[PATCH partly-for-4.20 v3 0/4] x86/HVM: emulation (MMIO) improvements

[PATCH partly-for-4.20 v3 0/4] x86/HVM: emulation (MMIO) improvements

[Jan Beulich]

The main fix is patch 2, with the earlier patch setting the stage and
the latter ones simplifying other things at least a little in exchange.

1: allocate emulation cache entries dynamically
2: correct read/write split at page boundaries
3: slightly improve CMPXCHG16B emulation
4: drop redundant access splitting

Oleksii - the first two patches (plus the simple one from v2 that I just
committed) are backporting material, and hence I'd like to ask for a
release ack for at least those two.

Jan

[PATCH for-4.20 v3 1/4] x86/HVM: allocate emulation cache entries dynamically

[Jan Beulich]

Both caches may need higher capacity, and the upper bound will need to
be determined dynamically based on CPUID policy (for AMX'es TILELOAD /
TILESTORE at least).

Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
---
This is a patch taken from the AMX series, but wasn't part of the v3
submission. All I did is strip out the actual AMX bits (from
hvmemul_cache_init()), plus of course change the description. As a
result some local variables there may look unnecessary, but this way
it's going to be less churn when the AMX bits are added. The next patch
pretty strongly depends on the changed approach (contextually, not so
much functionally), and I'd really like to avoid rebasing that one ahead
of this one, and then this one on top of that.

TBD: For AMX hvmemul_cache_init() will become CPUID policy dependent. We
     could of course take the opportunity and also reduce overhead when
     AVX-512 (and maybe even AVX) is unavailable (in the future: to the
     guest).
---
v3: Use xvmalloc variants throughout. Add comments.
v2: Split off cache bounds check fix.

--- a/xen/arch/x86/hvm/emulate.c
+++ b/xen/arch/x86/hvm/emulate.c
@@ -26,6 +26,18 @@
 #include <asm/iocap.h>
 #include <asm/vm_event.h>
 
+/*
+ * We may read or write up to m512 or up to a tile row as a number of
+ * device-model transactions.
+ */
+struct hvm_mmio_cache {
+    unsigned long gla;
+    unsigned int size;     /* Amount of buffer[] actually used. */
+    unsigned int space:31; /* Allocated size of buffer[]. */
+    unsigned int dir:1;
+    uint8_t buffer[] __aligned(sizeof(long));
+};
+
 struct hvmemul_cache
 {
     /* The cache is disabled as long as num_ents > max_ents. */
@@ -935,7 +947,7 @@ static int hvmemul_phys_mmio_access(
     }
 
     /* Accesses must not overflow the cache's buffer. */
-    if ( offset + size > sizeof(cache->buffer) )
+    if ( offset + size > cache->space )
     {
         ASSERT_UNREACHABLE();
         return X86EMUL_UNHANDLEABLE;
@@ -1011,7 +1023,7 @@ static struct hvm_mmio_cache *hvmemul_fi
 
     for ( i = 0; i < hvio->mmio_cache_count; i ++ )
     {
-        cache = &hvio->mmio_cache[i];
+        cache = hvio->mmio_cache[i];
 
         if ( gla == cache->gla &&
              dir == cache->dir )
@@ -1027,10 +1039,11 @@ static struct hvm_mmio_cache *hvmemul_fi
 
     ++hvio->mmio_cache_count;
 
-    cache = &hvio->mmio_cache[i];
-    memset(cache, 0, sizeof (*cache));
+    cache = hvio->mmio_cache[i];
+    memset(cache->buffer, 0, cache->space);
 
     cache->gla = gla;
+    cache->size = 0;
     cache->dir = dir;
 
     return cache;
@@ -2980,16 +2993,21 @@ void hvm_dump_emulation_state(const char
 int hvmemul_cache_init(struct vcpu *v)
 {
     /*
-     * No insn can access more than 16 independent linear addresses (AVX512F
-     * scatters/gathers being the worst). Each such linear range can span a
-     * page boundary, i.e. may require two page walks. Account for each insn
-     * byte individually, for simplicity.
+     * AVX512F scatter/gather insns can access up to 16 independent linear
+     * addresses, up to 8 bytes size. Each such linear range can span a page
+     * boundary, i.e. may require two page walks.
+     */
+    unsigned int nents = 16 * 2 * (CONFIG_PAGING_LEVELS + 1);
+    unsigned int i, max_bytes = 64;
+    struct hvmemul_cache *cache;
+
+    /*
+     * Account for each insn byte individually, both for simplicity and to
+     * leave some slack space.
      */
-    const unsigned int nents = (CONFIG_PAGING_LEVELS + 1) *
-                               (MAX_INST_LEN + 16 * 2);
-    struct hvmemul_cache *cache = xmalloc_flex_struct(struct hvmemul_cache,
-                                                      ents, nents);
+    nents += MAX_INST_LEN * (CONFIG_PAGING_LEVELS + 1);
 
+    cache = xvmalloc_flex_struct(struct hvmemul_cache, ents, nents);
     if ( !cache )
         return -ENOMEM;
 
@@ -2999,6 +3017,15 @@ int hvmemul_cache_init(struct vcpu *v)
 
     v->arch.hvm.hvm_io.cache = cache;
 
+    for ( i = 0; i < ARRAY_SIZE(v->arch.hvm.hvm_io.mmio_cache); ++i )
+    {
+        v->arch.hvm.hvm_io.mmio_cache[i] =
+            xvmalloc_flex_struct(struct hvm_mmio_cache, buffer, max_bytes);
+        if ( !v->arch.hvm.hvm_io.mmio_cache[i] )
+            return -ENOMEM;
+        v->arch.hvm.hvm_io.mmio_cache[i]->space = max_bytes;
+    }
+
     return 0;
 }
 
--- a/xen/arch/x86/include/asm/hvm/emulate.h
+++ b/xen/arch/x86/include/asm/hvm/emulate.h
@@ -15,6 +15,7 @@
 #include <xen/err.h>
 #include <xen/mm.h>
 #include <xen/sched.h>
+#include <xen/xvmalloc.h>
 #include <asm/hvm/hvm.h>
 #include <asm/x86_emulate.h>
 
@@ -119,7 +120,11 @@ int hvmemul_do_pio_buffer(uint16_t port,
 int __must_check hvmemul_cache_init(struct vcpu *v);
 static inline void hvmemul_cache_destroy(struct vcpu *v)
 {
-    XFREE(v->arch.hvm.hvm_io.cache);
+    unsigned int i;
+
+    for ( i = 0; i < ARRAY_SIZE(v->arch.hvm.hvm_io.mmio_cache); ++i )
+        XFREE(v->arch.hvm.hvm_io.mmio_cache[i]);
+    XVFREE(v->arch.hvm.hvm_io.cache);
 }
 bool hvmemul_read_cache(const struct vcpu *v, paddr_t gpa,
                         void *buffer, unsigned int size);
--- a/xen/arch/x86/include/asm/hvm/vcpu.h
+++ b/xen/arch/x86/include/asm/hvm/vcpu.h
@@ -22,17 +22,6 @@ struct hvm_vcpu_asid {
     uint32_t asid;
 };
 
-/*
- * We may read or write up to m512 as a number of device-model
- * transactions.
- */
-struct hvm_mmio_cache {
-    unsigned long gla;
-    unsigned int size;
-    uint8_t dir;
-    uint8_t buffer[64] __aligned(sizeof(long));
-};
-
 struct hvm_vcpu_io {
     /*
      * HVM emulation:
@@ -48,7 +37,7 @@ struct hvm_vcpu_io {
      * We may need to handle up to 3 distinct memory accesses per
      * instruction.
      */
-    struct hvm_mmio_cache mmio_cache[3];
+    struct hvm_mmio_cache *mmio_cache[3];
     unsigned int mmio_cache_count;
 
     /* For retries we shouldn't re-fetch the instruction. */

[PATCH for-4.20 v3 2/4] x86/HVM: correct read/write split at page boundaries

[Jan Beulich]

The MMIO cache is intended to have one entry used per independent memory
access that an insn does. This, in particular, is supposed to be
ignoring any page boundary crossing. Therefore when looking up a cache
entry, the access'es starting (linear) address is relevant, not the one
possibly advanced past a page boundary.

In order for the same offset-into-buffer variable to be usable in
hvmemul_phys_mmio_access() for both the caller's buffer and the cache
entry's it is further necessary to have the un-adjusted caller buffer
passed into there.

Fixes: 2d527ba310dc ("x86/hvm: split all linear reads and writes at page boundary")
Reported-by: Manuel Andreas <manuel.andreas@tum.de>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
---
This way problematic overlaps are only reduced (to ones starting at the
same address), not eliminated: Assumptions in hvmemul_phys_mmio_access()
go further - if a subsequent access is larger than an earlier one, but
the splitting results in a chunk to cross the end "boundary" of the
earlier access, an assertion will still trigger. Explicit memory
accesses (ones encoded in an insn by explicit or implicit memory
operands) match the assumption afaict (i.e. all those accesses are of
uniform size, and hence they either fully overlap or are mapped to
distinct cache entries).

Insns accessing descriptor tables, otoh, don't fulfill these
expectations: The selector read (if coming from memory) will always be
smaller than the descriptor being read, and if both (insanely) start at
the same linear address (in turn mapping MMIO), said assertion will kick
in. (The same would be true for an insn trying to access itself as data,
as long as certain size "restrictions" between insn and memory operand
are met. Except that linear_read() disallows insn fetches from MMIO.) To
deal with such, I expect we will need to further qualify (tag) cache
entries, such that reads/writes won't use insn fetch entries, and
implicit-supervisor-mode accesses won't use entries of ordinary
accesses. (Page table accesses don't need considering here for now, as
our page walking code demands page tables to be mappable, implying
they're in guest RAM; such accesses also don't take the path here.)
Thoughts anyone, before I get to making another patch?

Considering the insn fetch aspect mentioned above I'm having trouble
following why the cache has 3 entries. With insn fetches permitted,
descriptor table accesses where the accessed bit needs setting may also
fail because of that limited capacity of the cache, due to the way the
accesses are done. The read and write (cmpxchg) are independent accesses
from the cache's perspective, and hence we'd need another entry there.
If, otoh, the 3 entries are there to account for precisely this (which
seems unlikely with commit e101123463d2 ["x86/hvm: track large memory
mapped accesses by buffer offset"] not saying anything at all), then we
should be fine in this regard. If we were to permit insn fetches, which
way to overcome this (possibly by allowing the write to re-use the
earlier read's entry in this special situation) would remain to be
determined.
---
v3: Rename "start" function parameters to "start_gla". Adjust wording of
    a comment. Re-base over comment addition in an earlier patch.

--- a/xen/arch/x86/hvm/emulate.c
+++ b/xen/arch/x86/hvm/emulate.c
@@ -31,8 +31,9 @@
  * device-model transactions.
  */
 struct hvm_mmio_cache {
-    unsigned long gla;
-    unsigned int size;     /* Amount of buffer[] actually used. */
+    unsigned long gla;     /* Start of original access (e.g. insn operand). */
+    unsigned int skip;     /* Offset to start of MMIO */
+    unsigned int size;     /* Amount of buffer[] actually used, incl @skip. */
     unsigned int space:31; /* Allocated size of buffer[]. */
     unsigned int dir:1;
     uint8_t buffer[] __aligned(sizeof(long));
@@ -953,6 +954,13 @@ static int hvmemul_phys_mmio_access(
         return X86EMUL_UNHANDLEABLE;
     }
 
+    /* Accesses must not be to the unused leading space. */
+    if ( offset < cache->skip )
+    {
+        ASSERT_UNREACHABLE();
+        return X86EMUL_UNHANDLEABLE;
+    }
+
     /*
      * hvmemul_do_io() cannot handle non-power-of-2 accesses or
      * accesses larger than sizeof(long), so choose the highest power
@@ -1010,13 +1018,15 @@ static int hvmemul_phys_mmio_access(
 
 /*
  * Multi-cycle MMIO handling is based upon the assumption that emulation
- * of the same instruction will not access the same MMIO region more
- * than once. Hence we can deal with re-emulation (for secondary or
- * subsequent cycles) by looking up the result or previous I/O in a
- * cache indexed by linear MMIO address.
+ * of the same instruction will not access the exact same MMIO region
+ * more than once in exactly the same way (if it does, the accesses will
+ * be "folded"). Hence we can deal with re-emulation (for secondary or
+ * subsequent cycles) by looking up the result of previous I/O in a cache
+ * indexed by linear address and access type.
  */
 static struct hvm_mmio_cache *hvmemul_find_mmio_cache(
-    struct hvm_vcpu_io *hvio, unsigned long gla, uint8_t dir, bool create)
+    struct hvm_vcpu_io *hvio, unsigned long gla, uint8_t dir,
+    unsigned int skip)
 {
     unsigned int i;
     struct hvm_mmio_cache *cache;
@@ -1030,7 +1040,11 @@ static struct hvm_mmio_cache *hvmemul_fi
             return cache;
     }
 
-    if ( !create )
+    /*
+     * Bail if a new entry shouldn't be allocated, relying on ->space having
+     * the same value for all entries.
+     */
+    if ( skip >= hvio->mmio_cache[0]->space )
         return NULL;
 
     i = hvio->mmio_cache_count;
@@ -1043,7 +1057,8 @@ static struct hvm_mmio_cache *hvmemul_fi
     memset(cache->buffer, 0, cache->space);
 
     cache->gla = gla;
-    cache->size = 0;
+    cache->skip = skip;
+    cache->size = skip;
     cache->dir = dir;
 
     return cache;
@@ -1064,12 +1079,14 @@ static void latch_linear_to_phys(struct
 
 static int hvmemul_linear_mmio_access(
     unsigned long gla, unsigned int size, uint8_t dir, void *buffer,
-    uint32_t pfec, struct hvm_emulate_ctxt *hvmemul_ctxt, bool known_gpfn)
+    uint32_t pfec, struct hvm_emulate_ctxt *hvmemul_ctxt,
+    unsigned long start_gla, bool known_gpfn)
 {
     struct hvm_vcpu_io *hvio = &current->arch.hvm.hvm_io;
     unsigned long offset = gla & ~PAGE_MASK;
-    struct hvm_mmio_cache *cache = hvmemul_find_mmio_cache(hvio, gla, dir, true);
-    unsigned int chunk, buffer_offset = 0;
+    unsigned int chunk, buffer_offset = gla - start_gla;
+    struct hvm_mmio_cache *cache = hvmemul_find_mmio_cache(hvio, start_gla,
+                                                           dir, buffer_offset);
     paddr_t gpa;
     unsigned long one_rep = 1;
     int rc;
@@ -1117,19 +1134,19 @@ static int hvmemul_linear_mmio_access(
 static inline int hvmemul_linear_mmio_read(
     unsigned long gla, unsigned int size, void *buffer,
     uint32_t pfec, struct hvm_emulate_ctxt *hvmemul_ctxt,
-    bool translate)
+    unsigned long start_gla, bool translate)
 {
-    return hvmemul_linear_mmio_access(gla, size, IOREQ_READ, buffer,
-                                      pfec, hvmemul_ctxt, translate);
+    return hvmemul_linear_mmio_access(gla, size, IOREQ_READ, buffer, pfec,
+                                      hvmemul_ctxt, start_gla, translate);
 }
 
 static inline int hvmemul_linear_mmio_write(
     unsigned long gla, unsigned int size, void *buffer,
     uint32_t pfec, struct hvm_emulate_ctxt *hvmemul_ctxt,
-    bool translate)
+    unsigned long start_gla, bool translate)
 {
-    return hvmemul_linear_mmio_access(gla, size, IOREQ_WRITE, buffer,
-                                      pfec, hvmemul_ctxt, translate);
+    return hvmemul_linear_mmio_access(gla, size, IOREQ_WRITE, buffer, pfec,
+                                      hvmemul_ctxt, start_gla, translate);
 }
 
 static bool known_gla(unsigned long addr, unsigned int bytes, uint32_t pfec)
@@ -1158,7 +1175,10 @@ static int linear_read(unsigned long add
 {
     pagefault_info_t pfinfo;
     struct hvm_vcpu_io *hvio = &current->arch.hvm.hvm_io;
+    void *buffer = p_data;
+    unsigned long start = addr;
     unsigned int offset = addr & ~PAGE_MASK;
+    const struct hvm_mmio_cache *cache;
     int rc;
 
     if ( offset + bytes > PAGE_SIZE )
@@ -1182,8 +1202,17 @@ static int linear_read(unsigned long add
      * an access that was previously handled as MMIO. Thus it is imperative that
      * we handle this access in the same way to guarantee completion and hence
      * clean up any interim state.
+     *
+     * Care must be taken, however, to correctly deal with crossing RAM/MMIO or
+     * MMIO/RAM boundaries. While we want to use a single cache entry (tagged
+     * by the starting linear address), we need to continue issuing (i.e. also
+     * upon replay) the RAM access for anything that's ahead of or past MMIO,
+     * i.e. in RAM.
      */
-    if ( !hvmemul_find_mmio_cache(hvio, addr, IOREQ_READ, false) )
+    cache = hvmemul_find_mmio_cache(hvio, start, IOREQ_READ, ~0);
+    if ( !cache ||
+         addr + bytes <= start + cache->skip ||
+         addr >= start + cache->size )
         rc = hvm_copy_from_guest_linear(p_data, addr, bytes, pfec, &pfinfo);
 
     switch ( rc )
@@ -1199,8 +1228,8 @@ static int linear_read(unsigned long add
         if ( pfec & PFEC_insn_fetch )
             return X86EMUL_UNHANDLEABLE;
 
-        return hvmemul_linear_mmio_read(addr, bytes, p_data, pfec,
-                                        hvmemul_ctxt,
+        return hvmemul_linear_mmio_read(addr, bytes, buffer, pfec,
+                                        hvmemul_ctxt, start,
                                         known_gla(addr, bytes, pfec));
 
     case HVMTRANS_gfn_paged_out:
@@ -1217,7 +1246,10 @@ static int linear_write(unsigned long ad
 {
     pagefault_info_t pfinfo;
     struct hvm_vcpu_io *hvio = &current->arch.hvm.hvm_io;
+    void *buffer = p_data;
+    unsigned long start = addr;
     unsigned int offset = addr & ~PAGE_MASK;
+    const struct hvm_mmio_cache *cache;
     int rc;
 
     if ( offset + bytes > PAGE_SIZE )
@@ -1236,13 +1268,11 @@ static int linear_write(unsigned long ad
 
     rc = HVMTRANS_bad_gfn_to_mfn;
 
-    /*
-     * If there is an MMIO cache entry for the access then we must be re-issuing
-     * an access that was previously handled as MMIO. Thus it is imperative that
-     * we handle this access in the same way to guarantee completion and hence
-     * clean up any interim state.
-     */
-    if ( !hvmemul_find_mmio_cache(hvio, addr, IOREQ_WRITE, false) )
+    /* See commentary in linear_read(). */
+    cache = hvmemul_find_mmio_cache(hvio, start, IOREQ_WRITE, ~0);
+    if ( !cache ||
+         addr + bytes <= start + cache->skip ||
+         addr >= start + cache->size )
         rc = hvm_copy_to_guest_linear(addr, p_data, bytes, pfec, &pfinfo);
 
     switch ( rc )
@@ -1255,8 +1285,8 @@ static int linear_write(unsigned long ad
         return X86EMUL_EXCEPTION;
 
     case HVMTRANS_bad_gfn_to_mfn:
-        return hvmemul_linear_mmio_write(addr, bytes, p_data, pfec,
-                                         hvmemul_ctxt,
+        return hvmemul_linear_mmio_write(addr, bytes, buffer, pfec,
+                                         hvmemul_ctxt, start,
                                          known_gla(addr, bytes, pfec));
 
     case HVMTRANS_gfn_paged_out:
@@ -1643,7 +1673,7 @@ static int cf_check hvmemul_cmpxchg(
     {
         /* Fix this in case the guest is really relying on r-m-w atomicity. */
         return hvmemul_linear_mmio_write(addr, bytes, p_new, pfec,
-                                         hvmemul_ctxt,
+                                         hvmemul_ctxt, addr,
                                          hvio->mmio_access.write_access &&
                                          hvio->mmio_gla == (addr & PAGE_MASK));
     }

Re: [PATCH for-4.20 v3 2/4] x86/HVM: correct read/write split at page boundaries

[Roger Pau Monné]

On Thu, Jan 23, 2025 at 11:31:19AM +0100, Jan Beulich wrote:
> The MMIO cache is intended to have one entry used per independent memory
> access that an insn does. This, in particular, is supposed to be
> ignoring any page boundary crossing. Therefore when looking up a cache
> entry, the access'es starting (linear) address is relevant, not the one
> possibly advanced past a page boundary.
> 
> In order for the same offset-into-buffer variable to be usable in
> hvmemul_phys_mmio_access() for both the caller's buffer and the cache
> entry's it is further necessary to have the un-adjusted caller buffer
> passed into there.
> 
> Fixes: 2d527ba310dc ("x86/hvm: split all linear reads and writes at page boundary")
> Reported-by: Manuel Andreas <manuel.andreas@tum.de>
> Signed-off-by: Jan Beulich <jbeulich@suse.com>

Acked-by: Roger Pau Monné <roger.pau@citrix.com>

Thanks, Roger.

[PATCH v3 3/4] x86/HVM: slightly improve CMPXCHG16B emulation

[Jan Beulich]

Using hvmemul_linear_mmio_write() directly (as fallback when mapping the
memory operand isn't possible) won't work properly when the access
crosses a RAM/MMIO boundary. Use linear_write() instead, which splits at
such boundaries as necessary.

Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>

--- a/xen/arch/x86/hvm/emulate.c
+++ b/xen/arch/x86/hvm/emulate.c
@@ -1645,10 +1645,8 @@ static int cf_check hvmemul_cmpxchg(
 {
     struct hvm_emulate_ctxt *hvmemul_ctxt =
         container_of(ctxt, struct hvm_emulate_ctxt, ctxt);
-    struct vcpu *curr = current;
     unsigned long addr;
     uint32_t pfec = PFEC_page_present | PFEC_write_access;
-    struct hvm_vcpu_io *hvio = &curr->arch.hvm.hvm_io;
     int rc;
     void *mapping = NULL;
 
@@ -1672,10 +1670,7 @@ static int cf_check hvmemul_cmpxchg(
     if ( !mapping )
     {
         /* Fix this in case the guest is really relying on r-m-w atomicity. */
-        return hvmemul_linear_mmio_write(addr, bytes, p_new, pfec,
-                                         hvmemul_ctxt, addr,
-                                         hvio->mmio_access.write_access &&
-                                         hvio->mmio_gla == (addr & PAGE_MASK));
+        return linear_write(addr, bytes, p_new, pfec, hvmemul_ctxt);
     }
 
     switch ( bytes )

[PATCH v3 4/4] x86/HVM: drop redundant access splitting

[Jan Beulich]

With all paths into hvmemul_linear_mmio_access() coming through
linear_{read,write}(), there's no need anymore to split accesses at
page boundaries there. Leave an assertion, though.

Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
---
v3: Transform an expression. Re-base over parameter renaming in an
    earlier patch.
v2: Replace ASSERT() by more safe construct.

--- a/xen/arch/x86/hvm/emulate.c
+++ b/xen/arch/x86/hvm/emulate.c
@@ -1084,7 +1084,7 @@ static int hvmemul_linear_mmio_access(
 {
     struct hvm_vcpu_io *hvio = &current->arch.hvm.hvm_io;
     unsigned long offset = gla & ~PAGE_MASK;
-    unsigned int chunk, buffer_offset = gla - start_gla;
+    unsigned int buffer_offset = gla - start_gla;
     struct hvm_mmio_cache *cache = hvmemul_find_mmio_cache(hvio, start_gla,
                                                            dir, buffer_offset);
     paddr_t gpa;
@@ -1094,13 +1094,17 @@ static int hvmemul_linear_mmio_access(
     if ( cache == NULL )
         return X86EMUL_UNHANDLEABLE;
 
-    chunk = min_t(unsigned int, size, PAGE_SIZE - offset);
+    if ( size + offset > PAGE_SIZE )
+    {
+        ASSERT_UNREACHABLE();
+        return X86EMUL_UNHANDLEABLE;
+    }
 
     if ( known_gpfn )
         gpa = pfn_to_paddr(hvio->mmio_gpfn) | offset;
     else
     {
-        rc = hvmemul_linear_to_phys(gla, &gpa, chunk, &one_rep, pfec,
+        rc = hvmemul_linear_to_phys(gla, &gpa, size, &one_rep, pfec,
                                     hvmemul_ctxt);
         if ( rc != X86EMUL_OKAY )
             return rc;
@@ -1108,27 +1112,8 @@ static int hvmemul_linear_mmio_access(
         latch_linear_to_phys(hvio, gla, gpa, dir == IOREQ_WRITE);
     }
 
-    for ( ;; )
-    {
-        rc = hvmemul_phys_mmio_access(cache, gpa, chunk, dir, buffer, buffer_offset);
-        if ( rc != X86EMUL_OKAY )
-            break;
-
-        gla += chunk;
-        buffer_offset += chunk;
-        size -= chunk;
-
-        if ( size == 0 )
-            break;
-
-        chunk = min_t(unsigned int, size, PAGE_SIZE);
-        rc = hvmemul_linear_to_phys(gla, &gpa, chunk, &one_rep, pfec,
-                                    hvmemul_ctxt);
-        if ( rc != X86EMUL_OKAY )
-            return rc;
-    }
-
-    return rc;
+    return hvmemul_phys_mmio_access(cache, gpa, size, dir, buffer,
+                                    buffer_offset);
 }
 
 static inline int hvmemul_linear_mmio_read(

Re: [PATCH partly-for-4.20 v3 0/4] x86/HVM: emulation (MMIO) improvements

[Oleksii Kurochko]

On 1/23/25 11:29 AM, Jan Beulich wrote:
> The main fix is patch 2, with the earlier patch setting the stage and
> the latter ones simplifying other things at least a little in exchange.
>
> 1: allocate emulation cache entries dynamically
> 2: correct read/write split at page boundaries
> 3: slightly improve CMPXCHG16B emulation
> 4: drop redundant access splitting
>
> Oleksii - the first two patches (plus the simple one from v2 that I just
> committed) are backporting material, and hence I'd like to ask for a
> release ack for at least those two.

R-Acked-by: Oleksii Kurochko <oleksii.kurochko@gmail.com> for first two 
patches.

Thanks.

~ Oleksii