[Intel-xe] [PATCH 1/6] drm/xe: Use a define to set initial seqno for fences

Thomas Hellström thomas.hellstrom at linux.intel.com
Wed Mar 15 15:32:50 UTC 2023


On 3/15/23 16:26, Matthew Brost wrote:
> On Tue, Mar 14, 2023 at 11:54:29AM +0000, Matthew Auld wrote:
>> On Mon, 13 Mar 2023 at 20:46, Thomas Hellström
>> <thomas.hellstrom at linux.intel.com> wrote:
>>> Also for HW fences, write the initial seqno - 1 to the HW completed
>>> seqno to initialize.
>>>
>>> Signed-off-by: Thomas Hellström <thomas.hellstrom at linux.intel.com>
>>> ---
>>>   drivers/gpu/drm/xe/xe_engine.c   | 5 +++--
>>>   drivers/gpu/drm/xe/xe_hw_fence.c | 5 ++---
>>>   drivers/gpu/drm/xe/xe_hw_fence.h | 2 ++
>>>   drivers/gpu/drm/xe/xe_lrc.c      | 3 +++
>>>   4 files changed, 10 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/xe/xe_engine.c b/drivers/gpu/drm/xe/xe_engine.c
>>> index 3b6f8a25112a..7e6367a84f1f 100644
>>> --- a/drivers/gpu/drm/xe/xe_engine.c
>>> +++ b/drivers/gpu/drm/xe/xe_engine.c
>>> @@ -13,6 +13,7 @@
>>>
>>>   #include "xe_device.h"
>>>   #include "xe_gt.h"
>>> +#include "xe_hw_fence.h"
>>>   #include "xe_lrc.h"
>>>   #include "xe_macros.h"
>>>   #include "xe_migrate.h"
>>> @@ -57,11 +58,11 @@ static struct xe_engine *__xe_engine_create(struct xe_device *xe,
>>>
>>>          if (xe_engine_is_parallel(e)) {
>>>                  e->parallel.composite_fence_ctx = dma_fence_context_alloc(1);
>>> -               e->parallel.composite_fence_seqno = 1;
>>> +               e->parallel.composite_fence_seqno = XE_FENCE_INITIAL_SEQNO;
>>>          }
>>>          if (e->flags & ENGINE_FLAG_VM) {
>>>                  e->bind.fence_ctx = dma_fence_context_alloc(1);
>>> -               e->bind.fence_seqno = 1;
>>> +               e->bind.fence_seqno = XE_FENCE_INITIAL_SEQNO;
>>>          }
>>>
>>>          for (i = 0; i < width; ++i) {
>>> diff --git a/drivers/gpu/drm/xe/xe_hw_fence.c b/drivers/gpu/drm/xe/xe_hw_fence.c
>>> index e56ca2867545..ca8304650713 100644
>>> --- a/drivers/gpu/drm/xe/xe_hw_fence.c
>>> +++ b/drivers/gpu/drm/xe/xe_hw_fence.c
>>> @@ -129,7 +129,7 @@ void xe_hw_fence_ctx_init(struct xe_hw_fence_ctx *ctx, struct xe_gt *gt,
>>>          ctx->gt = gt;
>>>          ctx->irq = irq;
>>>          ctx->dma_fence_ctx = dma_fence_context_alloc(1);
>>> -       ctx->next_seqno = 1;
>>> +       ctx->next_seqno = XE_FENCE_INITIAL_SEQNO;
>>>          sprintf(ctx->name, "%s", name);
>>>   }
>>>
>>> @@ -164,8 +164,7 @@ static bool xe_hw_fence_signaled(struct dma_fence *dma_fence)
>>>          struct xe_device *xe = gt_to_xe(fence->ctx->gt);
>>>          u32 seqno = xe_map_rd(xe, &fence->seqno_map, 0, u32);
>>>
>>> -       return dma_fence->error ||
>>> -               (s32)fence->dma.seqno <= (s32)seqno;
>>> +       return dma_fence->error || (s32)(u32)fence->dma.seqno <= (s32)seqno;
>> Why do we need this change? Also the hw fence dma.seqno should always
>> be u32, right?
>>
> Not exactly on topic but we should change the seqno to u64 both dma-fences and
> our hardware supports this.

I agree. Although I think we might need to be careful if we think we 
ever going to run the driver on 32-bit systems.

/Thomas


>
> Matt
>
>>>   }
>>>
>>>   static bool xe_hw_fence_enable_signaling(struct dma_fence *dma_fence)
>>> diff --git a/drivers/gpu/drm/xe/xe_hw_fence.h b/drivers/gpu/drm/xe/xe_hw_fence.h
>>> index 07f202db6526..523c2611ef5d 100644
>>> --- a/drivers/gpu/drm/xe/xe_hw_fence.h
>>> +++ b/drivers/gpu/drm/xe/xe_hw_fence.h
>>> @@ -8,6 +8,8 @@
>>>
>>>   #include "xe_hw_fence_types.h"
>>>
>>> +#define XE_FENCE_INITIAL_SEQNO 1
>>> +
>>>   int xe_hw_fence_module_init(void);
>>>   void xe_hw_fence_module_exit(void);
>>>
>>> diff --git a/drivers/gpu/drm/xe/xe_lrc.c b/drivers/gpu/drm/xe/xe_lrc.c
>>> index 9140b057a5ba..fb8c6f7d6528 100644
>>> --- a/drivers/gpu/drm/xe/xe_lrc.c
>>> +++ b/drivers/gpu/drm/xe/xe_lrc.c
>>> @@ -697,6 +697,9 @@ int xe_lrc_init(struct xe_lrc *lrc, struct xe_hw_engine *hwe,
>>>          arb_enable = MI_ARB_ON_OFF | MI_ARB_ENABLE;
>>>          xe_lrc_write_ring(lrc, &arb_enable, sizeof(arb_enable));
>>>
>>> +       map = __xe_lrc_seqno_map(lrc);
>>> +       xe_map_write32(lrc_to_xe(lrc), &map, lrc->fence_ctx.next_seqno - 1);
>>> +
>>>          return 0;
>>>
>>>   err_lrc_finish:
>>> --
>>> 2.39.2
>>>


More information about the Intel-xe mailing list