[PATCH 4/4] drm/amdgpu: Implement OS triggered MCBP(v2)

Luben Tuikov luben.tuikov at amd.com
Tue Sep 13 15:47:15 UTC 2022


Inlined:

On 2022-09-08 21:50, jiadong.zhu at amd.com wrote:
> From: "Jiadong.Zhu" <Jiadong.Zhu at amd.com>
> 
> Trigger MCBP according to the priroty of the

"priority"

Spell out MCBP here, "Mid-Command Buffer Preemption."

> software rings and the hw fence signaling
> condition.

"signalling"

> 
> The muxer records some lastest locations from the

"lastest"? ENOENT
Please use an actual word.

Run your patches through scripts/checkpatch.pl.

> software ring which is used to resubmit packages
> in preemption scenarios.
> 
> v2: update comment style
> 
> Signed-off-by: Jiadong.Zhu <Jiadong.Zhu at amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/Makefile          |   2 +-
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c       |   2 +
>  drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.c     | 101 ++++++++++++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.h     |  29 ++++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c     |  12 ++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h     |   3 +
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c | 163 ++++++++++++++++++-
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h |  16 +-
>  drivers/gpu/drm/amd/amdgpu/amdgpu_sw_ring.c  |  26 +++
>  9 files changed, 351 insertions(+), 3 deletions(-)
>  create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.c
>  create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.h
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/Makefile b/drivers/gpu/drm/amd/amdgpu/Makefile
> index 85224bc81ce5..24c5aa19bbf2 100644
> --- a/drivers/gpu/drm/amd/amdgpu/Makefile
> +++ b/drivers/gpu/drm/amd/amdgpu/Makefile
> @@ -59,7 +59,7 @@ amdgpu-y += amdgpu_device.o amdgpu_kms.o \
>  	amdgpu_umc.o smu_v11_0_i2c.o amdgpu_fru_eeprom.o amdgpu_rap.o \
>  	amdgpu_fw_attestation.o amdgpu_securedisplay.o \
>  	amdgpu_eeprom.o amdgpu_mca.o amdgpu_psp_ta.o amdgpu_lsdma.o \
> -	amdgpu_sw_ring.o amdgpu_ring_mux.o
> +	amdgpu_sw_ring.o amdgpu_ring_mux.o amdgpu_mcbp.o
>  
>  amdgpu-$(CONFIG_PROC_FS) += amdgpu_fdinfo.o
>  
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> index 258cffe3c06a..af86d87e2f3b 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> @@ -211,6 +211,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
>  		}
>  	}
>  
> +	amdgpu_ring_ib_begin(ring);
>  	if (job && ring->funcs->init_cond_exec)
>  		patch_offset = amdgpu_ring_init_cond_exec(ring);
>  
> @@ -285,6 +286,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
>  	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
>  		ring->funcs->emit_wave_limit(ring, false);
>  
> +	amdgpu_ring_ib_end(ring);
>  	amdgpu_ring_commit(ring);
>  	return 0;
>  }
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.c
> new file mode 100644
> index 000000000000..2a12101a7699
> --- /dev/null
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.c
> @@ -0,0 +1,101 @@
> +/*
> + * Copyright 2022 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + *
> + */
> +
> +#include <linux/delay.h>
> +#include <linux/kernel.h>
> +#include <linux/firmware.h>
> +#include <linux/module.h>
> +#include <linux/pci.h>
> +#include <drm/gpu_scheduler.h>
> +
> +#include "amdgpu.h"
> +#include "amdgpu_mcbp.h"
> +#include "amdgpu_ring.h"
> +
> +/* trigger mcbp and find if we need resubmit */
> +int amdgpu_mcbp_trigger_preempt(struct amdgpu_ring_mux *mux)
> +{
> +	struct amdgpu_mux_entry *e;
> +	struct amdgpu_ring *ring = NULL;
> +	int i;
> +
> +	DRM_INFO("%s in\n", __func__);
> +
> +	spin_lock(&mux->lock);
> +
> +	amdgpu_ring_preempt_ib(mux->real_ring);
> +
> +	ring = NULL;
> +	for (i = 0; i < mux->num_ring_entries; i++) {
> +		e = &mux->ring_entries[i];
> +		if (e->ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT) {
> +			ring = e->ring;
> +			break;
> +		}
> +	}
> +
> +	if (!ring) {
> +		DRM_ERROR("cannot find low priority ring\n");
> +		return -ENOENT;
> +	}
> +
> +	amdgpu_fence_process(ring);
> +
> +	DRM_INFO("after preempted ring_prio(%d) last_seq(%x) sync_seq(%x)\n",
> +		ring->hw_prio, atomic_read(&ring->fence_drv.last_seq), ring->fence_drv.sync_seq);
> +
> +	if (atomic_read(&ring->fence_drv.last_seq) !=
> +	    ring->fence_drv.sync_seq) {
> +		DRM_INFO("schedule resubmit\n");
> +		mux->s_resubmit = true;
> +		amdgpu_ring_mux_schedule_resubmit(mux);
> +	}
> +
> +	spin_unlock(&mux->lock);
> +	return 0;
> +}
> +
> +
> +/*scan on low prio rings to have unsignaled fence and high ring has no fence.*/
> +int amdgpu_mcbp_scan(struct amdgpu_ring_mux *mux)
> +{
> +	struct amdgpu_ring *ring;
> +	uint32_t seq, last_seq;
> +	int i, need_preempt;
> +
> +	need_preempt = 0;
> +	for (i = 0; i < mux->num_ring_entries; i++) {
> +		ring = mux->ring_entries[i].ring;
> +		last_seq = atomic_read(&ring->fence_drv.last_seq);
> +		seq = READ_ONCE(ring->fence_drv.sync_seq);
> +		DRM_INFO("ring(%p) prio(%d) last_seq(%x) seq(%x)\n",
> +			ring, ring->hw_prio, last_seq, seq);
> +		if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT && last_seq < seq)
> +			return 0;
> +		if (ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT && last_seq < seq)
> +			need_preempt = 1;
> +	}
> +
> +	DRM_INFO("%s return %d\n", __func__, need_preempt && !mux->s_resubmit);
> +	return need_preempt && !mux->s_resubmit;
> +}

DRM_INFO() seem here to be debug messages and I feel they should be removed.
Sometimes we enable INFO level messages and this will overflow the log.

> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.h
> new file mode 100644
> index 000000000000..0033bcba8d03
> --- /dev/null
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mcbp.h
> @@ -0,0 +1,29 @@
> +/*
> + * Copyright 2022 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + *
> + */
> +
> +#ifndef __AMDGPU_MCBP_H__
> +#define __AMDGPU_MCBP_H__
> +
> +int amdgpu_mcbp_trigger_preempt(struct amdgpu_ring_mux *mux);
> +int amdgpu_mcbp_scan(struct amdgpu_ring_mux *mux);
> +#endif
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> index 5b70a2c36d81..6d7f8a40e308 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> @@ -583,3 +583,15 @@ int amdgpu_ring_init_mqd(struct amdgpu_ring *ring)
>  
>  	return mqd_mgr->init_mqd(adev, ring->mqd_ptr, &prop);
>  }
> +
> +void amdgpu_ring_ib_begin(struct amdgpu_ring *ring)
> +{
> +	if (ring->is_sw_ring)
> +		amdgpu_sw_ring_ib_begin(ring);
> +}
> +
> +void amdgpu_ring_ib_end(struct amdgpu_ring *ring)
> +{
> +	if (ring->is_sw_ring)
> +		amdgpu_sw_ring_ib_end(ring);
> +}
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> index d3155dc86c07..399037b0d6e6 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> @@ -311,6 +311,9 @@ struct amdgpu_ring {
>  #define amdgpu_ring_preempt_ib(r) (r)->funcs->preempt_ib(r)
>  
>  int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw);
> +void amdgpu_ring_ib_begin(struct amdgpu_ring *ring);
> +void amdgpu_ring_ib_end(struct amdgpu_ring *ring);
> +
>  void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count);
>  void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
>  void amdgpu_ring_commit(struct amdgpu_ring *ring);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c
> index ea4a3c66119a..0c9b639b844e 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c
> @@ -20,28 +20,60 @@
>   * OTHER DEALINGS IN THE SOFTWARE.
>   *
>   */
> -
> +#include <linux/slab.h>
>  #include <drm/drm_print.h>
>  
>  #include "amdgpu_ring_mux.h"
> +#include "amdgpu_mcbp.h"
>  #include "amdgpu_ring.h"
>  
>  #define AMDGPU_MUX_RESUBMIT_JIFFIES_TIMEOUT (HZ/2)
>  
> +static struct kmem_cache *amdgpu_mux_chunk_slab;
> +
>  static int copy_pkt_from_sw_ring(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring,
>  	u64 s_begin, u64 s_end);
> +static void amdgpu_mux_resubmit_chunks(struct amdgpu_ring_mux *mux);
> +static void amdgpu_mux_resubmit_fallback(struct timer_list *t);
>  
>  int amdgpu_ring_mux_init(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring)
>  {
>  	mux->real_ring = ring;
> +
>  	memset(mux->ring_entries, 0, sizeof(mux->ring_entries));
>  	mux->num_ring_entries = 0;
> +
> +	mux->s_resubmit = false;
> +
> +	amdgpu_mux_chunk_slab = kmem_cache_create(
> +		"amdgpu_mux_chunk", sizeof(struct amdgpu_mux_chunk), 0,
> +		SLAB_HWCACHE_ALIGN, NULL);
> +	if (!amdgpu_mux_chunk_slab) {
> +		DRM_ERROR("create amdgpu_mux_chunk cache failed\n");
> +		return -ENOMEM;
> +	}
> +
>  	spin_lock_init(&mux->lock);
> +
> +	timer_setup(&mux->resubmit_timer, amdgpu_mux_resubmit_fallback, 0);
> +
>  	return 0;
>  }
>  
>  void amdgpu_ring_mux_fini(struct amdgpu_ring_mux *mux)
>  {
> +	struct amdgpu_mux_entry *e;
> +	struct amdgpu_mux_chunk *chunk, *chunk2;
> +	int i;
> +
> +	for (i = 0; i < mux->num_ring_entries; i++) {
> +		e = &mux->ring_entries[i];
> +		list_for_each_entry_safe(chunk, chunk2, &e->list, entry) {
> +			list_del(&chunk->entry);
> +			kmem_cache_free(amdgpu_mux_chunk_slab, chunk);
> +		}
> +	}
> +	kmem_cache_destroy(amdgpu_mux_chunk_slab);
>  	memset(mux->ring_entries, 0, sizeof(mux->ring_entries));
>  	mux->num_ring_entries = 0;
>  }
> @@ -64,6 +96,8 @@ int amdgpu_ring_mux_add_sw_ring(struct amdgpu_ring_mux *mux, struct amdgpu_ring
>  	e->sw_rptr = 0;
>  	e->sw_wptr = 0;
>  
> +	INIT_LIST_HEAD(&e->list);
> +
>  	return 0;
>  }
>  
> @@ -180,3 +214,130 @@ static int copy_pkt_from_sw_ring(struct amdgpu_ring_mux *mux, struct amdgpu_ring
>  
>  	return 0;
>  }
> +
> +void amdgpu_ring_mux_schedule_resubmit(struct amdgpu_ring_mux *mux)
> +{
> +	mod_timer(&mux->resubmit_timer, jiffies + AMDGPU_MUX_RESUBMIT_JIFFIES_TIMEOUT);
> +}
> +
> +void amdgpu_ring_mux_start_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring)
> +{
> +	struct amdgpu_mux_entry *e;
> +	struct amdgpu_mux_chunk *chunk;
> +
> +	if (mux->s_resubmit)
> +		amdgpu_mux_resubmit_chunks(mux);
> +
> +	e = amdgpu_get_sw_entry(mux, ring);
> +	if (!e) {
> +		DRM_ERROR("cannot find entry!\n");
> +		return;
> +	}
> +
> +	chunk = kmem_cache_alloc(amdgpu_mux_chunk_slab, GFP_KERNEL);
> +	if (!chunk) {
> +		DRM_ERROR("alloc amdgpu_mux_chunk_slab failed\n");
> +		return;
> +	}
> +
> +	chunk->start = ring->wptr;
> +	list_add_tail(&chunk->entry, &e->list);
> +}
> +
> +static void scan_and_remove_signaled_chunk(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring)
> +{
> +	uint32_t last_seq, size = 0;
> +	struct amdgpu_mux_entry *e;
> +	struct amdgpu_mux_chunk *chunk, *tmp;
> +
> +	e = amdgpu_get_sw_entry(mux, ring);
> +	if (!e) {
> +		DRM_ERROR("cannot find entry!\n");
> +		return;
> +	}
> +
> +	last_seq = atomic_read(&ring->fence_drv.last_seq);
> +
> +	list_for_each_entry_safe(chunk, tmp, &e->list, entry) {
> +		if (chunk->sync_seq <= last_seq) {
> +			list_del(&chunk->entry);
> +			kmem_cache_free(amdgpu_mux_chunk_slab, chunk);
> +		} else {
> +			size++;
> +		}
> +	}
> +}
> +
> +void amdgpu_ring_mux_end_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring)
> +{
> +	struct amdgpu_mux_entry *e;
> +	struct amdgpu_mux_chunk *chunk;
> +
> +	e = amdgpu_get_sw_entry(mux, ring);
> +	if (!e) {
> +		DRM_ERROR("cannot find entry!\n");
> +		return;
> +	}
> +
> +	chunk = list_last_entry(&e->list, struct amdgpu_mux_chunk, entry);
> +	if (!chunk) {
> +		DRM_ERROR("cannot find chunk!\n");
> +		return;
> +	}
> +
> +	chunk->end = ring->wptr;
> +	chunk->sync_seq = READ_ONCE(ring->fence_drv.sync_seq);
> +
> +	scan_and_remove_signaled_chunk(mux, ring);
> +}
> +
> +static void amdgpu_mux_resubmit_chunks(struct amdgpu_ring_mux *mux)
> +{
> +	struct amdgpu_mux_entry *e = NULL;
> +	struct amdgpu_mux_chunk *chunk;
> +	uint32_t seq, last_seq;
> +	int i;
> +
> +	/*find low priority entries:*/
> +	spin_lock(&mux->lock);
> +
> +	for (i = 0; i < mux->num_ring_entries; i++) {
> +		if (mux->ring_entries[i].ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT) {
> +				e = &mux->ring_entries[i];
> +			break;
> +		}
> +	}
> +
> +	if (!e) {
> +		DRM_ERROR("%s no low priority ring found\n", __func__);
> +		return;
> +	}
> +
> +	last_seq = atomic_read(&e->ring->fence_drv.last_seq);
> +	seq = READ_ONCE(e->ring->fence_drv.sync_seq);
> +	if (seq == last_seq) {
> +		DRM_INFO("skip as fence signaled seq=%x\n", seq);
> +		return;
> +	}
> +	DRM_INFO("begin to copy resubmit chunks\n");

Those two DRM_INFO() should be removed.

> +
> +	/*resubmit all the fences between (last_seq, seq]*/
> +	list_for_each_entry(chunk, &e->list, entry) {
> +		if (chunk->sync_seq > last_seq) {
> +			copy_pkt_from_sw_ring(mux, e->ring, chunk->start, chunk->end);
> +			amdgpu_ring_commit(mux->real_ring);
> +		}
> +	}
> +	spin_unlock(&mux->lock);
> +
> +	del_timer(&mux->resubmit_timer);
> +	mux->s_resubmit = false;
> +}
> +
> +static void amdgpu_mux_resubmit_fallback(struct timer_list *t)
> +{
> +	struct amdgpu_ring_mux *mux = from_timer(mux, t, resubmit_timer);
> +
> +	DRM_INFO("calling %s\n", __func__);
> +	amdgpu_mux_resubmit_chunks(mux);
> +}
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h
> index d058c43bb063..1d91c235061a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h
> @@ -44,17 +44,27 @@ struct amdgpu_mux_entry {
>  	u64 sw_cptr;
>  	u64 sw_rptr;
>  	u64 sw_wptr;
> +
> +	struct list_head list;
>  };
>  
>  struct amdgpu_ring_mux {
>  	struct amdgpu_ring *real_ring;
>  
>  	struct amdgpu_mux_entry ring_entries[AMDGPU_MAX_GFX_RINGS];
> -
>  	unsigned num_ring_entries;
>  
>  	spinlock_t			lock;
>  
> +	bool s_resubmit;
> +	struct timer_list		resubmit_timer;
> +};
> +
> +struct amdgpu_mux_chunk {
> +	struct list_head entry;
> +	uint32_t sync_seq;
> +	u64 start;
> +	u64 end;
>  };

I'd generally include a comment on struct members--makes it clear what the members
are.

>  
>  int amdgpu_ring_mux_init(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring);
> @@ -64,4 +74,8 @@ void amdgpu_ring_set_wptr_to_mux(struct amdgpu_ring_mux *mux, struct amdgpu_ring
>  u64 amdgpu_ring_get_wptr_from_mux(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring);
>  u64 amdgpu_ring_get_rptr_from_mux(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring);
>  
> +void amdgpu_ring_mux_start_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring);
> +void amdgpu_ring_mux_end_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring);
> +void amdgpu_ring_mux_schedule_resubmit(struct amdgpu_ring_mux *mux);
> +
>  #endif
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sw_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sw_ring.c
> index 452d0ff37758..143a84c18534 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sw_ring.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sw_ring.c
> @@ -26,6 +26,7 @@
>  
>  #include "amdgpu_sw_ring.h"
>  #include "amdgpu_ring_mux.h"
> +#include "amdgpu_mcbp.h"
>  
>  #define amdgpu_ring_get_gpu_addr(ring, offset)				\
>  	(ring->is_mes_queue ?						\
> @@ -202,3 +203,28 @@ void amdgpu_sw_ring_fini(struct amdgpu_ring *ring)
>  	ring->adev->rings[ring->idx] = NULL;
>  }
>  
> +void amdgpu_sw_ring_ib_begin(struct amdgpu_ring *ring)
> +{
> +	struct amdgpu_device *adev = ring->adev;
> +	struct amdgpu_ring_mux *mux = &adev->gfx.muxer;
> +
> +	BUG_ON(!ring->is_sw_ring);
> +	if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT) {
> +		if (amdgpu_mcbp_scan(mux) > 0)
> +			amdgpu_mcbp_trigger_preempt(mux);
> +		return;
> +	}
> +
> +	amdgpu_ring_mux_start_ib(mux, ring);
> +}
> +
> +void amdgpu_sw_ring_ib_end(struct amdgpu_ring *ring)
> +{
> +	struct amdgpu_device *adev = ring->adev;
> +	struct amdgpu_ring_mux *mux = &adev->gfx.muxer;
> +
> +	BUG_ON(!ring->is_sw_ring);
> +	if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT)
> +		return;
> +	amdgpu_ring_mux_end_ib(mux, ring);
> +}

Regards,
-- 
Luben


More information about the amd-gfx mailing list