> -----Original Message-----
> From: Jan Beulich [mailto:[email protected]]
> Sent: 24 September 2018 13:16
> To: Paul Durrant <[email protected]>
> Cc: Brian Woods <[email protected]>; Suravee Suthikulpanit
> <[email protected]>; Andrew Cooper
> <[email protected]>; Roger Pau Monne <[email protected]>; Wei
> Liu <[email protected]>; Xen-devel <[email protected]>
> Subject: RE: [PATCH 1/2] x86/SVM-IOMMU: Don't opencode memcpy() in
> queue_iommu_command()
> 
> >>> On 24.09.18 at 14:09, <[email protected]> wrote:
> >>  -----Original Message-----
> >> From: Andrew Cooper
> >> Sent: 24 September 2018 13:06
> >> To: Paul Durrant <[email protected]>; Xen-devel <xen-
> >> [email protected]>
> >> Cc: Jan Beulich <[email protected]>; Wei Liu <[email protected]>;
> Roger
> >> Pau Monne <[email protected]>; Suravee Suthikulpanit
> >> <[email protected]>; Brian Woods <[email protected]>
> >> Subject: Re: [PATCH 1/2] x86/SVM-IOMMU: Don't opencode memcpy() in
> >> queue_iommu_command()
> >>
> >> On 24/09/18 12:59, Paul Durrant wrote:
> >> >> -----Original Message-----
> >> >> From: Andrew Cooper [mailto:[email protected]]
> >> >> Sent: 24 September 2018 11:55
> >> >> To: Xen-devel <[email protected]>
> >> >> Cc: Andrew Cooper <[email protected]>; Jan Beulich
> >> >> <[email protected]>; Wei Liu <[email protected]>; Roger Pau Monne
> >> >> <[email protected]>; Paul Durrant <[email protected]>;
> Suravee
> >> >> Suthikulpanit <[email protected]>; Brian Woods
> >> >> <[email protected]>
> >> >> Subject: [PATCH 1/2] x86/SVM-IOMMU: Don't opencode memcpy() in
> >> >> queue_iommu_command()
> >> >>
> >> >> In practice, this allows the compiler to replace the loop with a
> pair
> >> of
> >> >> movs.
> >> >>
> >> >> No functional change.
> >> > Well there is a potential functional change...
> >> >
> >> >> Signed-off-by: Andrew Cooper <[email protected]>
> >> >> ---
> >> >> CC: Jan Beulich <[email protected]>
> >> >> CC: Wei Liu <[email protected]>
> >> >> CC: Roger Pau MonnĂ© <[email protected]>
> >> >> CC: Paul Durrant <[email protected]>
> >> >> CC: Suravee Suthikulpanit <[email protected]>
> >> >> CC: Brian Woods <[email protected]>
> >> >> ---
> >> >>  xen/drivers/passthrough/amd/iommu_cmd.c      | 12 ++++--------
> >> >>  xen/include/asm-x86/hvm/svm/amd-iommu-defs.h |  1 -
> >> >>  2 files changed, 4 insertions(+), 9 deletions(-)
> >> >>
> >> >> diff --git a/xen/drivers/passthrough/amd/iommu_cmd.c
> >> >> b/xen/drivers/passthrough/amd/iommu_cmd.c
> >> >> index 08247fa..c6c0b4f 100644
> >> >> --- a/xen/drivers/passthrough/amd/iommu_cmd.c
> >> >> +++ b/xen/drivers/passthrough/amd/iommu_cmd.c
> >> >> @@ -24,8 +24,7 @@
> >> >>
> >> >>  static int queue_iommu_command(struct amd_iommu *iommu, u32 cmd[])
> >> >>  {
> >> >> -    u32 tail, head, *cmd_buffer;
> >> >> -    int i;
> >> >> +    uint32_t tail, head;
> >> >>
> >> >>      tail = iommu->cmd_buffer.tail;
> >> >>      if ( ++tail == iommu->cmd_buffer.entries )
> >> >> @@ -35,12 +34,9 @@ static int queue_iommu_command(struct amd_iommu
> >> *iommu,
> >> >> u32 cmd[])
> >> >>
> IOMMU_CMD_BUFFER_HEAD_OFFSET));
> >> >>      if ( head != tail )
> >> >>      {
> >> >> -        cmd_buffer = (u32 *)(iommu->cmd_buffer.buffer +
> >> >> -                             (iommu->cmd_buffer.tail *
> >> >> -                             IOMMU_CMD_BUFFER_ENTRY_SIZE));
> >> >> -
> >> >> -        for ( i = 0; i < IOMMU_CMD_BUFFER_U32_PER_ENTRY; i++ )
> >> >> -            cmd_buffer[i] = cmd[i];
> >> >> +        memcpy(iommu->cmd_buffer.buffer +
> >> >> +               (iommu->cmd_buffer.tail *
> IOMMU_CMD_BUFFER_ENTRY_SIZE),
> >> >> +               cmd, IOMMU_CMD_BUFFER_ENTRY_SIZE);
> >> > ...since the built-in memcpy may not guarantee to the copy in 4 byte
> >> chunks in ascending order.
> >>
> >> "No functional change" != "The binary is identical".
> >>
> >> The functionality of queue_iommu_command() does not change, even if
> it's
> >> code generation does.  It is just copying bytes into a shared ring,
> >> bounded later by updating the tail pointer.
> >
> > Yes, my point is that the ring is shared and so DMA by the h/w may race.
> > This is clearly not a good situation but the fact that the code
> generation
> > may change may have side effects.
> 
> All writes to the ring occur strictly before the update of the tail
> pointer

...unless the compiler decides to re-order. There's no barrier.

  Paul

> (in MMIO), no matter how the copying gets done.
> 
> Jan

_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel

Reply via email to