From ded946f3f6399003ea0bdcc8911bc2fc3a7313c6 Mon Sep 17 00:00:00 2001 From: "Jiadong.Zhu" Date: Wed, 7 Sep 2022 09:40:47 +0800 Subject: drm/amdgpu: Introduce gfx software ring (v9) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The software ring is created to support priority context while there is only one hardware queue for gfx. Every software ring has its fence driver and could be used as an ordinary ring for the GPU scheduler. Multiple software rings are bound to a real ring with the ring muxer. The packages committed on the software ring are copied to the real ring. v2: Use array to store software ring entry. v3: Remove unnecessary prints. v4: Remove amdgpu_ring_sw_init/fini functions, using gtt for sw ring buffer for later dma copy optimization. v5: Allocate ring entry dynamically in the muxer. v6: Update comments for the ring muxer. v7: Modify for function naming. v8: Combine software ring functions into amdgpu_ring_mux.c v9: Use kernel-doc comment on the get_rptr function. Cc: Christian Koenig Cc: Luben Tuikov Cc: Andrey Grodzovsky Cc: Michel Dänzer Signed-off-by: Jiadong.Zhu Acked-by: Huang Rui Acked-by: Luben Tuikov Acked-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c | 221 +++++++++++++++++++++++++++ 1 file changed, 221 insertions(+) create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c new file mode 100644 index 000000000000..6fbf71451e29 --- /dev/null +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c @@ -0,0 +1,221 @@ +/* + * 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 +#include + +#include "amdgpu_ring_mux.h" +#include "amdgpu_ring.h" +#include "amdgpu.h" + +#define AMDGPU_MUX_RESUBMIT_JIFFIES_TIMEOUT (HZ / 2) + +int amdgpu_ring_mux_init(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, + unsigned int entry_size) +{ + mux->real_ring = ring; + mux->num_ring_entries = 0; + mux->ring_entry = kcalloc(entry_size, sizeof(struct amdgpu_mux_entry), GFP_KERNEL); + if (!mux->ring_entry) + return -ENOMEM; + + mux->ring_entry_size = entry_size; + spin_lock_init(&mux->lock); + + return 0; +} + +void amdgpu_ring_mux_fini(struct amdgpu_ring_mux *mux) +{ + kfree(mux->ring_entry); + mux->ring_entry = NULL; + mux->num_ring_entries = 0; + mux->ring_entry_size = 0; +} + +int amdgpu_ring_mux_add_sw_ring(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring) +{ + struct amdgpu_mux_entry *e; + + if (mux->num_ring_entries >= mux->ring_entry_size) { + DRM_ERROR("add sw ring exceeding max entry size\n"); + return -ENOENT; + } + + e = &mux->ring_entry[mux->num_ring_entries]; + ring->entry_index = mux->num_ring_entries; + e->ring = ring; + + mux->num_ring_entries += 1; + return 0; +} + +static inline struct amdgpu_mux_entry *amdgpu_ring_mux_sw_entry(struct amdgpu_ring_mux *mux, + struct amdgpu_ring *ring) +{ + return ring->entry_index < mux->ring_entry_size ? + &mux->ring_entry[ring->entry_index] : NULL; +} + +/* copy packages on sw ring range[begin, end) */ +static void amdgpu_ring_mux_copy_pkt_from_sw_ring(struct amdgpu_ring_mux *mux, + struct amdgpu_ring *ring, + u64 s_start, u64 s_end) +{ + u64 start, end; + struct amdgpu_ring *real_ring = mux->real_ring; + + start = s_start & ring->buf_mask; + end = s_end & ring->buf_mask; + + if (start == end) { + DRM_ERROR("no more data copied from sw ring\n"); + return; + } + if (start > end) { + amdgpu_ring_alloc(real_ring, (ring->ring_size >> 2) + end - start); + amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], + (ring->ring_size >> 2) - start); + amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[0], end); + } else { + amdgpu_ring_alloc(real_ring, end - start); + amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], end - start); + } +} + +void amdgpu_ring_mux_set_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, u64 wptr) +{ + struct amdgpu_mux_entry *e; + + e = amdgpu_ring_mux_sw_entry(mux, ring); + if (!e) { + DRM_ERROR("cannot find entry for sw ring\n"); + return; + } + + spin_lock(&mux->lock); + e->sw_cptr = e->sw_wptr; + e->sw_wptr = wptr; + e->start_ptr_in_hw_ring = mux->real_ring->wptr; + + amdgpu_ring_mux_copy_pkt_from_sw_ring(mux, ring, e->sw_cptr, wptr); + e->end_ptr_in_hw_ring = mux->real_ring->wptr; + amdgpu_ring_commit(mux->real_ring); + + spin_unlock(&mux->lock); +} + +u64 amdgpu_ring_mux_get_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring) +{ + struct amdgpu_mux_entry *e; + + e = amdgpu_ring_mux_sw_entry(mux, ring); + if (!e) { + DRM_ERROR("cannot find entry for sw ring\n"); + return 0; + } + + return e->sw_wptr; +} + +/** + * amdgpu_ring_mux_get_rptr - get the readptr of the software ring + * @mux: the multiplexer the software rings attach to + * @ring: the software ring of which we calculate the readptr + * + * The return value of the readptr is not precise while the other rings could + * write data onto the real ring buffer.After overwriting on the real ring, we + * can not decide if our packages have been excuted or not read yet. However, + * this function is only called by the tools such as umr to collect the latest + * packages for the hang analysis. We assume the hang happens near our latest + * submit. Thus we could use the following logic to give the clue: + * If the readptr is between start and end, then we return the copy pointer + * plus the distance from start to readptr. If the readptr is before start, we + * return the copy pointer. Lastly, if the readptr is past end, we return the + * write pointer. + */ +u64 amdgpu_ring_mux_get_rptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring) +{ + struct amdgpu_mux_entry *e; + u64 readp, offset, start, end; + + e = amdgpu_ring_mux_sw_entry(mux, ring); + if (!e) { + DRM_ERROR("no sw entry found!\n"); + return 0; + } + + readp = amdgpu_ring_get_rptr(mux->real_ring); + + start = e->start_ptr_in_hw_ring & mux->real_ring->buf_mask; + end = e->end_ptr_in_hw_ring & mux->real_ring->buf_mask; + if (start > end) { + if (readp <= end) + readp += mux->real_ring->ring_size >> 2; + end += mux->real_ring->ring_size >> 2; + } + + if (start <= readp && readp <= end) { + offset = readp - start; + e->sw_rptr = (e->sw_cptr + offset) & ring->buf_mask; + } else if (readp < start) { + e->sw_rptr = e->sw_cptr; + } else { + /* end < readptr */ + e->sw_rptr = e->sw_wptr; + } + + return e->sw_rptr; +} + +u64 amdgpu_sw_ring_get_rptr_gfx(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + struct amdgpu_ring_mux *mux = &adev->gfx.muxer; + + WARN_ON(!ring->is_sw_ring); + return amdgpu_ring_mux_get_rptr(mux, ring); +} + +u64 amdgpu_sw_ring_get_wptr_gfx(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + struct amdgpu_ring_mux *mux = &adev->gfx.muxer; + + WARN_ON(!ring->is_sw_ring); + return amdgpu_ring_mux_get_wptr(mux, ring); +} + +void amdgpu_sw_ring_set_wptr_gfx(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + struct amdgpu_ring_mux *mux = &adev->gfx.muxer; + + WARN_ON(!ring->is_sw_ring); + amdgpu_ring_mux_set_wptr(mux, ring, ring->wptr); +} + +/* Override insert_nop to prevent emitting nops to the software rings */ +void amdgpu_sw_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) +{ + WARN_ON(!ring->is_sw_ring); +} -- cgit From 0c97a19aa9940491f0374166fb2f2c47667b4e3d Mon Sep 17 00:00:00 2001 From: "Jiadong.Zhu" Date: Wed, 7 Sep 2022 10:03:10 +0800 Subject: drm/amdgpu: Add software ring callbacks for gfx9 (v8) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Set ring functions with software ring callbacks on gfx9. The software ring could be tested by debugfs_test_ib case. v2: Set sw_ring 2 to enable software ring by default. v3: Remove the parameter for software ring enablement. v4: Use amdgpu_ring_init/fini for software rings. v5: Update for code format. Fix conflict. v6: Remove unnecessary checks and enable software ring on gfx9 by default. v7: Use static array for software ring names and priorities. v8: Stop creating software rings if no gfx ring existed. Cc: Christian Koenig Cc: Luben Tuikov Cc: Andrey Grodzovsky Cc: Michel Dänzer Cc: Likun Gao Signed-off-by: Jiadong.Zhu Acked-by: Luben Tuikov Acked-by: Huang Rui Acked-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c | 20 +++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h | 2 + drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 113 ++++++++++++++++++++++++++- 5 files changed, 136 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h index 5610c53080a1..b3df4787877e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h @@ -354,6 +354,7 @@ struct amdgpu_gfx { bool is_poweron; + struct amdgpu_ring sw_gfx_ring[AMDGPU_MAX_SW_GFX_RINGS]; struct amdgpu_ring_mux muxer; }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h index 8be51ebfedd5..a744aa9bac95 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h @@ -39,6 +39,7 @@ struct amdgpu_vm; #define AMDGPU_MAX_RINGS 28 #define AMDGPU_MAX_HWIP_RINGS 8 #define AMDGPU_MAX_GFX_RINGS 2 +#define AMDGPU_MAX_SW_GFX_RINGS 2 #define AMDGPU_MAX_COMPUTE_RINGS 8 #define AMDGPU_MAX_VCE_RINGS 3 #define AMDGPU_MAX_UVD_ENC_RINGS 2 diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c index 6fbf71451e29..a52c4e7e8c39 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c @@ -29,6 +29,14 @@ #define AMDGPU_MUX_RESUBMIT_JIFFIES_TIMEOUT (HZ / 2) +static const struct ring_info { + unsigned int hw_pio; + const char *ring_name; +} sw_ring_info[] = { + { AMDGPU_RING_PRIO_DEFAULT, "gfx_low"}, + { AMDGPU_RING_PRIO_2, "gfx_high"}, +}; + int amdgpu_ring_mux_init(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, unsigned int entry_size) { @@ -219,3 +227,15 @@ void amdgpu_sw_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) { WARN_ON(!ring->is_sw_ring); } + +const char *amdgpu_sw_ring_name(int idx) +{ + return idx < ARRAY_SIZE(sw_ring_info) ? + sw_ring_info[idx].ring_name : NULL; +} + +unsigned int amdgpu_sw_ring_priority(int idx) +{ + return idx < ARRAY_SIZE(sw_ring_info) ? + sw_ring_info[idx].hw_pio : AMDGPU_RING_PRIO_DEFAULT; +} diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h index f58672a8e0ad..f67970dc3dbd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h @@ -73,4 +73,6 @@ void amdgpu_sw_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count); void amdgpu_sw_ring_ib_begin(struct amdgpu_ring *ring); void amdgpu_sw_ring_ib_end(struct amdgpu_ring *ring); +const char *amdgpu_sw_ring_name(int idx); +unsigned int amdgpu_sw_ring_priority(int idx); #endif diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c index 676832da75eb..cc28a087c3d1 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c @@ -47,6 +47,7 @@ #include "amdgpu_ras.h" +#include "amdgpu_ring_mux.h" #include "gfx_v9_4.h" #include "gfx_v9_0.h" #include "gfx_v9_4_2.h" @@ -56,6 +57,7 @@ #include "asic_reg/gc/gc_9_0_default.h" #define GFX9_NUM_GFX_RINGS 1 +#define GFX9_NUM_SW_GFX_RINGS 2 #define GFX9_MEC_HPD_SIZE 4096 #define RLCG_UCODE_LOADING_START_ADDRESS 0x00002000L #define RLC_SAVE_RESTORE_ADDR_STARTING_OFFSET 0x00000000L @@ -2103,6 +2105,7 @@ static int gfx_v9_0_sw_init(void *handle) struct amdgpu_ring *ring; struct amdgpu_kiq *kiq; struct amdgpu_device *adev = (struct amdgpu_device *)handle; + unsigned int hw_prio; switch (adev->ip_versions[GC_HWIP][0]) { case IP_VERSION(9, 0, 1): @@ -2186,6 +2189,9 @@ static int gfx_v9_0_sw_init(void *handle) sprintf(ring->name, "gfx_%d", i); ring->use_doorbell = true; ring->doorbell_index = adev->doorbell_index.gfx_ring0 << 1; + + /* disable scheduler on the real ring */ + ring->no_scheduler = true; r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP, AMDGPU_RING_PRIO_DEFAULT, NULL); @@ -2193,6 +2199,41 @@ static int gfx_v9_0_sw_init(void *handle) return r; } + /* set up the software rings */ + if (adev->gfx.num_gfx_rings) { + for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++) { + ring = &adev->gfx.sw_gfx_ring[i]; + ring->ring_obj = NULL; + sprintf(ring->name, amdgpu_sw_ring_name(i)); + ring->use_doorbell = true; + ring->doorbell_index = adev->doorbell_index.gfx_ring0 << 1; + ring->is_sw_ring = true; + hw_prio = amdgpu_sw_ring_priority(i); + r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, + AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP, hw_prio, + NULL); + if (r) + return r; + ring->wptr = 0; + } + + /* init the muxer and add software rings */ + r = amdgpu_ring_mux_init(&adev->gfx.muxer, &adev->gfx.gfx_ring[0], + GFX9_NUM_SW_GFX_RINGS); + if (r) { + DRM_ERROR("amdgpu_ring_mux_init failed(%d)\n", r); + return r; + } + for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++) { + r = amdgpu_ring_mux_add_sw_ring(&adev->gfx.muxer, + &adev->gfx.sw_gfx_ring[i]); + if (r) { + DRM_ERROR("amdgpu_ring_mux_add_sw_ring failed(%d)\n", r); + return r; + } + } + } + /* set up the compute queues - allocate horizontally across pipes */ ring_id = 0; for (i = 0; i < adev->gfx.mec.num_mec; ++i) { @@ -2243,6 +2284,12 @@ static int gfx_v9_0_sw_fini(void *handle) int i; struct amdgpu_device *adev = (struct amdgpu_device *)handle; + if (adev->gfx.num_gfx_rings) { + for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++) + amdgpu_ring_fini(&adev->gfx.sw_gfx_ring[i]); + amdgpu_ring_mux_fini(&adev->gfx.muxer); + } + for (i = 0; i < adev->gfx.num_gfx_rings; i++) amdgpu_ring_fini(&adev->gfx.gfx_ring[i]); for (i = 0; i < adev->gfx.num_compute_rings; i++) @@ -5712,7 +5759,11 @@ static int gfx_v9_0_eop_irq(struct amdgpu_device *adev, switch (me_id) { case 0: - amdgpu_fence_process(&adev->gfx.gfx_ring[0]); + /* Fence signals are handled on the software rings*/ + if (adev->gfx.num_gfx_rings) { + for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++) + amdgpu_fence_process(&adev->gfx.sw_gfx_ring[i]); + } break; case 1: case 2: @@ -6717,6 +6768,61 @@ static const struct amdgpu_ring_funcs gfx_v9_0_ring_funcs_gfx = { .emit_mem_sync = gfx_v9_0_emit_mem_sync, }; +static const struct amdgpu_ring_funcs gfx_v9_0_sw_ring_funcs_gfx = { + .type = AMDGPU_RING_TYPE_GFX, + .align_mask = 0xff, + .nop = PACKET3(PACKET3_NOP, 0x3FFF), + .support_64bit_ptrs = true, + .secure_submission_supported = true, + .vmhub = AMDGPU_GFXHUB_0, + .get_rptr = amdgpu_sw_ring_get_rptr_gfx, + .get_wptr = amdgpu_sw_ring_get_wptr_gfx, + .set_wptr = amdgpu_sw_ring_set_wptr_gfx, + .emit_frame_size = /* totally 242 maximum if 16 IBs */ + 5 + /* COND_EXEC */ + 7 + /* PIPELINE_SYNC */ + SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + + SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + + 2 + /* VM_FLUSH */ + 8 + /* FENCE for VM_FLUSH */ + 20 + /* GDS switch */ + 4 + /* double SWITCH_BUFFER, + * the first COND_EXEC jump to the place just + * prior to this double SWITCH_BUFFER + */ + 5 + /* COND_EXEC */ + 7 + /* HDP_flush */ + 4 + /* VGT_flush */ + 14 + /* CE_META */ + 31 + /* DE_META */ + 3 + /* CNTX_CTRL */ + 5 + /* HDP_INVL */ + 8 + 8 + /* FENCE x2 */ + 2 + /* SWITCH_BUFFER */ + 7, /* gfx_v9_0_emit_mem_sync */ + .emit_ib_size = 4, /* gfx_v9_0_ring_emit_ib_gfx */ + .emit_ib = gfx_v9_0_ring_emit_ib_gfx, + .emit_fence = gfx_v9_0_ring_emit_fence, + .emit_pipeline_sync = gfx_v9_0_ring_emit_pipeline_sync, + .emit_vm_flush = gfx_v9_0_ring_emit_vm_flush, + .emit_gds_switch = gfx_v9_0_ring_emit_gds_switch, + .emit_hdp_flush = gfx_v9_0_ring_emit_hdp_flush, + .test_ring = gfx_v9_0_ring_test_ring, + .test_ib = gfx_v9_0_ring_test_ib, + .insert_nop = amdgpu_sw_ring_insert_nop, + .pad_ib = amdgpu_ring_generic_pad_ib, + .emit_switch_buffer = gfx_v9_ring_emit_sb, + .emit_cntxcntl = gfx_v9_ring_emit_cntxcntl, + .init_cond_exec = gfx_v9_0_ring_emit_init_cond_exec, + .patch_cond_exec = gfx_v9_0_ring_emit_patch_cond_exec, + .emit_frame_cntl = gfx_v9_0_ring_emit_frame_cntl, + .emit_wreg = gfx_v9_0_ring_emit_wreg, + .emit_reg_wait = gfx_v9_0_ring_emit_reg_wait, + .emit_reg_write_reg_wait = gfx_v9_0_ring_emit_reg_write_reg_wait, + .soft_recovery = gfx_v9_0_ring_soft_recovery, + .emit_mem_sync = gfx_v9_0_emit_mem_sync, +}; + static const struct amdgpu_ring_funcs gfx_v9_0_ring_funcs_compute = { .type = AMDGPU_RING_TYPE_COMPUTE, .align_mask = 0xff, @@ -6794,6 +6900,11 @@ static void gfx_v9_0_set_ring_funcs(struct amdgpu_device *adev) for (i = 0; i < adev->gfx.num_gfx_rings; i++) adev->gfx.gfx_ring[i].funcs = &gfx_v9_0_ring_funcs_gfx; + if (adev->gfx.num_gfx_rings) { + for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++) + adev->gfx.sw_gfx_ring[i].funcs = &gfx_v9_0_sw_ring_funcs_gfx; + } + for (i = 0; i < adev->gfx.num_compute_rings; i++) adev->gfx.compute_ring[i].funcs = &gfx_v9_0_ring_funcs_compute; } -- cgit From 3f4c175d62d89819121cbbd5a0a30f4b80862025 Mon Sep 17 00:00:00 2001 From: "Jiadong.Zhu" Date: Wed, 7 Sep 2022 10:24:37 +0800 Subject: drm/amdgpu: MCBP based on DRM scheduler (v9) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Trigger Mid-Command Buffer Preemption according to the priority of the software rings and the hw fence signalling condition. The muxer saves the locations of the indirect buffer frames from the software ring together with the fence sequence number in its fifo queue, and pops out those records when the fences are signalled. The locations are used to resubmit packages in preemption scenarios by coping the chunks from the software ring. v2: Update comment style. v3: Fix conflict caused by previous modifications. v4: Remove unnecessary prints. v5: Fix corner cases for resubmission cases. v6: Refactor functions for resubmission, calling fence_process in irq handler. v7: Solve conflict for removing amdgpu_sw_ring.c. v8: Add time threshold to judge if preemption request is needed. v9: Correct comment spelling. Set fence emit timestamp before rsu assignment. Cc: Christian Koenig Cc: Luben Tuikov Cc: Andrey Grodzovsky Cc: Michel Dänzer Signed-off-by: Jiadong.Zhu Acked-by: Luben Tuikov Acked-by: Huang Rui Acked-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 54 +++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 12 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 8 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c | 351 ++++++++++++++++++++++++--- drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h | 29 ++- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 + drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 7 +- 8 files changed, 422 insertions(+), 43 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index d0d99ed607dd..00444203220d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -55,6 +55,7 @@ struct amdgpu_fence { /* RB, DMA, etc. */ struct amdgpu_ring *ring; + ktime_t start_timestamp; }; static struct kmem_cache *amdgpu_fence_slab; @@ -199,6 +200,8 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd } } + to_amdgpu_fence(fence)->start_timestamp = ktime_get(); + /* This function can't be called concurrently anyway, otherwise * emitting the fence would mess up the hardware ring buffer. */ @@ -406,6 +409,57 @@ unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring) return lower_32_bits(emitted); } +/** + * amdgpu_fence_last_unsignaled_time_us - the time fence emitted until now + * @ring: ring the fence is associated with + * + * Find the earliest fence unsignaled until now, calculate the time delta + * between the time fence emitted and now. + */ +u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring) +{ + struct amdgpu_fence_driver *drv = &ring->fence_drv; + struct dma_fence *fence; + uint32_t last_seq, sync_seq; + + last_seq = atomic_read(&ring->fence_drv.last_seq); + sync_seq = READ_ONCE(ring->fence_drv.sync_seq); + if (last_seq == sync_seq) + return 0; + + ++last_seq; + last_seq &= drv->num_fences_mask; + fence = drv->fences[last_seq]; + if (!fence) + return 0; + + return ktime_us_delta(ktime_get(), + to_amdgpu_fence(fence)->start_timestamp); +} + +/** + * amdgpu_fence_update_start_timestamp - update the timestamp of the fence + * @ring: ring the fence is associated with + * @seq: the fence seq number to update. + * @timestamp: the start timestamp to update. + * + * The function called at the time the fence and related ib is about to + * resubmit to gpu in MCBP scenario. Thus we do not consider race condition + * with amdgpu_fence_process to modify the same fence. + */ +void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq, ktime_t timestamp) +{ + struct amdgpu_fence_driver *drv = &ring->fence_drv; + struct dma_fence *fence; + + seq &= drv->num_fences_mask; + fence = drv->fences[seq]; + if (!fence) + return; + + to_amdgpu_fence(fence)->start_timestamp = timestamp; +} + /** * amdgpu_fence_driver_start_ring - make the fence driver * ready for use on the requested ring. diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c index 774c77bb8f4e..bcccc348dbe2 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_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index d3558c34d406..dc474b809604 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -569,3 +569,15 @@ 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 073e767c057f..f752c7ae7f60 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h @@ -145,8 +145,13 @@ signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring, uint32_t wait_seq, signed long timeout); unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring); + void amdgpu_fence_driver_isr_toggle(struct amdgpu_device *adev, bool stop); +u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring); +void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq, + ktime_t timestamp); + /* * Rings. */ @@ -313,6 +318,9 @@ 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 a52c4e7e8c39..3348337379da 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c @@ -28,6 +28,7 @@ #include "amdgpu.h" #define AMDGPU_MUX_RESUBMIT_JIFFIES_TIMEOUT (HZ / 2) +#define AMDGPU_MAX_LAST_UNSIGNALED_THRESHOLD_US 10000 static const struct ring_info { unsigned int hw_pio; @@ -37,23 +38,145 @@ static const struct ring_info { { AMDGPU_RING_PRIO_2, "gfx_high"}, }; +static struct kmem_cache *amdgpu_mux_chunk_slab; + +static inline struct amdgpu_mux_entry *amdgpu_ring_mux_sw_entry(struct amdgpu_ring_mux *mux, + struct amdgpu_ring *ring) +{ + return ring->entry_index < mux->ring_entry_size ? + &mux->ring_entry[ring->entry_index] : NULL; +} + +/* copy packages on sw ring range[begin, end) */ +static void amdgpu_ring_mux_copy_pkt_from_sw_ring(struct amdgpu_ring_mux *mux, + struct amdgpu_ring *ring, + u64 s_start, u64 s_end) +{ + u64 start, end; + struct amdgpu_ring *real_ring = mux->real_ring; + + start = s_start & ring->buf_mask; + end = s_end & ring->buf_mask; + + if (start == end) { + DRM_ERROR("no more data copied from sw ring\n"); + return; + } + if (start > end) { + amdgpu_ring_alloc(real_ring, (ring->ring_size >> 2) + end - start); + amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], + (ring->ring_size >> 2) - start); + amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[0], end); + } else { + amdgpu_ring_alloc(real_ring, end - start); + amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], end - start); + } +} + +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:*/ + if (!mux->s_resubmit) + return; + + for (i = 0; i < mux->num_ring_entries; i++) { + if (mux->ring_entry[i].ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT) { + e = &mux->ring_entry[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 = mux->seqno_to_resubmit; + if (last_seq < seq) { + /*resubmit all the fences between (last_seq, seq]*/ + list_for_each_entry(chunk, &e->list, entry) { + if (chunk->sync_seq > last_seq && chunk->sync_seq <= seq) { + amdgpu_fence_update_start_timestamp(e->ring, + chunk->sync_seq, + ktime_get()); + amdgpu_ring_mux_copy_pkt_from_sw_ring(mux, e->ring, + chunk->start, + chunk->end); + mux->wptr_resubmit = chunk->end; + amdgpu_ring_commit(mux->real_ring); + } + } + } + + del_timer(&mux->resubmit_timer); + mux->s_resubmit = false; +} + +static void amdgpu_ring_mux_schedule_resubmit(struct amdgpu_ring_mux *mux) +{ + mod_timer(&mux->resubmit_timer, jiffies + AMDGPU_MUX_RESUBMIT_JIFFIES_TIMEOUT); +} + +static void amdgpu_mux_resubmit_fallback(struct timer_list *t) +{ + struct amdgpu_ring_mux *mux = from_timer(mux, t, resubmit_timer); + + if (!spin_trylock(&mux->lock)) { + amdgpu_ring_mux_schedule_resubmit(mux); + DRM_ERROR("reschedule resubmit\n"); + return; + } + amdgpu_mux_resubmit_chunks(mux); + spin_unlock(&mux->lock); +} + int amdgpu_ring_mux_init(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, unsigned int entry_size) { mux->real_ring = ring; mux->num_ring_entries = 0; + mux->ring_entry = kcalloc(entry_size, sizeof(struct amdgpu_mux_entry), GFP_KERNEL); if (!mux->ring_entry) return -ENOMEM; mux->ring_entry_size = entry_size; + 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_entry[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); kfree(mux->ring_entry); mux->ring_entry = NULL; mux->num_ring_entries = 0; @@ -73,62 +196,48 @@ int amdgpu_ring_mux_add_sw_ring(struct amdgpu_ring_mux *mux, struct amdgpu_ring ring->entry_index = mux->num_ring_entries; e->ring = ring; + INIT_LIST_HEAD(&e->list); mux->num_ring_entries += 1; return 0; } -static inline struct amdgpu_mux_entry *amdgpu_ring_mux_sw_entry(struct amdgpu_ring_mux *mux, - struct amdgpu_ring *ring) -{ - return ring->entry_index < mux->ring_entry_size ? - &mux->ring_entry[ring->entry_index] : NULL; -} - -/* copy packages on sw ring range[begin, end) */ -static void amdgpu_ring_mux_copy_pkt_from_sw_ring(struct amdgpu_ring_mux *mux, - struct amdgpu_ring *ring, - u64 s_start, u64 s_end) -{ - u64 start, end; - struct amdgpu_ring *real_ring = mux->real_ring; - - start = s_start & ring->buf_mask; - end = s_end & ring->buf_mask; - - if (start == end) { - DRM_ERROR("no more data copied from sw ring\n"); - return; - } - if (start > end) { - amdgpu_ring_alloc(real_ring, (ring->ring_size >> 2) + end - start); - amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], - (ring->ring_size >> 2) - start); - amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[0], end); - } else { - amdgpu_ring_alloc(real_ring, end - start); - amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], end - start); - } -} - void amdgpu_ring_mux_set_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, u64 wptr) { struct amdgpu_mux_entry *e; + spin_lock(&mux->lock); + + if (ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT) + amdgpu_mux_resubmit_chunks(mux); + e = amdgpu_ring_mux_sw_entry(mux, ring); if (!e) { DRM_ERROR("cannot find entry for sw ring\n"); + spin_unlock(&mux->lock); + return; + } + + /* We could skip this set wptr as preemption in process. */ + if (ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT && mux->pending_trailing_fence_signaled) { + spin_unlock(&mux->lock); return; } - spin_lock(&mux->lock); e->sw_cptr = e->sw_wptr; + /* Update cptr if the package already copied in resubmit functions */ + if (ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT && e->sw_cptr < mux->wptr_resubmit) + e->sw_cptr = mux->wptr_resubmit; e->sw_wptr = wptr; e->start_ptr_in_hw_ring = mux->real_ring->wptr; - amdgpu_ring_mux_copy_pkt_from_sw_ring(mux, ring, e->sw_cptr, wptr); - e->end_ptr_in_hw_ring = mux->real_ring->wptr; - amdgpu_ring_commit(mux->real_ring); - + /* Skip copying for the packages already resubmitted.*/ + if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT || mux->wptr_resubmit < wptr) { + amdgpu_ring_mux_copy_pkt_from_sw_ring(mux, ring, e->sw_cptr, wptr); + e->end_ptr_in_hw_ring = mux->real_ring->wptr; + amdgpu_ring_commit(mux->real_ring); + } else { + e->end_ptr_in_hw_ring = mux->real_ring->wptr; + } spin_unlock(&mux->lock); } @@ -239,3 +348,169 @@ unsigned int amdgpu_sw_ring_priority(int idx) return idx < ARRAY_SIZE(sw_ring_info) ? sw_ring_info[idx].hw_pio : AMDGPU_RING_PRIO_DEFAULT; } + +/*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; + int i, need_preempt; + + need_preempt = 0; + for (i = 0; i < mux->num_ring_entries; i++) { + ring = mux->ring_entry[i].ring; + if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT && + amdgpu_fence_count_emitted(ring) > 0) + return 0; + if (ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT && + amdgpu_fence_last_unsignaled_time_us(ring) > + AMDGPU_MAX_LAST_UNSIGNALED_THRESHOLD_US) + need_preempt = 1; + } + return need_preempt && !mux->s_resubmit; +} + +/* Trigger Mid-Command Buffer Preemption (MCBP) and find if we need to resubmit. */ +int amdgpu_mcbp_trigger_preempt(struct amdgpu_ring_mux *mux) +{ + int r; + + spin_lock(&mux->lock); + mux->pending_trailing_fence_signaled = true; + r = amdgpu_ring_preempt_ib(mux->real_ring); + spin_unlock(&mux->lock); + return r; +} + +void amdgpu_sw_ring_ib_begin(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + struct amdgpu_ring_mux *mux = &adev->gfx.muxer; + + WARN_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; + + WARN_ON(!ring->is_sw_ring); + if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT) + return; + amdgpu_ring_mux_end_ib(mux, ring); +} + +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; + + spin_lock(&mux->lock); + amdgpu_mux_resubmit_chunks(mux); + spin_unlock(&mux->lock); + + e = amdgpu_ring_mux_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_ring_mux_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_ring_mux_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); +} + +bool amdgpu_mcbp_handle_trailing_fence_irq(struct amdgpu_ring_mux *mux) +{ + struct amdgpu_mux_entry *e; + struct amdgpu_ring *ring = NULL; + int i; + + if (!mux->pending_trailing_fence_signaled) + return false; + + if (mux->real_ring->trail_seq != le32_to_cpu(*mux->real_ring->trail_fence_cpu_addr)) + return false; + + for (i = 0; i < mux->num_ring_entries; i++) { + e = &mux->ring_entry[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 false; + } + + amdgpu_fence_process(ring); + if (amdgpu_fence_count_emitted(ring) > 0) { + mux->s_resubmit = true; + mux->seqno_to_resubmit = ring->fence_drv.sync_seq; + amdgpu_ring_mux_schedule_resubmit(mux); + } + + mux->pending_trailing_fence_signaled = false; + return true; +} diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h index f67970dc3dbd..4be45fc14954 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h @@ -29,6 +29,7 @@ #include "amdgpu_ring.h" struct amdgpu_ring; + /** * struct amdgpu_mux_entry - the entry recording software rings copying information. * @ring: the pointer to the software ring. @@ -37,6 +38,7 @@ struct amdgpu_ring; * @sw_cptr: the position of the copy pointer in the sw ring. * @sw_rptr: the read pointer in software ring. * @sw_wptr: the write pointer in software ring. + * @list: list head for amdgpu_mux_chunk */ struct amdgpu_mux_entry { struct amdgpu_ring *ring; @@ -45,6 +47,7 @@ struct amdgpu_mux_entry { u64 sw_cptr; u64 sw_rptr; u64 sw_wptr; + struct list_head list; }; struct amdgpu_ring_mux { @@ -55,6 +58,26 @@ struct amdgpu_ring_mux { unsigned int ring_entry_size; /*the lock for copy data from different software rings*/ spinlock_t lock; + bool s_resubmit; + uint32_t seqno_to_resubmit; + u64 wptr_resubmit; + struct timer_list resubmit_timer; + + bool pending_trailing_fence_signaled; +}; + +/** + * struct amdgpu_mux_chunk - save the location of indirect buffer's package on softare rings. + * @entry: the list entry. + * @sync_seq: the fence seqno related with the saved IB. + * @start:- start location on the software ring. + * @end:- end location on the software ring. + */ +struct amdgpu_mux_chunk { + struct list_head entry; + uint32_t sync_seq; + u64 start; + u64 end; }; int amdgpu_ring_mux_init(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, @@ -64,15 +87,17 @@ int amdgpu_ring_mux_add_sw_ring(struct amdgpu_ring_mux *mux, struct amdgpu_ring void amdgpu_ring_mux_set_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, u64 wptr); u64 amdgpu_ring_mux_get_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring); u64 amdgpu_ring_mux_get_rptr(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); +bool amdgpu_mcbp_handle_trailing_fence_irq(struct amdgpu_ring_mux *mux); u64 amdgpu_sw_ring_get_rptr_gfx(struct amdgpu_ring *ring); u64 amdgpu_sw_ring_get_wptr_gfx(struct amdgpu_ring *ring); void amdgpu_sw_ring_set_wptr_gfx(struct amdgpu_ring *ring); - void amdgpu_sw_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count); void amdgpu_sw_ring_ib_begin(struct amdgpu_ring *ring); void amdgpu_sw_ring_ib_end(struct amdgpu_ring *ring); - const char *amdgpu_sw_ring_name(int idx); unsigned int amdgpu_sw_ring_priority(int idx); + #endif diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 003aa9e47085..37b461a19117 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -541,6 +541,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync) return 0; + amdgpu_ring_ib_begin(ring); if (ring->funcs->init_cond_exec) patch_offset = amdgpu_ring_init_cond_exec(ring); @@ -601,6 +602,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, amdgpu_ring_emit_switch_buffer(ring); amdgpu_ring_emit_switch_buffer(ring); } + amdgpu_ring_ib_end(ring); return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c index afcc6e397c72..f202b45c413c 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c @@ -5456,7 +5456,7 @@ static int gfx_v9_0_ring_preempt_ib(struct amdgpu_ring *ring) ring->trail_seq += 1; amdgpu_ring_alloc(ring, 13); gfx_v9_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr, - ring->trail_seq, AMDGPU_FENCE_FLAG_EXEC); + ring->trail_seq, AMDGPU_FENCE_FLAG_EXEC | AMDGPU_FENCE_FLAG_INT); /*reset the CP_VMID_PREEMPT after trailing fence*/ amdgpu_ring_emit_wreg(ring, SOC15_REG_OFFSET(GC, 0, mmCP_VMID_PREEMPT), @@ -5882,8 +5882,9 @@ static int gfx_v9_0_eop_irq(struct amdgpu_device *adev, switch (me_id) { case 0: - /* Fence signals are handled on the software rings*/ - if (adev->gfx.num_gfx_rings) { + if (adev->gfx.num_gfx_rings && + !amdgpu_mcbp_handle_trailing_fence_irq(&adev->gfx.muxer)) { + /* Fence signals are handled on the software rings*/ for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++) amdgpu_fence_process(&adev->gfx.sw_gfx_ring[i]); } -- cgit From 89f3f24233c886c5662439fd829822adf3b88167 Mon Sep 17 00:00:00 2001 From: Jiadong Zhu Date: Mon, 5 Dec 2022 09:55:32 +0800 Subject: drm/amdgpu: Make amdgpu_ring_mux functions as static MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit lkp robot reported missing-prototypes and unused-but-set-variable warnings on some functions of amdgpu_mcbp_mux.c. Make them static and remove the unused variable. Reported-by: kernel test robot Signed-off-by: Jiadong Zhu Reviewed-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c | 8 +++----- 1 file changed, 3 insertions(+), 5 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c index 3348337379da..62079f0e3ee8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c @@ -350,7 +350,7 @@ unsigned int amdgpu_sw_ring_priority(int idx) } /*Scan on low prio rings to have unsignaled fence and high ring has no fence.*/ -int amdgpu_mcbp_scan(struct amdgpu_ring_mux *mux) +static int amdgpu_mcbp_scan(struct amdgpu_ring_mux *mux) { struct amdgpu_ring *ring; int i, need_preempt; @@ -370,7 +370,7 @@ int amdgpu_mcbp_scan(struct amdgpu_ring_mux *mux) } /* Trigger Mid-Command Buffer Preemption (MCBP) and find if we need to resubmit. */ -int amdgpu_mcbp_trigger_preempt(struct amdgpu_ring_mux *mux) +static int amdgpu_mcbp_trigger_preempt(struct amdgpu_ring_mux *mux) { int r; @@ -434,7 +434,7 @@ void amdgpu_ring_mux_start_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *r static void scan_and_remove_signaled_chunk(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring) { - uint32_t last_seq, size = 0; + uint32_t last_seq = 0; struct amdgpu_mux_entry *e; struct amdgpu_mux_chunk *chunk, *tmp; @@ -450,8 +450,6 @@ static void scan_and_remove_signaled_chunk(struct amdgpu_ring_mux *mux, struct a if (chunk->sync_seq <= last_seq) { list_del(&chunk->entry); kmem_cache_free(amdgpu_mux_chunk_slab, chunk); - } else { - size++; } } } -- cgit