diff options
Diffstat (limited to 'drivers/gpu/drm/radeon/radeon_ring.c')
| -rw-r--r-- | drivers/gpu/drm/radeon/radeon_ring.c | 410 |
1 files changed, 367 insertions, 43 deletions
diff --git a/drivers/gpu/drm/radeon/radeon_ring.c b/drivers/gpu/drm/radeon/radeon_ring.c index 983658c91358..ec79b3750430 100644 --- a/drivers/gpu/drm/radeon/radeon_ring.c +++ b/drivers/gpu/drm/radeon/radeon_ring.c @@ -35,47 +35,97 @@ #include "atom.h" /* - * IB. + * IB + * IBs (Indirect Buffers) and areas of GPU accessible memory where + * commands are stored. You can put a pointer to the IB in the + * command ring and the hw will fetch the commands from the IB + * and execute them. Generally userspace acceleration drivers + * produce command buffers which are send to the kernel and + * put in IBs for execution by the requested ring. */ int radeon_debugfs_sa_init(struct radeon_device *rdev); +/** + * radeon_ib_get - request an IB (Indirect Buffer) + * + * @rdev: radeon_device pointer + * @ring: ring index the IB is associated with + * @ib: IB object returned + * @size: requested IB size + * + * Request an IB (all asics). IBs are allocated using the + * suballocator. + * Returns 0 on success, error on failure. + */ int radeon_ib_get(struct radeon_device *rdev, int ring, struct radeon_ib *ib, unsigned size) { - int r; + int i, r; r = radeon_sa_bo_new(rdev, &rdev->ring_tmp_bo, &ib->sa_bo, size, 256, true); if (r) { dev_err(rdev->dev, "failed to get a new IB (%d)\n", r); return r; } - r = radeon_fence_create(rdev, &ib->fence, ring); + + r = radeon_semaphore_create(rdev, &ib->semaphore); if (r) { - dev_err(rdev->dev, "failed to create fence for new IB (%d)\n", r); - radeon_sa_bo_free(rdev, &ib->sa_bo, NULL); return r; } + ib->ring = ring; + ib->fence = NULL; ib->ptr = radeon_sa_bo_cpu_addr(ib->sa_bo); ib->gpu_addr = radeon_sa_bo_gpu_addr(ib->sa_bo); ib->vm_id = 0; ib->is_const_ib = false; - ib->semaphore = NULL; + for (i = 0; i < RADEON_NUM_RINGS; ++i) + ib->sync_to[i] = NULL; return 0; } +/** + * radeon_ib_free - free an IB (Indirect Buffer) + * + * @rdev: radeon_device pointer + * @ib: IB object to free + * + * Free an IB (all asics). + */ void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib *ib) { - radeon_semaphore_free(rdev, ib->semaphore, ib->fence); + radeon_semaphore_free(rdev, &ib->semaphore, ib->fence); radeon_sa_bo_free(rdev, &ib->sa_bo, ib->fence); radeon_fence_unref(&ib->fence); } -int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib) +/** + * radeon_ib_schedule - schedule an IB (Indirect Buffer) on the ring + * + * @rdev: radeon_device pointer + * @ib: IB object to schedule + * @const_ib: Const IB to schedule (SI only) + * + * Schedule an IB on the associated ring (all asics). + * Returns 0 on success, error on failure. + * + * On SI, there are two parallel engines fed from the primary ring, + * the CE (Constant Engine) and the DE (Drawing Engine). Since + * resource descriptors have moved to memory, the CE allows you to + * prime the caches while the DE is updating register state so that + * the resource descriptors will be already in cache when the draw is + * processed. To accomplish this, the userspace driver submits two + * IBs, one for the CE and one for the DE. If there is a CE IB (called + * a CONST_IB), it will be put on the ring prior to the DE IB. Prior + * to SI there was just a DE IB. + */ +int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib, + struct radeon_ib *const_ib) { - struct radeon_ring *ring = &rdev->ring[ib->fence->ring]; - int r = 0; + struct radeon_ring *ring = &rdev->ring[ib->ring]; + bool need_sync = false; + int i, r = 0; if (!ib->length_dw || !ring->ready) { /* TODO: Nothings in the ib we should report. */ @@ -84,17 +134,51 @@ int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib) } /* 64 dwords should be enough for fence too */ - r = radeon_ring_lock(rdev, ring, 64); + r = radeon_ring_lock(rdev, ring, 64 + RADEON_NUM_RINGS * 8); if (r) { dev_err(rdev->dev, "scheduling IB failed (%d).\n", r); return r; } - radeon_ring_ib_execute(rdev, ib->fence->ring, ib); - radeon_fence_emit(rdev, ib->fence); + for (i = 0; i < RADEON_NUM_RINGS; ++i) { + struct radeon_fence *fence = ib->sync_to[i]; + if (radeon_fence_need_sync(fence, ib->ring)) { + need_sync = true; + radeon_semaphore_sync_rings(rdev, ib->semaphore, + fence->ring, ib->ring); + radeon_fence_note_sync(fence, ib->ring); + } + } + /* immediately free semaphore when we don't need to sync */ + if (!need_sync) { + radeon_semaphore_free(rdev, &ib->semaphore, NULL); + } + if (const_ib) { + radeon_ring_ib_execute(rdev, const_ib->ring, const_ib); + radeon_semaphore_free(rdev, &const_ib->semaphore, NULL); + } + radeon_ring_ib_execute(rdev, ib->ring, ib); + r = radeon_fence_emit(rdev, &ib->fence, ib->ring); + if (r) { + dev_err(rdev->dev, "failed to emit fence for new IB (%d)\n", r); + radeon_ring_unlock_undo(rdev, ring); + return r; + } + if (const_ib) { + const_ib->fence = radeon_fence_ref(ib->fence); + } radeon_ring_unlock_commit(rdev, ring); return 0; } +/** + * radeon_ib_pool_init - Init the IB (Indirect Buffer) pool + * + * @rdev: radeon_device pointer + * + * Initialize the suballocator to manage a pool of memory + * for use as IBs (all asics). + * Returns 0 on success, error on failure. + */ int radeon_ib_pool_init(struct radeon_device *rdev) { int r; @@ -108,6 +192,12 @@ int radeon_ib_pool_init(struct radeon_device *rdev) if (r) { return r; } + + r = radeon_sa_bo_manager_start(rdev, &rdev->ring_tmp_bo); + if (r) { + return r; + } + rdev->ib_pool_ready = true; if (radeon_debugfs_sa_init(rdev)) { dev_err(rdev->dev, "failed to register debugfs file for SA\n"); @@ -115,24 +205,33 @@ int radeon_ib_pool_init(struct radeon_device *rdev) return 0; } +/** + * radeon_ib_pool_fini - Free the IB (Indirect Buffer) pool + * + * @rdev: radeon_device pointer + * + * Tear down the suballocator managing the pool of memory + * for use as IBs (all asics). + */ void radeon_ib_pool_fini(struct radeon_device *rdev) { if (rdev->ib_pool_ready) { + radeon_sa_bo_manager_suspend(rdev, &rdev->ring_tmp_bo); radeon_sa_bo_manager_fini(rdev, &rdev->ring_tmp_bo); rdev->ib_pool_ready = false; } } -int radeon_ib_pool_start(struct radeon_device *rdev) -{ - return radeon_sa_bo_manager_start(rdev, &rdev->ring_tmp_bo); -} - -int radeon_ib_pool_suspend(struct radeon_device *rdev) -{ - return radeon_sa_bo_manager_suspend(rdev, &rdev->ring_tmp_bo); -} - +/** + * radeon_ib_ring_tests - test IBs on the rings + * + * @rdev: radeon_device pointer + * + * Test an IB (Indirect Buffer) on each ring. + * If the test fails, disable the ring. + * Returns 0 on success, error if the primary GFX ring + * IB test fails. + */ int radeon_ib_ring_tests(struct radeon_device *rdev) { unsigned i; @@ -164,10 +263,28 @@ int radeon_ib_ring_tests(struct radeon_device *rdev) } /* - * Ring. + * Rings + * Most engines on the GPU are fed via ring buffers. Ring + * buffers are areas of GPU accessible memory that the host + * writes commands into and the GPU reads commands out of. + * There is a rptr (read pointer) that determines where the + * GPU is currently reading, and a wptr (write pointer) + * which determines where the host has written. When the + * pointers are equal, the ring is idle. When the host + * writes commands to the ring buffer, it increments the + * wptr. The GPU then starts fetching commands and executes + * them until the pointers are equal again. */ int radeon_debugfs_ring_init(struct radeon_device *rdev, struct radeon_ring *ring); +/** + * radeon_ring_write - write a value to the ring + * + * @ring: radeon_ring structure holding ring information + * @v: dword (dw) value to write + * + * Write a value to the requested ring buffer (all asics). + */ void radeon_ring_write(struct radeon_ring *ring, uint32_t v) { #if DRM_DEBUG_CODE @@ -181,21 +298,37 @@ void radeon_ring_write(struct radeon_ring *ring, uint32_t v) ring->ring_free_dw--; } -int radeon_ring_index(struct radeon_device *rdev, struct radeon_ring *ring) +/** + * radeon_ring_supports_scratch_reg - check if the ring supports + * writing to scratch registers + * + * @rdev: radeon_device pointer + * @ring: radeon_ring structure holding ring information + * + * Check if a specific ring supports writing to scratch registers (all asics). + * Returns true if the ring supports writing to scratch regs, false if not. + */ +bool radeon_ring_supports_scratch_reg(struct radeon_device *rdev, + struct radeon_ring *ring) { - /* r1xx-r5xx only has CP ring */ - if (rdev->family < CHIP_R600) - return RADEON_RING_TYPE_GFX_INDEX; - - if (rdev->family >= CHIP_CAYMAN) { - if (ring == &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX]) - return CAYMAN_RING_TYPE_CP1_INDEX; - else if (ring == &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX]) - return CAYMAN_RING_TYPE_CP2_INDEX; + switch (ring->idx) { + case RADEON_RING_TYPE_GFX_INDEX: + case CAYMAN_RING_TYPE_CP1_INDEX: + case CAYMAN_RING_TYPE_CP2_INDEX: + return true; + default: + return false; } - return RADEON_RING_TYPE_GFX_INDEX; } +/** + * radeon_ring_free_size - update the free size + * + * @rdev: radeon_device pointer + * @ring: radeon_ring structure holding ring information + * + * Update the free dw slots in the ring buffer (all asics). + */ void radeon_ring_free_size(struct radeon_device *rdev, struct radeon_ring *ring) { u32 rptr; @@ -214,7 +347,16 @@ void radeon_ring_free_size(struct radeon_device *rdev, struct radeon_ring *ring) } } - +/** + * radeon_ring_alloc - allocate space on the ring buffer + * + * @rdev: radeon_device pointer + * @ring: radeon_ring structure holding ring information + * @ndw: number of dwords to allocate in the ring buffer + * + * Allocate @ndw dwords in the ring buffer (all asics). + * Returns 0 on success, error on failure. + */ int radeon_ring_alloc(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw) { int r; @@ -227,7 +369,7 @@ int radeon_ring_alloc(struct radeon_device *rdev, struct radeon_ring *ring, unsi if (ndw < ring->ring_free_dw) { break; } - r = radeon_fence_wait_next_locked(rdev, radeon_ring_index(rdev, ring)); + r = radeon_fence_wait_next_locked(rdev, ring->idx); if (r) return r; } @@ -236,6 +378,17 @@ int radeon_ring_alloc(struct radeon_device *rdev, struct radeon_ring *ring, unsi return 0; } +/** + * radeon_ring_lock - lock the ring and allocate space on it + * + * @rdev: radeon_device pointer + * @ring: radeon_ring structure holding ring information + * @ndw: number of dwords to allocate in the ring buffer + * + * Lock the ring and allocate @ndw dwords in the ring buffer + * (all asics). + * Returns 0 on success, error on failure. + */ int radeon_ring_lock(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw) { int r; @@ -249,15 +402,20 @@ int radeon_ring_lock(struct radeon_device *rdev, struct radeon_ring *ring, unsig return 0; } +/** + * radeon_ring_commit - tell the GPU to execute the new + * commands on the ring buffer + * + * @rdev: radeon_device pointer + * @ring: radeon_ring structure holding ring information + * + * Update the wptr (write pointer) to tell the GPU to + * execute new commands on the ring buffer (all asics). + */ void radeon_ring_commit(struct radeon_device *rdev, struct radeon_ring *ring) { - unsigned count_dw_pad; - unsigned i; - /* We pad to match fetch size */ - count_dw_pad = (ring->align_mask + 1) - - (ring->wptr & ring->align_mask); - for (i = 0; i < count_dw_pad; i++) { + while (ring->wptr & ring->align_mask) { radeon_ring_write(ring, ring->nop); } DRM_MEMORYBARRIER(); @@ -265,23 +423,55 @@ void radeon_ring_commit(struct radeon_device *rdev, struct radeon_ring *ring) (void)RREG32(ring->wptr_reg); } +/** + * radeon_ring_unlock_commit - tell the GPU to execute the new + * commands on the ring buffer and unlock it + * + * @rdev: radeon_device pointer + * @ring: radeon_ring structure holding ring information + * + * Call radeon_ring_commit() then unlock the ring (all asics). + */ void radeon_ring_unlock_commit(struct radeon_device *rdev, struct radeon_ring *ring) { radeon_ring_commit(rdev, ring); mutex_unlock(&rdev->ring_lock); } +/** + * radeon_ring_undo - reset the wptr + * + * @ring: radeon_ring structure holding ring information + * + * Reset the driver's copy of the wtpr (all asics). + */ void radeon_ring_undo(struct radeon_ring *ring) { ring->wptr = ring->wptr_old; } +/** + * radeon_ring_unlock_undo - reset the wptr and unlock the ring + * + * @ring: radeon_ring structure holding ring information + * + * Call radeon_ring_undo() then unlock the ring (all asics). + */ void radeon_ring_unlock_undo(struct radeon_device *rdev, struct radeon_ring *ring) { radeon_ring_undo(ring); mutex_unlock(&rdev->ring_lock); } +/** + * radeon_ring_force_activity - add some nop packets to the ring + * + * @rdev: radeon_device pointer + * @ring: radeon_ring structure holding ring information + * + * Add some nop packets to the ring to force activity (all asics). + * Used for lockup detection to see if the rptr is advancing. + */ void radeon_ring_force_activity(struct radeon_device *rdev, struct radeon_ring *ring) { int r; @@ -296,6 +486,13 @@ void radeon_ring_force_activity(struct radeon_device *rdev, struct radeon_ring * } } +/** + * radeon_ring_force_activity - update lockup variables + * + * @ring: radeon_ring structure holding ring information + * + * Update the last rptr value and timestamp (all asics). + */ void radeon_ring_lockup_update(struct radeon_ring *ring) { ring->last_rptr = ring->rptr; @@ -349,6 +546,116 @@ bool radeon_ring_test_lockup(struct radeon_device *rdev, struct radeon_ring *rin return false; } +/** + * radeon_ring_backup - Back up the content of a ring + * + * @rdev: radeon_device pointer + * @ring: the ring we want to back up + * + * Saves all unprocessed commits from a ring, returns the number of dwords saved. + */ +unsigned radeon_ring_backup(struct radeon_device *rdev, struct radeon_ring *ring, + uint32_t **data) +{ + unsigned size, ptr, i; + + /* just in case lock the ring */ + mutex_lock(&rdev->ring_lock); + *data = NULL; + + if (ring->ring_obj == NULL) { + mutex_unlock(&rdev->ring_lock); + return 0; + } + + /* it doesn't make sense to save anything if all fences are signaled */ + if (!radeon_fence_count_emitted(rdev, ring->idx)) { + mutex_unlock(&rdev->ring_lock); + return 0; + } + + /* calculate the number of dw on the ring */ + if (ring->rptr_save_reg) + ptr = RREG32(ring->rptr_save_reg); + else if (rdev->wb.enabled) + ptr = le32_to_cpu(*ring->next_rptr_cpu_addr); + else { + /* no way to read back the next rptr */ + mutex_unlock(&rdev->ring_lock); + return 0; + } + + size = ring->wptr + (ring->ring_size / 4); + size -= ptr; + size &= ring->ptr_mask; + if (size == 0) { + mutex_unlock(&rdev->ring_lock); + return 0; + } + + /* and then save the content of the ring */ + *data = kmalloc_array(size, sizeof(uint32_t), GFP_KERNEL); + if (!*data) { + mutex_unlock(&rdev->ring_lock); + return 0; + } + for (i = 0; i < size; ++i) { + (*data)[i] = ring->ring[ptr++]; + ptr &= ring->ptr_mask; + } + + mutex_unlock(&rdev->ring_lock); + return size; +} + +/** + * radeon_ring_restore - append saved commands to the ring again + * + * @rdev: radeon_device pointer + * @ring: ring to append commands to + * @size: number of dwords we want to write + * @data: saved commands + * + * Allocates space on the ring and restore the previously saved commands. + */ +int radeon_ring_restore(struct radeon_device *rdev, struct radeon_ring *ring, + unsigned size, uint32_t *data) +{ + int i, r; + + if (!size || !data) + return 0; + + /* restore the saved ring content */ + r = radeon_ring_lock(rdev, ring, size); + if (r) + return r; + + for (i = 0; i < size; ++i) { + radeon_ring_write(ring, data[i]); + } + + radeon_ring_unlock_commit(rdev, ring); + kfree(data); + return 0; +} + +/** + * radeon_ring_init - init driver ring struct. + * + * @rdev: radeon_device pointer + * @ring: radeon_ring structure holding ring information + * @ring_size: size of the ring + * @rptr_offs: offset of the rptr writeback location in the WB buffer + * @rptr_reg: MMIO offset of the rptr register + * @wptr_reg: MMIO offset of the wptr register + * @ptr_reg_shift: bit offset of the rptr/wptr values + * @ptr_reg_mask: bit mask of the rptr/wptr values + * @nop: nop packet for this ring + * + * Initialize the driver information for the selected ring (all asics). + * Returns 0 on success, error on failure. + */ int radeon_ring_init(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ring_size, unsigned rptr_offs, unsigned rptr_reg, unsigned wptr_reg, u32 ptr_reg_shift, u32 ptr_reg_mask, u32 nop) @@ -391,12 +698,25 @@ int radeon_ring_init(struct radeon_device *rdev, struct radeon_ring *ring, unsig } ring->ptr_mask = (ring->ring_size / 4) - 1; ring->ring_free_dw = ring->ring_size / 4; + if (rdev->wb.enabled) { + u32 index = RADEON_WB_RING0_NEXT_RPTR + (ring->idx * 4); + ring->next_rptr_gpu_addr = rdev->wb.gpu_addr + index; + ring->next_rptr_cpu_addr = &rdev->wb.wb[index/4]; + } if (radeon_debugfs_ring_init(rdev, ring)) { DRM_ERROR("Failed to register debugfs file for rings !\n"); } return 0; } +/** + * radeon_ring_fini - tear down the driver ring struct. + * + * @rdev: radeon_device pointer + * @ring: radeon_ring structure holding ring information + * + * Tear down the driver information for the selected ring (all asics). + */ void radeon_ring_fini(struct radeon_device *rdev, struct radeon_ring *ring) { int r; @@ -438,6 +758,10 @@ static int radeon_debugfs_ring_info(struct seq_file *m, void *data) count = (ring->ring_size / 4) - ring->ring_free_dw; seq_printf(m, "wptr(0x%04x): 0x%08x\n", ring->wptr_reg, RREG32(ring->wptr_reg)); seq_printf(m, "rptr(0x%04x): 0x%08x\n", ring->rptr_reg, RREG32(ring->rptr_reg)); + if (ring->rptr_save_reg) { + seq_printf(m, "rptr next(0x%04x): 0x%08x\n", ring->rptr_save_reg, + RREG32(ring->rptr_save_reg)); + } seq_printf(m, "driver's copy of the wptr: 0x%08x\n", ring->wptr); seq_printf(m, "driver's copy of the rptr: 0x%08x\n", ring->rptr); seq_printf(m, "%u free dwords in ring\n", ring->ring_free_dw); |