Skip to contentSkip to conversation

vhost-user: fix VHOST_USER_ADD/REM_MEM_REG truncation

PATCH In review Stefan Hajnoczi 3 participants 4 messages Nov 9, 2020 – Nov 10, 2020

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 vhost-user: fix VHOST_USER_ADD/REM_MEM_REG truncation

QEMU currently truncates the mmap_offset field when sending VHOST_USER_ADD_MEM_REG and VHOST_USER_REM_MEM_REG messages. The struct layout looks like this:

  typedef struct VhostUserMemoryRegion {
      uint64_t guest_phys_addr;
      uint64_t memory_size;
      uint64_t userspace_addr;
      uint64_t mmap_offset;
  } VhostUserMemoryRegion;

  typedef struct VhostUserMemRegMsg {
      uint32_t padding;
      /* WARNING: there is a 32-bit hole here! */
      VhostUserMemoryRegion region;
  } VhostUserMemRegMsg;

The payload size is calculated as follows when sending the message in hw/virtio/vhost-user.c:

  msg->hdr.size = sizeof(msg->payload.mem_reg.padding) +
      sizeof(VhostUserMemoryRegion);

This calculation produces an incorrect result of only 36 bytes. sizeof(VhostUserMemRegMsg) is actually 40 bytes.

The consequence of this is that the final field, mmap_offset, is truncated. This breaks x86_64 TCG guests on s390 hosts. Other guest/host combinations may get lucky if either of the following holds:
1. The guest memory layout does not need mmap_offset != 0.
2. The host is little-endian and mmap_offset <= 0xffffffff so the truncation has no effect.

Fix this by extending the existing 32-bit padding field to 64-bit. Now the padding reflects the actual compiler padding. This can be verified using pahole(1).

Also document the layout properly in the vhost-user specification. The vhost-user spec did not document the exact layout. It would be impossible to implement the spec without looking at the QEMU source code.

Existing vhost-user frontends and device backends continue to work after this fix has been applied. The only change in the wire protocol is that QEMU now sets hdr.size to 40 instead of 36. If a vhost-user implementation has a hardcoded size check for 36 bytes, then it will fail with new QEMUs. Both QEMU and DPDK/SPDK don't check the exact payload size, so they continue to work.

Fixes f1aeb14b0809e313c74244d838645ed25e85ea63 ("Transmit vhost-user memory regions individually")
Cc Raphael Norwitz <[email protected]>
Cc Cornelia Huck <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Christian Borntraeger <[email protected]>
---
 docs/interop/vhost-user.rst           | 21 +++++++++++++++++++--
 contrib/libvhost-user/libvhost-user.h |  2 +-
 hw/virtio/vhost-user.c                |  5 ++---
 3 files changed, 22 insertions(+), 6 deletions(-)
docs/interop/vhost-user.rst +19 -2
index 988f154144..6d4025ba6a 100644
@@ -131,6 +131,23 @@ A region is:
 
 :mmap offset: 64-bit offset where region starts in the mapped memory
 
+Single memory region description
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
++---------+---------------+------+--------------+-------------+
+| padding | guest address | size | user address | mmap offset |
++---------+---------------+------+--------------+-------------+
+
+:padding: 64-bit
+
+:guest address: a 64-bit guest address of the region
+
+:size: a 64-bit size
+
+:user address: a 64-bit user address
+
+:mmap offset: 64-bit offset where region starts in the mapped memory
+
 Log description
 ^^^^^^^^^^^^^^^
 
@@ -1281,7 +1298,7 @@ Master message types
 ``VHOST_USER_ADD_MEM_REG``
   :id: 37
   :equivalent ioctl: N/A
-  :slave payload: memory region
+  :slave payload: single memory region description
 
   When the ``VHOST_USER_PROTOCOL_F_CONFIGURE_MEM_SLOTS`` protocol
   feature has been successfully negotiated, this message is submitted
@@ -1296,7 +1313,7 @@ Master message types
 ``VHOST_USER_REM_MEM_REG``
   :id: 38
   :equivalent ioctl: N/A
-  :slave payload: memory region
+  :slave payload: single memory region description
 
   When the ``VHOST_USER_PROTOCOL_F_CONFIGURE_MEM_SLOTS`` protocol
   feature has been successfully negotiated, this message is submitted
contrib/libvhost-user/libvhost-user.h +1 -1
index a1539dbb69..7d47f1364a 100644
@@ -136,7 +136,7 @@ typedef struct VhostUserMemory {
 } VhostUserMemory;
 
 typedef struct VhostUserMemRegMsg {
-    uint32_t padding;
+    uint64_t padding;
     VhostUserMemoryRegion region;
 } VhostUserMemRegMsg;
 
hw/virtio/vhost-user.c +2 -3
index 9c5b4f7fbc..2fdd5daf74 100644
@@ -149,7 +149,7 @@ typedef struct VhostUserMemory {
 } VhostUserMemory;
 
 typedef struct VhostUserMemRegMsg {
-    uint32_t padding;
+    uint64_t padding;
     VhostUserMemoryRegion region;
 } VhostUserMemRegMsg;
 
@@ -800,8 +800,7 @@ static int vhost_user_add_remove_regions(struct vhost_dev *dev,
     uint64_t shadow_pcb[VHOST_USER_MAX_RAM_SLOTS] = {};
     int nr_add_reg, nr_rem_reg;
 
-    msg->hdr.size = sizeof(msg->payload.mem_reg.padding) +
-        sizeof(VhostUserMemoryRegion);
+    msg->hdr.size = sizeof(msg->payload.mem_reg);
 
     /* Find the regions which need to be removed or added. */
     scrub_shadow_regions(dev, add_reg, &nr_add_reg, rem_reg, &nr_rem_reg,
Signature · 1 line
2.28.0
Reviewed-by on v1 1/1
#1 lore raw
Quoted 42 lines · On Mon, 9 Nov 2020 17:43:55 +0000 Stefan Hajnoczi <[email protected]> wrote:

QEMU currently truncates the mmap_offset field when sending VHOST_USER_ADD_MEM_REG and VHOST_USER_REM_MEM_REG messages. The struct layout looks like this:

  typedef struct VhostUserMemoryRegion {
      uint64_t guest_phys_addr;
      uint64_t memory_size;
      uint64_t userspace_addr;
      uint64_t mmap_offset;
  } VhostUserMemoryRegion;

  typedef struct VhostUserMemRegMsg {
      uint32_t padding;
      /* WARNING: there is a 32-bit hole here! */
      VhostUserMemoryRegion region;
  } VhostUserMemRegMsg;

The payload size is calculated as follows when sending the message in hw/virtio/vhost-user.c:

  msg->hdr.size = sizeof(msg->payload.mem_reg.padding) +
      sizeof(VhostUserMemoryRegion);

This calculation produces an incorrect result of only 36 bytes. sizeof(VhostUserMemRegMsg) is actually 40 bytes.

The consequence of this is that the final field, mmap_offset, is truncated. This breaks x86_64 TCG guests on s390 hosts. Other guest/host combinations may get lucky if either of the following holds:
1. The guest memory layout does not need mmap_offset != 0.
2. The host is little-endian and mmap_offset <= 0xffffffff so the truncation has no effect.

Fix this by extending the existing 32-bit padding field to 64-bit. Now the padding reflects the actual compiler padding. This can be verified using pahole(1).

Also document the layout properly in the vhost-user specification. The vhost-user spec did not document the exact layout. It would be impossible to implement the spec without looking at the QEMU source code.

Quoted 6 lines

Existing vhost-user frontends and device backends continue to work after this fix has been applied. The only change in the wire protocol is that QEMU now sets hdr.size to 40 instead of 36. If a vhost-user implementation has a hardcoded size check for 36 bytes, then it will fail with new QEMUs. Both QEMU and DPDK/SPDK don't check the exact payload size, so they continue to work.

Seems we are lucky, then.

Quoted 2 lines
Fixes f1aeb14b0809e313c74244d838645ed25e85ea63 ("Transmit vhost-user memory regions individually")

I think the canonical format is

Fixes f1aeb14b0809 ("Transmit vhost-user memory regions individually")

Maybe cc:stable as well?

Quoted 10 lines
Cc Raphael Norwitz <[email protected]>
Cc Cornelia Huck <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Christian Borntraeger <[email protected]>
---
 docs/interop/vhost-user.rst           | 21 +++++++++++++++++++--
 contrib/libvhost-user/libvhost-user.h |  2 +-
 hw/virtio/vhost-user.c                |  5 ++---
 3 files changed, 22 insertions(+), 6 deletions(-)
Reviewed-by on v1 1/1
#3 lore raw
In reply to Stefan Hajnoczi
Quoted 56 lines · Stefan Hajnoczi

QEMU currently truncates the mmap_offset field when sending VHOST_USER_ADD_MEM_REG and VHOST_USER_REM_MEM_REG messages. The struct layout looks like this:

  typedef struct VhostUserMemoryRegion {
      uint64_t guest_phys_addr;
      uint64_t memory_size;
      uint64_t userspace_addr;
      uint64_t mmap_offset;
  } VhostUserMemoryRegion;

  typedef struct VhostUserMemRegMsg {
      uint32_t padding;
      /* WARNING: there is a 32-bit hole here! */
      VhostUserMemoryRegion region;
  } VhostUserMemRegMsg;

The payload size is calculated as follows when sending the message in hw/virtio/vhost-user.c:

  msg->hdr.size = sizeof(msg->payload.mem_reg.padding) +
      sizeof(VhostUserMemoryRegion);

This calculation produces an incorrect result of only 36 bytes. sizeof(VhostUserMemRegMsg) is actually 40 bytes.

The consequence of this is that the final field, mmap_offset, is truncated. This breaks x86_64 TCG guests on s390 hosts. Other guest/host combinations may get lucky if either of the following holds:
1. The guest memory layout does not need mmap_offset != 0.
2. The host is little-endian and mmap_offset <= 0xffffffff so the truncation has no effect.

Fix this by extending the existing 32-bit padding field to 64-bit. Now the padding reflects the actual compiler padding. This can be verified using pahole(1).

Also document the layout properly in the vhost-user specification. The vhost-user spec did not document the exact layout. It would be impossible to implement the spec without looking at the QEMU source code.

Existing vhost-user frontends and device backends continue to work after this fix has been applied. The only change in the wire protocol is that QEMU now sets hdr.size to 40 instead of 36. If a vhost-user implementation has a hardcoded size check for 36 bytes, then it will fail with new QEMUs. Both QEMU and DPDK/SPDK don't check the exact payload size, so they continue to work.

Fixes f1aeb14b0809e313c74244d838645ed25e85ea63 ("Transmit vhost-user memory regions individually")
Cc Raphael Norwitz <[email protected]>
Cc Cornelia Huck <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Christian Borntraeger <[email protected]>
Quoted 6 lines
---
 docs/interop/vhost-user.rst           | 21 +++++++++++++++++++--
 contrib/libvhost-user/libvhost-user.h |  2 +-
 hw/virtio/vhost-user.c                |  5 ++---
 3 files changed, 22 insertions(+), 6 deletions(-)
#2 lore raw
In reply to Cornelia Huck
Quoted 7 lines · Cornelia Huck
Quoted 1 line · On Mon, 9 Nov 2020 17:43:55 +0000 Stefan Hajnoczi <[email protected]> wrote:
Fixes f1aeb14b0809e313c74244d838645ed25e85ea63 ("Transmit vhost-user memory regions individually")

I think the canonical format is

Fixes f1aeb14b0809 ("Transmit vhost-user memory regions individually")

According to https://wiki.qemu.org/Contribute/SubmitAPatch:
"Fixes: <at-least-12-digits-of-SHA-commit-id> ("Fixed commit subject")"

Quoted 1 line

Maybe cc:stable as well?

Good point, this issue started in QEMU 5.1.