Re: [PATCH 6/8] drm/radeon: cope with foreign fences inside the reservation object

[Maarten Lankhorst <maarten.lankhorst@canonical.com>]

Op 17-09-14 om 15:09 schreef Christian König:
> Am 17.09.2014 um 14:35 schrieb Maarten Lankhorst:
>> Not the whole world is a radeon! :-)
>>
>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>> ---
>>   drivers/gpu/drm/radeon/cik.c              |  2 +-
>>   drivers/gpu/drm/radeon/cik_sdma.c         |  2 +-
>>   drivers/gpu/drm/radeon/evergreen_dma.c    |  2 +-
>>   drivers/gpu/drm/radeon/r600.c             |  2 +-
>>   drivers/gpu/drm/radeon/r600_dma.c         |  2 +-
>>   drivers/gpu/drm/radeon/radeon.h           |  7 ++++---
>>   drivers/gpu/drm/radeon/radeon_cs.c        | 27 ++++++++++++++++++++-------
>>   drivers/gpu/drm/radeon/radeon_fence.c     |  3 +++
>>   drivers/gpu/drm/radeon/radeon_semaphore.c | 24 ++++++++++++++++++------
>>   drivers/gpu/drm/radeon/radeon_vm.c        |  4 ++--
>>   drivers/gpu/drm/radeon/rv770_dma.c        |  2 +-
>>   drivers/gpu/drm/radeon/si_dma.c           |  2 +-
>>   12 files changed, 54 insertions(+), 25 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/radeon/cik.c b/drivers/gpu/drm/radeon/cik.c
>> index 0d761f73a7fa..1440b6e9281e 100644
>> --- a/drivers/gpu/drm/radeon/cik.c
>> +++ b/drivers/gpu/drm/radeon/cik.c
>> @@ -3993,7 +3993,7 @@ struct radeon_fence *cik_copy_cpdma(struct radeon_device *rdev,
>>           return ERR_PTR(r);
>>       }
>>   -    radeon_semaphore_sync_resv(sem, resv, false);
>> +    radeon_semaphore_sync_resv(rdev, sem, resv, false, false);
>>       radeon_semaphore_sync_rings(rdev, sem, ring->idx);
>>         for (i = 0; i < num_loops; i++) {
>> diff --git a/drivers/gpu/drm/radeon/cik_sdma.c b/drivers/gpu/drm/radeon/cik_sdma.c
>> index c01a6100c318..315c595418ec 100644
>> --- a/drivers/gpu/drm/radeon/cik_sdma.c
>> +++ b/drivers/gpu/drm/radeon/cik_sdma.c
>> @@ -571,7 +571,7 @@ struct radeon_fence *cik_copy_dma(struct radeon_device *rdev,
>>           return ERR_PTR(r);
>>       }
>>   -    radeon_semaphore_sync_resv(sem, resv, false);
>> +    radeon_semaphore_sync_resv(rdev, sem, resv, false, false);
>>       radeon_semaphore_sync_rings(rdev, sem, ring->idx);
>>         for (i = 0; i < num_loops; i++) {
>> diff --git a/drivers/gpu/drm/radeon/evergreen_dma.c b/drivers/gpu/drm/radeon/evergreen_dma.c
>> index 946f37d0b469..5a5686792068 100644
>> --- a/drivers/gpu/drm/radeon/evergreen_dma.c
>> +++ b/drivers/gpu/drm/radeon/evergreen_dma.c
>> @@ -133,7 +133,7 @@ struct radeon_fence *evergreen_copy_dma(struct radeon_device *rdev,
>>           return ERR_PTR(r);
>>       }
>>   -    radeon_semaphore_sync_resv(sem, resv, false);
>> +    radeon_semaphore_sync_resv(rdev, sem, resv, false, false);
>>       radeon_semaphore_sync_rings(rdev, sem, ring->idx);
>>         for (i = 0; i < num_loops; i++) {
>> diff --git a/drivers/gpu/drm/radeon/r600.c b/drivers/gpu/drm/radeon/r600.c
>> index 25f367ac4637..35c22ee9bc4a 100644
>> --- a/drivers/gpu/drm/radeon/r600.c
>> +++ b/drivers/gpu/drm/radeon/r600.c
>> @@ -2912,7 +2912,7 @@ struct radeon_fence *r600_copy_cpdma(struct radeon_device *rdev,
>>           return ERR_PTR(r);
>>       }
>>   -    radeon_semaphore_sync_resv(sem, resv, false);
>> +    radeon_semaphore_sync_resv(rdev, sem, resv, false, false);
>>       radeon_semaphore_sync_rings(rdev, sem, ring->idx);
>>         radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
>> diff --git a/drivers/gpu/drm/radeon/r600_dma.c b/drivers/gpu/drm/radeon/r600_dma.c
>> index fc54224ce87b..674af8db7a35 100644
>> --- a/drivers/gpu/drm/radeon/r600_dma.c
>> +++ b/drivers/gpu/drm/radeon/r600_dma.c
>> @@ -470,7 +470,7 @@ struct radeon_fence *r600_copy_dma(struct radeon_device *rdev,
>>           return ERR_PTR(r);
>>       }
>>   -    radeon_semaphore_sync_resv(sem, resv, false);
>> +    radeon_semaphore_sync_resv(rdev, sem, resv, false, false);
>>       radeon_semaphore_sync_rings(rdev, sem, ring->idx);
>>         for (i = 0; i < num_loops; i++) {
>> diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
>> index 9aa75c1af4f4..6cdc5e62fe12 100644
>> --- a/drivers/gpu/drm/radeon/radeon.h
>> +++ b/drivers/gpu/drm/radeon/radeon.h
>> @@ -587,9 +587,10 @@ bool radeon_semaphore_emit_wait(struct radeon_device *rdev, int ring,
>>                   struct radeon_semaphore *semaphore);
>>   void radeon_semaphore_sync_fence(struct radeon_semaphore *semaphore,
>>                    struct radeon_fence *fence);
>> -void radeon_semaphore_sync_resv(struct radeon_semaphore *semaphore,
>> -                struct reservation_object *resv,
>> -                bool shared);
>> +int radeon_semaphore_sync_resv(struct radeon_device *rdev,
>> +                   struct radeon_semaphore *semaphore,
>> +                   struct reservation_object *resv,
>> +                   bool shared, bool intr);
>>   int radeon_semaphore_sync_rings(struct radeon_device *rdev,
>>                   struct radeon_semaphore *semaphore,
>>                   int waiting_ring);
>> diff --git a/drivers/gpu/drm/radeon/radeon_cs.c b/drivers/gpu/drm/radeon/radeon_cs.c
>> index f662de41ba49..a4a608c8b1ba 100644
>> --- a/drivers/gpu/drm/radeon/radeon_cs.c
>> +++ b/drivers/gpu/drm/radeon/radeon_cs.c
>> @@ -249,20 +249,21 @@ static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority
>>       return 0;
>>   }
>>   -static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
>> +static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
>>   {
>> -    int i;
>> +    int i, r = 0;
>>   -    for (i = 0; i < p->nrelocs; i++) {
>> +    for (i = 0; !r && i < p->nrelocs; i++) {
>>           struct reservation_object *resv;
>>             if (!p->relocs[i].robj)
>>               continue;
>>             resv = p->relocs[i].robj->tbo.resv;
>> -        radeon_semaphore_sync_resv(p->ib.semaphore, resv,
>> -                       p->relocs[i].tv.shared);
>> +        r = radeon_semaphore_sync_resv(p->rdev, p->ib.semaphore, resv,
>> +                           p->relocs[i].tv.shared, true);
>
> Break on the first error?
It does, see the !r in the loop.

>>       }
>> +    return r;
>>   }
>>     /* XXX: note that this is called from the legacy UMS CS ioctl as well */
>> @@ -472,13 +473,19 @@ static int radeon_cs_ib_chunk(struct radeon_device *rdev,
>>           return r;
>>       }
>>   +    r = radeon_cs_sync_rings(parser);
>> +    if (r) {
>> +        if (r != -ERESTARTSYS)
>> +            DRM_ERROR("Failed to sync rings: %i\n", r);
>> +        return r;
>> +    }
>> +
>>       if (parser->ring == R600_RING_TYPE_UVD_INDEX)
>>           radeon_uvd_note_usage(rdev);
>>       else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
>>            (parser->ring == TN_RING_TYPE_VCE2_INDEX))
>>           radeon_vce_note_usage(rdev);
>>   -    radeon_cs_sync_rings(parser);
>>       r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
>>       if (r) {
>>           DRM_ERROR("Failed to schedule IB !\n");
>> @@ -565,7 +572,13 @@ static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
>>       if (r) {
>>           goto out;
>>       }
>> -    radeon_cs_sync_rings(parser);
>> +
>> +    r = radeon_cs_sync_rings(parser);
>> +    if (r) {
>> +        if (r != -ERESTARTSYS)
>> +            DRM_ERROR("Failed to sync rings: %i\n", r);
>> +        goto out;
>> +    }
>>       radeon_semaphore_sync_fence(parser->ib.semaphore, vm->fence);
>>         if ((rdev->family >= CHIP_TAHITI) &&
>> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c
>> index af9f2d6bd7d0..0262fe2580d2 100644
>> --- a/drivers/gpu/drm/radeon/radeon_fence.c
>> +++ b/drivers/gpu/drm/radeon/radeon_fence.c
>> @@ -541,6 +541,9 @@ int radeon_fence_wait(struct radeon_fence *fence, bool intr)
>>       uint64_t seq[RADEON_NUM_RINGS] = {};
>>       long r;
>>   +    if (unlikely(!to_radeon_fence(&fence->base)))
>> +        return fence_wait(&fence->base, intr);
>> +
>>       seq[fence->ring] = fence->seq;
>>       r = radeon_fence_wait_seq_timeout(fence->rdev, seq, intr, MAX_SCHEDULE_TIMEOUT);
>>       if (r < 0) {
>> diff --git a/drivers/gpu/drm/radeon/radeon_semaphore.c b/drivers/gpu/drm/radeon/radeon_semaphore.c
>> index 4d4b0773638a..68311da39c09 100644
>> --- a/drivers/gpu/drm/radeon/radeon_semaphore.c
>> +++ b/drivers/gpu/drm/radeon/radeon_semaphore.c
>> @@ -124,27 +124,39 @@ void radeon_semaphore_sync_fence(struct radeon_semaphore *semaphore,
>>    *
>>    * Sync to the fence using this semaphore object
>>    */
>> -void radeon_semaphore_sync_resv(struct radeon_semaphore *sema,
>> -                struct reservation_object *resv,
>> -                bool shared)
>> +int radeon_semaphore_sync_resv(struct radeon_device *rdev,
>> +                   struct radeon_semaphore *sema,
>> +                   struct reservation_object *resv,
>> +                   bool shared, bool intr)
>>   {
>>       struct reservation_object_list *flist;
>>       struct fence *f;
>> +    struct radeon_fence *fence;
>>       unsigned i;
>> +    int r = 0;
>>         /* always sync to the exclusive fence */
>>       f = reservation_object_get_excl(resv);
>> -    radeon_semaphore_sync_fence(sema, (struct radeon_fence*)f);
>> +    fence = f ? to_radeon_fence(f) : NULL;
>> +    if (fence && fence->rdev == rdev)
>> +        radeon_semaphore_sync_fence(sema, fence);
>> +    else if (f)
>> +        r = fence_wait(f, intr);
>
> Error handling? If anything goes wrong while waiting on the exclusive fence we would overwrite the error with the result of the shared fences.
The if (!r) fence_wait in the shared fences takes care of that.
>>         flist = reservation_object_get_list(resv);
>>       if (shared || !flist)
>> -        return;
>> +        return r;
>>         for (i = 0; i < flist->shared_count; ++i) {
>>           f = rcu_dereference_protected(flist->shared[i],
>>                             reservation_object_held(resv));
>> -        radeon_semaphore_sync_fence(sema, (struct radeon_fence*)f);
>> +        fence = to_radeon_fence(f);
>> +        if (fence && fence->rdev == rdev)
>> +            radeon_semaphore_sync_fence(sema, fence);
>> +        else if (!r)
>> +            r = fence_wait(f, intr);
>
> Again break on the first error.
Not all the functions check for errors, so if the wait fails I wanted to sync as much as possible.

But investigating them it looks like the places that don't check for errors are the .blit, .copy and radeon_vm calls,
Those bo's cannot be exported as dma-buf, so I think the resv call can't fail in that case, and I can return immediately if the calls fails after all.
I should be able to drop the intr parameter too for now, because it means the interruptible wait can be the default here.

~Maarten