + i);

i is the number of pages, not bytes. Also stray parentheses here.

>> +        if ( hvmemul_ctxt->ctxt.addr_size < 64 )
>> +            frame = (uint32_t)frame;
>> +        frame = frame << PAGE_SHIFT;
>> +
>> +    } while ( i < nr_frames );
>> +
>> +    /* Entire access within a single frame? */
>> +    if ( first == final )

nr_frames == 1

>> +        mapping = map_domain_page(hvmemul_ctxt->mfn[0]);
>> +    /* Multiple frames? Need to vmap(). */
>> +    else if ( (mapping = vmap(hvmemul_ctxt->mfn,
>> +                              nr_frames)) == NULL )
>> +        goto unhandleable;
>> +
>> +#ifndef NDEBUG /* Poision unused mfn[]s with INVALID_MFN. */
>> +    while ( mfn < hvmemul_ctxt->mfn + ARRAY_SIZE(hvmemul_ctxt->mfn) )
>> +    {
>> +        ASSERT(mfn_x(*mfn) == 0);
>> +        *mfn++ = INVALID_MFN;
>> +    }
>> +#endif
>> +
>> +    return mapping + (linear & ~PAGE_MASK);
>> +
>> + unhandleable:
>> +    err = ERR_PTR(~X86EMUL_UNHANDLEABLE);
>> +
>> + out:
>> +    /* Drop all held references. */
>> +    while ( mfn-- > hvmemul_ctxt->mfn )
>> +        put_page(mfn_to_page(mfn_x(*mfn)));
>> +
>> +    return err;
>> +}
>> +
>> +static void hvmemul_unmap_linear_addr(
>> +    void *mapping, unsigned long linear, unsigned int bytes,
>> +    struct hvm_emulate_ctxt *hvmemul_ctxt)
>> +{
>> +    struct domain *currd = current->domain;
>> +    unsigned long frame = linear >> PAGE_SHIFT;
>> +    unsigned long final = (linear + bytes - !!bytes) >> PAGE_SHIFT;
>> +    mfn_t *mfn = &hvmemul_ctxt->mfn[0];
>> +
> 
> Extraneous blank line.
> 
>> +
>> +    ASSERT(bytes > 0);
>> +
>> +    if ( frame == final )
>> +        unmap_domain_page(mapping);
>> +    else
>> +        vunmap(mapping);
>> +
>> +    do
>> +    {
>> +        ASSERT(mfn_valid(*mfn));
>> +        paging_mark_dirty(currd, *mfn);
>> +        put_page(mfn_to_page(mfn_x(*mfn)));
>> +
>> +        *mfn++ = _mfn(0); /* Clean slot for map()'s error checking. */
>> +    } while ( ++frame < final );
> 
> Do you not have the same overflow issues here? I would have though the 
> looping construct here should be near identical to that in 
> hvmemul_map_linear_addr().

Indeed, and I _specifically_ said so in reply to v6. Alexandru,
please be quite a bit more careful. Just to emphasize it again -
any comments made to the mapping code need to also be
checked for applicability to the unmapping side.

Jan
diff mbox

Patch

diff --git a/xen/arch/x86/hvm/emulate.c b/xen/arch/x86/hvm/emulate.c
index cc874ce..7d18bf5 100644
--- a/xen/arch/x86/hvm/emulate.c
+++ b/xen/arch/x86/hvm/emulate.c
@@ -498,6 +498,166 @@  static int hvmemul_do_mmio_addr(paddr_t mmio_gpa,
 }
 
 /*
+ * Map the frame(s) covering an individual linear access, for writeable
+ * access.  May return NULL for MMIO, or ERR_PTR(~X86EMUL_*) for other errors
+ * including ERR_PTR(~X86EMUL_OKAY) for write-discard mappings.
+ *
+ * In debug builds, map() checks that each slot in hvmemul_ctxt->mfn[] is
+ * clean before use, and poisions unused slots with INVALID_MFN.
+ */
+static void *hvmemul_map_linear_addr(
+    unsigned long linear, unsigned int bytes, uint32_t pfec,
+    struct hvm_emulate_ctxt *hvmemul_ctxt)
+{
+    struct vcpu *curr = current;
+    void *err, *mapping;
+
+    /* First and final gfns which need mapping. */
+    unsigned long frame = linear >> PAGE_SHIFT, first = frame;
+    unsigned long final = (linear + bytes - !!bytes) >> PAGE_SHIFT;
+    unsigned long nr_frames = ((linear + bytes) >> PAGE_SHIFT) - (linear >> PAGE_SHIFT) + 1;
+    int i = 0;
+
+    /*
+     * mfn points to the next free slot.  All used slots have a page reference
+     * held on them.
+     */
+    mfn_t *mfn = &hvmemul_ctxt->mfn[0];
+
+
+    /*
+     * The caller has no legitimate reason for trying a zero-byte write, but
+     * final is calculate to fail safe in release builds.
+     *
+     * The maximum write size depends on the number of adjacent mfns[] which
+     * can be vmap()'d, accouting for possible misalignment within the region.
+     * The higher level emulation callers are responsible for ensuring that
+     * mfns[] is large enough for the requested write size.
+     */
+    if ( bytes == 0 ||
+         nr_frames > ARRAY_SIZE(hvmemul_ctxt->mfn) )
+    {
+        ASSERT_UNREACHABLE();
+        goto unhandleable;
+    }
+
+    do {
+        enum hvm_translation_result res;
+        struct page_info *page;
+        pagefault_info_t pfinfo;
+        p2m_type_t p2mt;
+
+        /* Error checking.  Confirm that the current slot is clean. */
+        ASSERT(mfn_x(*mfn) == 0);
+
+        res = hvm_translate_get_page(curr, frame, true, pfec,
+                                     &pfinfo, &page, NULL, &p2mt);
+
+        switch ( res )
+        {
+        case HVMTRANS_okay:
+            break;
+
+        case HVMTRANS_bad_linear_to_gfn:
+            x86_emul_pagefault(pfinfo.ec, pfinfo.linear, &hvmemul_ctxt->ctxt);
+            err = ERR_PTR(~X86EMUL_EXCEPTION);
+            goto out;
+
+        case HVMTRANS_bad_gfn_to_mfn:
+            err = NULL;
+            goto out;
+
+        case HVMTRANS_gfn_paged_out:
+        case HVMTRANS_gfn_shared:
+            err = ERR_PTR(~X86EMUL_RETRY);
+            goto out;
+
+        default:
+            goto unhandleable;
+        }
+
+        *mfn++ = _mfn(page_to_mfn(page));
+        i++;
+
+        if ( p2m_is_discard_write(p2mt) )
+        {
+            err = ERR_PTR(~X86EMUL_OKAY);
+            goto out;
+        }
+
+        frame = (linear + i);
+        if ( hvmemul_ctxt->ctxt.addr_size < 64 )
+            frame = (uint32_t)frame;
+        frame = frame << PAGE_SHIFT;
+
+    } while ( i < nr_frames );
+
+    /* Entire access within a single frame? */
+    if ( first == final )
+        mapping = map_domain_page(hvmemul_ctxt->mfn[0]);
+    /* Multiple frames? Need to vmap(). */
+    else if ( (mapping = vmap(hvmemul_ctxt->mfn,
+                              nr_frames)) == NULL )
+        goto unhandleable;
+
+#ifndef NDEBUG /* Poision unused mfn[]s with INVALID_MFN. */
+    while ( mfn < hvmemul_ctxt->mfn + ARRAY_SIZE(hvmemul_ctxt->mfn) )
+    {
+        ASSERT(mfn_x(*mfn) == 0);
+        *mfn++ = INVALID_MFN;
+    }
+#endif
+
+    return mapping + (linear & ~PAGE_MASK);
+
+ unhandleable:
+    err = ERR_PTR(~X86EMUL_UNHANDLEABLE);
+
+ out:
+    /* Drop all held references. */
+    while ( mfn-- > hvmemul_ctxt->mfn )
+        put_page(mfn_to_page(mfn_x(*mfn)));
+
+    return err;
+}
+
+static void hvmemul_unmap_linear_addr(
+    void *mapping, unsigned long linear, unsigned int bytes,
+    struct hvm_emulate_ctxt *hvmemul_ctxt)
+{
+    struct domain *currd = current->domain;
+    unsigned long frame = linear >> PAGE_SHIFT;
+    unsigned long final = (linear + bytes - !!bytes) >> PAGE_SHIFT;
+    mfn_t *mfn = &hvmemul_ctxt->mfn[0];
+
+
+    ASSERT(bytes > 0);
+
+    if ( frame == final )
+        unmap_domain_page(mapping);
+    else
+        vunmap(mapping);
+
+    do
+    {
+        ASSERT(mfn_valid(*mfn));
+        paging_mark_dirty(currd, *mfn);
+        put_page(mfn_to_page(mfn_x(*mfn)));
+
+        *mfn++ = _mfn(0); /* Clean slot for map()'s error checking. */
+    } while ( ++frame < final );
+
+
+#ifndef NDEBUG /* Check (and clean) all unused mfns. */
+    while ( mfn < hvmemul_ctxt->mfn + ARRAY_SIZE(hvmemul_ctxt->mfn) )
+    {
+        ASSERT(mfn_eq(*mfn, INVALID_MFN));
+        *mfn++ = _mfn(0);
+    }
+#endif
+}
+
+/*
  * Convert addr from linear to physical form, valid over the range
  * [addr, addr + *reps * bytes_per_rep]. *reps is adjusted according to
  * the valid computed range. It is always >0 when X86EMUL_OKAY is returned.
@@ -988,11 +1148,11 @@  static int hvmemul_write(
     struct hvm_emulate_ctxt *hvmemul_ctxt =
         container_of(ctxt, struct hvm_emulate_ctxt, ctxt);
     struct vcpu *curr = current;
-    pagefault_info_t pfinfo;
     unsigned long addr, reps = 1;
     uint32_t pfec = PFEC_page_present | PFEC_write_access;
     struct hvm_vcpu_io *vio = &curr->arch.hvm_vcpu.hvm_io;
     int rc;
+    void *mapping;
 
     if ( is_x86_system_segment(seg) )
         pfec |= PFEC_implicit;
@@ -1008,23 +1168,16 @@  static int hvmemul_write(
          (vio->mmio_gla == (addr & PAGE_MASK)) )
         return hvmemul_linear_mmio_write(addr, bytes, p_data, pfec, hvmemul_ctxt, 1);
 
-    rc = hvm_copy_to_guest_linear(addr, p_data, bytes, pfec, &pfinfo);
+    mapping = hvmemul_map_linear_addr(addr, bytes, pfec, hvmemul_ctxt);
+    if ( IS_ERR(mapping) )
+        return ~PTR_ERR(mapping);
 
-    switch ( rc )
-    {
-    case HVMTRANS_okay:
-        break;
-    case HVMTRANS_bad_linear_to_gfn:
-        x86_emul_pagefault(pfinfo.ec, pfinfo.linear, &hvmemul_ctxt->ctxt);
-        return X86EMUL_EXCEPTION;
-    case HVMTRANS_bad_gfn_to_mfn:
+    if ( !mapping )
         return hvmemul_linear_mmio_write(addr, bytes, p_data, pfec, hvmemul_ctxt, 0);
-    case HVMTRANS_gfn_paged_out:
-    case HVMTRANS_gfn_shared:
-        return X86EMUL_RETRY;
-    default:
-        return X86EMUL_UNHANDLEABLE;
-    }
+
+    memcpy(mapping, p_data, bytes);
+
+    hvmemul_unmap_linear_addr(mapping, addr, bytes, hvmemul_ctxt);
 
     return X86EMUL_OKAY;
 }
diff --git a/xen/include/asm-x86/hvm/emulate.h b/xen/include/asm-x86/hvm/emulate.h
index 8864775..d379a4a 100644
--- a/xen/include/asm-x86/hvm/emulate.h
+++ b/xen/include/asm-x86/hvm/emulate.h
@@ -37,6 +37,13 @@  struct hvm_emulate_ctxt {
     unsigned long seg_reg_accessed;
     unsigned long seg_reg_dirty;
 
+    /*
+     * MFNs behind temporary mappings in the write callback.  The length is
+     * arbitrary, and can be increased if writes longer than PAGE_SIZE+1 are
+     * needed.
+     */
+    mfn_t mfn[2];
+
     uint32_t intr_shadow;
 
     bool_t set_context;