From a9f7b97ddae36d664d627b152d5c5a07b32ba816 Mon Sep 17 00:00:00 2001 From: John Harrison Date: Tue, 19 Nov 2024 16:02:22 -0800 Subject: [PATCH 01/16] drm/xe/guc: Add support for G2G communications Some features require inter-GuC communication channels on multi-tile devices. So allocate and enable such. v2: Correct use of xe_bo_get/put (review feedback from Matthew Brost) Add extra assert, re-order a calculation for better clarity and add comments to slot calculation (review feedback from Daniele). Also slightly re-work the slot calc to avoid code duplication. Signed-off-by: John Harrison Reviewed-by: Daniele Ceraolo Spurio Link: https://patchwork.freedesktop.org/patch/msgid/20241120000222.204095-3-John.C.Harrison@Intel.com --- drivers/gpu/drm/xe/abi/guc_actions_abi.h | 20 ++ drivers/gpu/drm/xe/xe_guc.c | 308 ++++++++++++++++++++++- drivers/gpu/drm/xe/xe_guc_types.h | 10 + 3 files changed, 337 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/xe/abi/guc_actions_abi.h b/drivers/gpu/drm/xe/abi/guc_actions_abi.h index b54fe40fc5a9..fee385532fb0 100644 --- a/drivers/gpu/drm/xe/abi/guc_actions_abi.h +++ b/drivers/gpu/drm/xe/abi/guc_actions_abi.h @@ -134,6 +134,8 @@ enum xe_guc_action { XE_GUC_ACTION_DEREGISTER_CONTEXT = 0x4503, XE_GUC_ACTION_REGISTER_COMMAND_TRANSPORT_BUFFER = 0x4505, XE_GUC_ACTION_DEREGISTER_COMMAND_TRANSPORT_BUFFER = 0x4506, + XE_GUC_ACTION_REGISTER_G2G = 0x4507, + XE_GUC_ACTION_DEREGISTER_G2G = 0x4508, XE_GUC_ACTION_DEREGISTER_CONTEXT_DONE = 0x4600, XE_GUC_ACTION_REGISTER_CONTEXT_MULTI_LRC = 0x4601, XE_GUC_ACTION_CLIENT_SOFT_RESET = 0x5507, @@ -218,4 +220,22 @@ enum xe_guc_tlb_inval_mode { XE_GUC_TLB_INVAL_MODE_LITE = 0x1, }; +/* + * GuC to GuC communication (de-)registration fields: + */ +enum xe_guc_g2g_type { + XE_G2G_TYPE_IN = 0x0, + XE_G2G_TYPE_OUT, + XE_G2G_TYPE_LIMIT, +}; + +#define XE_G2G_REGISTER_DEVICE REG_GENMASK(16, 16) +#define XE_G2G_REGISTER_TILE REG_GENMASK(15, 12) +#define XE_G2G_REGISTER_TYPE REG_GENMASK(11, 8) +#define XE_G2G_REGISTER_SIZE REG_GENMASK(7, 0) + +#define XE_G2G_DEREGISTER_DEVICE REG_GENMASK(16, 16) +#define XE_G2G_DEREGISTER_TILE REG_GENMASK(15, 12) +#define XE_G2G_DEREGISTER_TYPE REG_GENMASK(11, 8) + #endif diff --git a/drivers/gpu/drm/xe/xe_guc.c b/drivers/gpu/drm/xe/xe_guc.c index df1ba94cf4ca..4e2868efb620 100644 --- a/drivers/gpu/drm/xe/xe_guc.c +++ b/drivers/gpu/drm/xe/xe_guc.c @@ -44,7 +44,15 @@ static u32 guc_bo_ggtt_addr(struct xe_guc *guc, struct xe_bo *bo) { struct xe_device *xe = guc_to_xe(guc); - u32 addr = xe_bo_ggtt_addr(bo); + u32 addr; + + /* + * For most BOs, the address on the allocating tile is fine. However for + * some, e.g. G2G CTB, the address on a specific tile is required as it + * might be different for each tile. So, just always ask for the address + * on the target GuC. + */ + addr = __xe_bo_ggtt_addr(bo, gt_to_tile(guc_to_gt(guc))->id); /* GuC addresses above GUC_GGTT_TOP don't map through the GTT */ xe_assert(xe, addr >= xe_wopcm_size(guc_to_xe(guc))); @@ -244,6 +252,293 @@ static void guc_write_params(struct xe_guc *guc) xe_mmio_write32(>->mmio, SOFT_SCRATCH(1 + i), guc->params[i]); } +static int guc_action_register_g2g_buffer(struct xe_guc *guc, u32 type, u32 dst_tile, u32 dst_dev, + u32 desc_addr, u32 buff_addr, u32 size) +{ + struct xe_gt *gt = guc_to_gt(guc); + struct xe_device *xe = gt_to_xe(gt); + u32 action[] = { + XE_GUC_ACTION_REGISTER_G2G, + FIELD_PREP(XE_G2G_REGISTER_SIZE, size / SZ_4K - 1) | + FIELD_PREP(XE_G2G_REGISTER_TYPE, type) | + FIELD_PREP(XE_G2G_REGISTER_TILE, dst_tile) | + FIELD_PREP(XE_G2G_REGISTER_DEVICE, dst_dev), + desc_addr, + buff_addr, + }; + + xe_assert(xe, (type == XE_G2G_TYPE_IN) || (type == XE_G2G_TYPE_OUT)); + xe_assert(xe, !(size % SZ_4K)); + + return xe_guc_ct_send_block(&guc->ct, action, ARRAY_SIZE(action)); +} + +static int guc_action_deregister_g2g_buffer(struct xe_guc *guc, u32 type, u32 dst_tile, u32 dst_dev) +{ + struct xe_gt *gt = guc_to_gt(guc); + struct xe_device *xe = gt_to_xe(gt); + u32 action[] = { + XE_GUC_ACTION_DEREGISTER_G2G, + FIELD_PREP(XE_G2G_DEREGISTER_TYPE, type) | + FIELD_PREP(XE_G2G_DEREGISTER_TILE, dst_tile) | + FIELD_PREP(XE_G2G_DEREGISTER_DEVICE, dst_dev), + }; + + xe_assert(xe, (type == XE_G2G_TYPE_IN) || (type == XE_G2G_TYPE_OUT)); + + return xe_guc_ct_send_block(&guc->ct, action, ARRAY_SIZE(action)); +} + +#define G2G_DEV(gt) (((gt)->info.type == XE_GT_TYPE_MAIN) ? 0 : 1) + +#define G2G_BUFFER_SIZE (SZ_4K) +#define G2G_DESC_SIZE (64) +#define G2G_DESC_AREA_SIZE (SZ_4K) + +/* + * Generate a unique id for each bi-directional CTB for each pair of + * near and far tiles/devices. The id can then be used as an index into + * a single allocation that is sub-divided into multiple CTBs. + * + * For example, with two devices per tile and two tiles, the table should + * look like: + * Far . + * 0.0 0.1 1.0 1.1 + * N 0.0 --/-- 00/01 02/03 04/05 + * e 0.1 01/00 --/-- 06/07 08/09 + * a 1.0 03/02 07/06 --/-- 10/11 + * r 1.1 05/04 09/08 11/10 --/-- + * + * Where each entry is Rx/Tx channel id. + * + * So GuC #3 (tile 1, dev 1) talking to GuC #2 (tile 1, dev 0) would + * be reading from channel #11 and writing to channel #10. Whereas, + * GuC #2 talking to GuC #3 would be read on #10 and write to #11. + */ +static unsigned int g2g_slot(u32 near_tile, u32 near_dev, u32 far_tile, u32 far_dev, + u32 type, u32 max_inst, bool have_dev) +{ + u32 near = near_tile, far = far_tile; + u32 idx = 0, x, y, direction; + int i; + + if (have_dev) { + near = (near << 1) | near_dev; + far = (far << 1) | far_dev; + } + + /* No need to send to one's self */ + if (far == near) + return -1; + + if (far > near) { + /* Top right table half */ + x = far; + y = near; + + /* T/R is 'forwards' direction */ + direction = type; + } else { + /* Bottom left table half */ + x = near; + y = far; + + /* B/L is 'backwards' direction */ + direction = (1 - type); + } + + /* Count the rows prior to the target */ + for (i = y; i > 0; i--) + idx += max_inst - i; + + /* Count this row up to the target */ + idx += (x - 1 - y); + + /* Slots are in Rx/Tx pairs */ + idx *= 2; + + /* Pick Rx/Tx direction */ + idx += direction; + + return idx; +} + +static int guc_g2g_register(struct xe_guc *near_guc, struct xe_gt *far_gt, u32 type, bool have_dev) +{ + struct xe_gt *near_gt = guc_to_gt(near_guc); + struct xe_device *xe = gt_to_xe(near_gt); + struct xe_bo *g2g_bo; + u32 near_tile = gt_to_tile(near_gt)->id; + u32 near_dev = G2G_DEV(near_gt); + u32 far_tile = gt_to_tile(far_gt)->id; + u32 far_dev = G2G_DEV(far_gt); + u32 max = xe->info.gt_count; + u32 base, desc, buf; + int slot; + + /* G2G is not allowed between different cards */ + xe_assert(xe, xe == gt_to_xe(far_gt)); + + g2g_bo = near_guc->g2g.bo; + xe_assert(xe, g2g_bo); + + slot = g2g_slot(near_tile, near_dev, far_tile, far_dev, type, max, have_dev); + xe_assert(xe, slot >= 0); + + base = guc_bo_ggtt_addr(near_guc, g2g_bo); + desc = base + slot * G2G_DESC_SIZE; + buf = base + G2G_DESC_AREA_SIZE + slot * G2G_BUFFER_SIZE; + + xe_assert(xe, (desc - base + G2G_DESC_SIZE) <= G2G_DESC_AREA_SIZE); + xe_assert(xe, (buf - base + G2G_BUFFER_SIZE) <= g2g_bo->size); + + return guc_action_register_g2g_buffer(near_guc, type, far_tile, far_dev, + desc, buf, G2G_BUFFER_SIZE); +} + +static void guc_g2g_deregister(struct xe_guc *guc, u32 far_tile, u32 far_dev, u32 type) +{ + guc_action_deregister_g2g_buffer(guc, type, far_tile, far_dev); +} + +static u32 guc_g2g_size(struct xe_guc *guc) +{ + struct xe_gt *gt = guc_to_gt(guc); + struct xe_device *xe = gt_to_xe(gt); + unsigned int count = xe->info.gt_count; + u32 num_channels = (count * (count - 1)) / 2; + + xe_assert(xe, num_channels * XE_G2G_TYPE_LIMIT * G2G_DESC_SIZE <= G2G_DESC_AREA_SIZE); + + return num_channels * XE_G2G_TYPE_LIMIT * G2G_BUFFER_SIZE + G2G_DESC_AREA_SIZE; +} + +static bool xe_guc_g2g_wanted(struct xe_device *xe) +{ + /* Can't do GuC to GuC communication if there is only one GuC */ + if (xe->info.gt_count <= 1) + return false; + + /* No current user */ + return false; +} + +static int guc_g2g_alloc(struct xe_guc *guc) +{ + struct xe_gt *gt = guc_to_gt(guc); + struct xe_device *xe = gt_to_xe(gt); + struct xe_tile *tile = gt_to_tile(gt); + struct xe_bo *bo; + u32 g2g_size; + + if (guc->g2g.bo) + return 0; + + if (gt->info.id != 0) { + struct xe_gt *root_gt = xe_device_get_gt(xe, 0); + struct xe_guc *root_guc = &root_gt->uc.guc; + struct xe_bo *bo; + + bo = xe_bo_get(root_guc->g2g.bo); + if (!bo) + return -ENODEV; + + guc->g2g.bo = bo; + guc->g2g.owned = false; + return 0; + } + + g2g_size = guc_g2g_size(guc); + bo = xe_managed_bo_create_pin_map(xe, tile, g2g_size, + XE_BO_FLAG_VRAM_IF_DGFX(tile) | + XE_BO_FLAG_GGTT | + XE_BO_FLAG_GGTT_ALL | + XE_BO_FLAG_GGTT_INVALIDATE); + if (IS_ERR(bo)) + return PTR_ERR(bo); + + xe_map_memset(xe, &bo->vmap, 0, 0, g2g_size); + guc->g2g.bo = bo; + guc->g2g.owned = true; + + return 0; +} + +static void guc_g2g_fini(struct xe_guc *guc) +{ + if (!guc->g2g.bo) + return; + + /* Unpinning the owned object is handled by generic shutdown */ + if (!guc->g2g.owned) + xe_bo_put(guc->g2g.bo); + + guc->g2g.bo = NULL; +} + +static int guc_g2g_start(struct xe_guc *guc) +{ + struct xe_gt *far_gt, *gt = guc_to_gt(guc); + struct xe_device *xe = gt_to_xe(gt); + unsigned int i, j; + int t, err; + bool have_dev; + + if (!guc->g2g.bo) { + int ret; + + ret = guc_g2g_alloc(guc); + if (ret) + return ret; + } + + /* GuC interface will need extending if more GT device types are ever created. */ + xe_gt_assert(gt, (gt->info.type == XE_GT_TYPE_MAIN) || (gt->info.type == XE_GT_TYPE_MEDIA)); + + /* Channel numbering depends on whether there are multiple GTs per tile */ + have_dev = xe->info.gt_count > xe->info.tile_count; + + for_each_gt(far_gt, xe, i) { + u32 far_tile, far_dev; + + if (far_gt->info.id == gt->info.id) + continue; + + far_tile = gt_to_tile(far_gt)->id; + far_dev = G2G_DEV(far_gt); + + for (t = 0; t < XE_G2G_TYPE_LIMIT; t++) { + err = guc_g2g_register(guc, far_gt, t, have_dev); + if (err) { + while (--t >= 0) + guc_g2g_deregister(guc, far_tile, far_dev, t); + goto err_deregister; + } + } + } + + return 0; + +err_deregister: + for_each_gt(far_gt, xe, j) { + u32 tile, dev; + + if (far_gt->info.id == gt->info.id) + continue; + + if (j >= i) + break; + + tile = gt_to_tile(far_gt)->id; + dev = G2G_DEV(far_gt); + + for (t = 0; t < XE_G2G_TYPE_LIMIT; t++) + guc_g2g_deregister(guc, tile, dev, t); + } + + return err; +} + static void guc_fini_hw(void *arg) { struct xe_guc *guc = arg; @@ -253,6 +548,8 @@ static void guc_fini_hw(void *arg) fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FORCEWAKE_ALL); xe_uc_fini_hw(&guc_to_gt(guc)->uc); xe_force_wake_put(gt_to_fw(gt), fw_ref); + + guc_g2g_fini(guc); } /** @@ -423,7 +720,16 @@ int xe_guc_init_post_hwconfig(struct xe_guc *guc) int xe_guc_post_load_init(struct xe_guc *guc) { + int ret; + xe_guc_ads_populate_post_load(&guc->ads); + + if (xe_guc_g2g_wanted(guc_to_xe(guc))) { + ret = guc_g2g_start(guc); + if (ret) + return ret; + } + guc->submission_state.enabled = true; return 0; diff --git a/drivers/gpu/drm/xe/xe_guc_types.h b/drivers/gpu/drm/xe/xe_guc_types.h index fa75f57bf5da..83a41ebcdc91 100644 --- a/drivers/gpu/drm/xe/xe_guc_types.h +++ b/drivers/gpu/drm/xe/xe_guc_types.h @@ -64,6 +64,15 @@ struct xe_guc { struct xe_guc_pc pc; /** @dbm: GuC Doorbell Manager */ struct xe_guc_db_mgr dbm; + + /** @g2g: GuC to GuC communication state */ + struct { + /** @g2g.bo: Storage for GuC to GuC communication channels */ + struct xe_bo *bo; + /** @g2g.owned: Is the BO owned by this GT or just mapped in */ + bool owned; + } g2g; + /** @submission_state: GuC submission state */ struct { /** @submission_state.idm: GuC context ID Manager */ @@ -79,6 +88,7 @@ struct xe_guc { /** @submission_state.fini_wq: submit fini wait queue */ wait_queue_head_t fini_wq; } submission_state; + /** @hwconfig: Hardware config state */ struct { /** @hwconfig.bo: buffer object of the hardware config */ -- 2.51.0 From 4fe70f664a105391321c85b2af241001e8118d24 Mon Sep 17 00:00:00 2001 From: Matt Roper Date: Mon, 25 Nov 2024 11:48:39 -0800 Subject: [PATCH 02/16] drm/xe: Update xe2_graphics name string Since both Xe2 and Xe3 platforms currently use the same set of graphics IP feature flags, we associate the "graphics_xe2" structure with both IPs. Update the name string on that IP structure to clarify this and avoid confusion as Xe3 platforms start going into public CI. Fixes: 800d75bf20ae ("drm/xe/xe3: Define Xe3 feature flags") Signed-off-by: Matt Roper Reviewed-by: Vinay Belgaumkar Link: https://patchwork.freedesktop.org/patch/msgid/20241125194838.1190599-2-matthew.d.roper@intel.com --- drivers/gpu/drm/xe/xe_pci.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/drivers/gpu/drm/xe/xe_pci.c b/drivers/gpu/drm/xe/xe_pci.c index 64a8336ca437..2cfd1a4df92e 100644 --- a/drivers/gpu/drm/xe/xe_pci.c +++ b/drivers/gpu/drm/xe/xe_pci.c @@ -174,7 +174,7 @@ static const struct xe_graphics_desc graphics_xelpg = { GENMASK(XE_HW_ENGINE_CCS3, XE_HW_ENGINE_CCS0) static const struct xe_graphics_desc graphics_xe2 = { - .name = "Xe2_LPG / Xe2_HPG", + .name = "Xe2_LPG / Xe2_HPG / Xe3_LPG", XE2_GFX_FEATURES, }; -- 2.51.0 From f7e1fe4593511ff3abe70387235a57aff546adb4 Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Tue, 19 Nov 2024 10:19:27 +0000 Subject: [PATCH 03/16] drm/xe/vram: fix lpfn check MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Technically we should check the lfpn value and not the place->lpfn, for the case where the allocation itself could be as large as the entire region and not be range based, which might result in incorrectly doing a power-of-two roundup. The allocator itself will already ensure it's contiguous underneath for such an allocation. This shouldn't fix any current usecase, but never the less came up from some internal testing. Signed-off-by: Matthew Auld Cc: Piotr Piórkowski Cc: Matthew Brost Reviewed-by: Matthew Brost Link: https://patchwork.freedesktop.org/patch/msgid/20241119101926.190203-2-matthew.auld@intel.com --- drivers/gpu/drm/xe/xe_ttm_vram_mgr.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c b/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c index 1d39a8c53b3a..c95728c45ea4 100644 --- a/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c +++ b/drivers/gpu/drm/xe/xe_ttm_vram_mgr.c @@ -109,7 +109,7 @@ static int xe_ttm_vram_mgr_new(struct ttm_resource_manager *man, goto error_unlock; } - if (place->fpfn + (size >> PAGE_SHIFT) != place->lpfn && + if (place->fpfn + (size >> PAGE_SHIFT) != lpfn && place->flags & TTM_PL_FLAG_CONTIGUOUS) { size = roundup_pow_of_two(size); min_page_size = size; -- 2.51.0 From 6364a06c5e943b6761d1ffe5c177cf9cc9be867b Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Fri, 22 Nov 2024 16:19:15 +0000 Subject: [PATCH 04/16] drm/xe/trace: improve xe_sched_msg trace Also include the gt_id, that way we can ignore duplicate guc_id across different GTs when applying some filtering. Signed-off-by: Matthew Auld Cc: Matthew Brost Reviewed-by: Lucas De Marchi Link: https://patchwork.freedesktop.org/patch/msgid/20241122161914.321263-4-matthew.auld@intel.com --- drivers/gpu/drm/xe/xe_trace.h | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/xe/xe_trace.h b/drivers/gpu/drm/xe/xe_trace.h index de682978c4bf..d5281de04d54 100644 --- a/drivers/gpu/drm/xe/xe_trace.h +++ b/drivers/gpu/drm/xe/xe_trace.h @@ -284,6 +284,7 @@ DECLARE_EVENT_CLASS(xe_sched_msg, __string(dev, __dev_name_eq(((struct xe_exec_queue *)msg->private_data))) __field(u32, opcode) __field(u16, guc_id) + __field(u8, gt_id) ), TP_fast_assign( @@ -291,9 +292,11 @@ DECLARE_EVENT_CLASS(xe_sched_msg, __entry->opcode = msg->opcode; __entry->guc_id = ((struct xe_exec_queue *)msg->private_data)->guc->id; + __entry->gt_id = + ((struct xe_exec_queue *)msg->private_data)->gt->info.id; ), - TP_printk("dev=%s, guc_id=%d, opcode=%u", __get_str(dev), __entry->guc_id, + TP_printk("dev=%s, gt=%u guc_id=%d, opcode=%u", __get_str(dev), __entry->gt_id, __entry->guc_id, __entry->opcode) ); -- 2.51.0 From ddb106d2120a0bf1c5ff87c71d059d193814da41 Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Fri, 22 Nov 2024 16:19:16 +0000 Subject: [PATCH 05/16] drm/xe/guc_submit: fix race around pending_disable Currently in some testcases we can trigger: [drm] *ERROR* GT0: SCHED_DONE: Unexpected engine state 0x02b1, guc_id=8, runnable_state=0 [drm] *ERROR* GT0: G2H action 0x1002 failed (-EPROTO) len 3 msg 02 10 00 90 08 00 00 00 00 00 00 00 Looking at a snippet of corresponding ftrace for this GuC id we can see: 498.852891: xe_sched_msg_add: dev=0000:03:00.0, gt=0 guc_id=8, opcode=3 498.854083: xe_sched_msg_recv: dev=0000:03:00.0, gt=0 guc_id=8, opcode=3 498.855389: xe_exec_queue_kill: dev=0000:03:00.0, 5:0x1, gt=0, width=1, guc_id=8, guc_state=0x3, flags=0x0 498.855436: xe_exec_queue_lr_cleanup: dev=0000:03:00.0, 5:0x1, gt=0, width=1, guc_id=8, guc_state=0x83, flags=0x0 498.856767: xe_exec_queue_close: dev=0000:03:00.0, 5:0x1, gt=0, width=1, guc_id=8, guc_state=0x83, flags=0x0 498.862889: xe_exec_queue_scheduling_disable: dev=0000:03:00.0, 5:0x1, gt=0, width=1, guc_id=8, guc_state=0xa9, flags=0x0 498.863032: xe_exec_queue_scheduling_disable: dev=0000:03:00.0, 5:0x1, gt=0, width=1, guc_id=8, guc_state=0x2b9, flags=0x0 498.875596: xe_exec_queue_scheduling_done: dev=0000:03:00.0, 5:0x1, gt=0, width=1, guc_id=8, guc_state=0x2b9, flags=0x0 498.875604: xe_exec_queue_deregister: dev=0000:03:00.0, 5:0x1, gt=0, width=1, guc_id=8, guc_state=0x2b1, flags=0x0 499.074483: xe_exec_queue_deregister_done: dev=0000:03:00.0, 5:0x1, gt=0, width=1, guc_id=8, guc_state=0x2b1, flags=0x0 This looks to be the two scheduling_disable racing with each other, one from the suspend (opcode=3) and then again during lr cleanup. While those two operations are serialized, the G2H portion is not, therefore when marking the queue as pending_disabled and then firing off the first request, we proceed do the same again, however the first disable response only fires after this which then clears the pending_disabled. At this point the second comes back and is processed, however the pending_disabled is no longer set, hence triggering the warning. To fix this wait for pending_disabled when doing the lr cleanup and calling disable_scheduling_deregister. Also do the same for all other disable_scheduling callers. Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs") Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/3515 Signed-off-by: Matthew Auld Cc: Matthew Brost Cc: # v6.8+ Reviewed-by: Matthew Brost Link: https://patchwork.freedesktop.org/patch/msgid/20241122161914.321263-5-matthew.auld@intel.com --- drivers/gpu/drm/xe/xe_guc_submit.c | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c index f9ecee5364d8..f3c22b101916 100644 --- a/drivers/gpu/drm/xe/xe_guc_submit.c +++ b/drivers/gpu/drm/xe/xe_guc_submit.c @@ -767,12 +767,15 @@ static void disable_scheduling_deregister(struct xe_guc *guc, set_min_preemption_timeout(guc, q); smp_rmb(); - ret = wait_event_timeout(guc->ct.wq, !exec_queue_pending_enable(q) || - xe_guc_read_stopped(guc), HZ * 5); + ret = wait_event_timeout(guc->ct.wq, + (!exec_queue_pending_enable(q) && + !exec_queue_pending_disable(q)) || + xe_guc_read_stopped(guc), + HZ * 5); if (!ret) { struct xe_gpu_scheduler *sched = &q->guc->sched; - xe_gt_warn(q->gt, "Pending enable failed to respond\n"); + xe_gt_warn(q->gt, "Pending enable/disable failed to respond\n"); xe_sched_submission_start(sched); xe_gt_reset_async(q->gt); xe_sched_tdr_queue_imm(sched); @@ -1099,7 +1102,8 @@ guc_exec_queue_timedout_job(struct drm_sched_job *drm_job) * modifying state */ ret = wait_event_timeout(guc->ct.wq, - !exec_queue_pending_enable(q) || + (!exec_queue_pending_enable(q) && + !exec_queue_pending_disable(q)) || xe_guc_read_stopped(guc), HZ * 5); if (!ret || xe_guc_read_stopped(guc)) goto trigger_reset; @@ -1329,8 +1333,8 @@ static void __guc_exec_queue_process_msg_suspend(struct xe_sched_msg *msg) if (guc_exec_queue_allowed_to_change_state(q) && !exec_queue_suspended(q) && exec_queue_enabled(q)) { - wait_event(guc->ct.wq, q->guc->resume_time != RESUME_PENDING || - xe_guc_read_stopped(guc)); + wait_event(guc->ct.wq, (q->guc->resume_time != RESUME_PENDING || + xe_guc_read_stopped(guc)) && !exec_queue_pending_disable(q)); if (!xe_guc_read_stopped(guc)) { s64 since_resume_ms = -- 2.51.0 From f161809b362f027b6d72bd998e47f8f0bad60a2e Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Fri, 22 Nov 2024 16:19:17 +0000 Subject: [PATCH 06/16] drm/xe/guc_submit: fix race around suspend_pending Currently in some testcases we can trigger: xe 0000:03:00.0: [drm] Assertion `exec_queue_destroyed(q)` failed! .... WARNING: CPU: 18 PID: 2640 at drivers/gpu/drm/xe/xe_guc_submit.c:1826 xe_guc_sched_done_handler+0xa54/0xef0 [xe] xe 0000:03:00.0: [drm] *ERROR* GT1: DEREGISTER_DONE: Unexpected engine state 0x00a1, guc_id=57 Looking at a snippet of corresponding ftrace for this GuC id we can see: 162.673311: xe_sched_msg_add: dev=0000:03:00.0, gt=1 guc_id=57, opcode=3 162.673317: xe_sched_msg_recv: dev=0000:03:00.0, gt=1 guc_id=57, opcode=3 162.673319: xe_exec_queue_scheduling_disable: dev=0000:03:00.0, 1:0x2, gt=1, width=1, guc_id=57, guc_state=0x29, flags=0x0 162.674089: xe_exec_queue_kill: dev=0000:03:00.0, 1:0x2, gt=1, width=1, guc_id=57, guc_state=0x29, flags=0x0 162.674108: xe_exec_queue_close: dev=0000:03:00.0, 1:0x2, gt=1, width=1, guc_id=57, guc_state=0xa9, flags=0x0 162.674488: xe_exec_queue_scheduling_done: dev=0000:03:00.0, 1:0x2, gt=1, width=1, guc_id=57, guc_state=0xa9, flags=0x0 162.678452: xe_exec_queue_deregister: dev=0000:03:00.0, 1:0x2, gt=1, width=1, guc_id=57, guc_state=0xa1, flags=0x0 It looks like we try to suspend the queue (opcode=3), setting suspend_pending and triggering a disable_scheduling. The user then closes the queue. However the close will also forcefully signal the suspend fence after killing the queue, later when the G2H response for disable_scheduling comes back we have now cleared suspend_pending when signalling the suspend fence, so the disable_scheduling now incorrectly tries to also deregister the queue. This leads to warnings since the queue has yet to even be marked for destruction. We also seem to trigger errors later with trying to double unregister the same queue. To fix this tweak the ordering when handling the response to ensure we don't race with a disable_scheduling that didn't actually intend to perform an unregister. The destruction path should now also correctly wait for any pending_disable before marking as destroyed. Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs") Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/3371 Signed-off-by: Matthew Auld Cc: Matthew Brost Cc: # v6.8+ Reviewed-by: Matthew Brost Link: https://patchwork.freedesktop.org/patch/msgid/20241122161914.321263-6-matthew.auld@intel.com --- drivers/gpu/drm/xe/xe_guc_submit.c | 17 +++++++++++++++-- 1 file changed, 15 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_guc_submit.c b/drivers/gpu/drm/xe/xe_guc_submit.c index f3c22b101916..d00af8d8acbe 100644 --- a/drivers/gpu/drm/xe/xe_guc_submit.c +++ b/drivers/gpu/drm/xe/xe_guc_submit.c @@ -1867,16 +1867,29 @@ static void handle_sched_done(struct xe_guc *guc, struct xe_exec_queue *q, xe_gt_assert(guc_to_gt(guc), runnable_state == 0); xe_gt_assert(guc_to_gt(guc), exec_queue_pending_disable(q)); - clear_exec_queue_pending_disable(q); if (q->guc->suspend_pending) { suspend_fence_signal(q); + clear_exec_queue_pending_disable(q); } else { if (exec_queue_banned(q) || check_timeout) { smp_wmb(); wake_up_all(&guc->ct.wq); } - if (!check_timeout) + if (!check_timeout && exec_queue_destroyed(q)) { + /* + * Make sure to clear the pending_disable only + * after sampling the destroyed state. We want + * to ensure we don't trigger the unregister too + * early with something intending to only + * disable scheduling. The caller doing the + * destroy must wait for an ongoing + * pending_disable before marking as destroyed. + */ + clear_exec_queue_pending_disable(q); deregister_exec_queue(guc, q); + } else { + clear_exec_queue_pending_disable(q); + } } } } -- 2.51.0 From f8c91d966372d7796b9908d4dd933f6900f27b76 Mon Sep 17 00:00:00 2001 From: Apoorva Singh Date: Thu, 7 Nov 2024 13:51:58 +0530 Subject: [PATCH 07/16] drm/xe/xe3lpg: Add Wa_16024792527 Force Sampler Tile64 Overfetch via MMIO Signed-off-by: Apoorva Singh Reviewed-by: Gustavo Sousa Link: https://patchwork.freedesktop.org/patch/msgid/20241107082158.1436637-1-apoorva.singh@intel.com Signed-off-by: Himal Prasad Ghimiray Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/regs/xe_gt_regs.h | 2 ++ drivers/gpu/drm/xe/xe_wa.c | 6 ++++++ 2 files changed, 8 insertions(+) diff --git a/drivers/gpu/drm/xe/regs/xe_gt_regs.h b/drivers/gpu/drm/xe/regs/xe_gt_regs.h index 0c9e4b2fafab..162f18e975da 100644 --- a/drivers/gpu/drm/xe/regs/xe_gt_regs.h +++ b/drivers/gpu/drm/xe/regs/xe_gt_regs.h @@ -445,6 +445,8 @@ #define SAMPLER_MODE XE_REG_MCR(0xe18c, XE_REG_OPTION_MASKED) #define ENABLE_SMALLPL REG_BIT(15) +#define SMP_WAIT_FETCH_MERGING_COUNTER REG_GENMASK(11, 10) +#define SMP_FORCE_128B_OVERFETCH REG_FIELD_PREP(SMP_WAIT_FETCH_MERGING_COUNTER, 1) #define SC_DISABLE_POWER_OPTIMIZATION_EBB REG_BIT(9) #define SAMPLER_ENABLE_HEADLESS_MSG REG_BIT(5) #define INDIRECT_STATE_BASE_ADDR_OVERRIDE REG_BIT(0) diff --git a/drivers/gpu/drm/xe/xe_wa.c b/drivers/gpu/drm/xe/xe_wa.c index 02cf647f86d8..570fe0376402 100644 --- a/drivers/gpu/drm/xe/xe_wa.c +++ b/drivers/gpu/drm/xe/xe_wa.c @@ -607,6 +607,12 @@ static const struct xe_rtp_entry_sr engine_was[] = { FUNC(xe_rtp_match_first_render_or_compute)), XE_RTP_ACTIONS(SET(ROW_CHICKEN4, DISABLE_TDL_PUSH)) }, + { XE_RTP_NAME("16024792527"), + XE_RTP_RULES(GRAPHICS_VERSION(3000), GRAPHICS_STEP(A0, B0), + FUNC(xe_rtp_match_first_render_or_compute)), + XE_RTP_ACTIONS(FIELD_SET(SAMPLER_MODE, SMP_WAIT_FETCH_MERGING_COUNTER, + SMP_FORCE_128B_OVERFETCH)) + }, {} }; -- 2.51.0 From f3dc9246f9c3cd5a7d8fd70cfd805bfc52214e2e Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Tue, 26 Nov 2024 18:13:00 +0000 Subject: [PATCH 08/16] drm/xe/migrate: fix pat index usage XE_CACHE_WB must be converted into the per-platform pat index for that particular caching mode, otherwise we are just encoding whatever happens to be the value of that enum. Fixes: e8babb280b5e ("drm/xe: Convert multiple bind ops into single job") Signed-off-by: Matthew Auld Cc: Matthew Brost Cc: Nirmoy Das Cc: # v6.12+ Reviewed-by: Nirmoy Das Reviewed-by: Matthew Brost Link: https://patchwork.freedesktop.org/patch/msgid/20241126181259.159713-3-matthew.auld@intel.com --- drivers/gpu/drm/xe/xe_migrate.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c index cfd31ae49cc1..48e205a40fd2 100644 --- a/drivers/gpu/drm/xe/xe_migrate.c +++ b/drivers/gpu/drm/xe/xe_migrate.c @@ -1350,6 +1350,7 @@ __xe_migrate_update_pgtables(struct xe_migrate *m, /* For sysmem PTE's, need to map them in our hole.. */ if (!IS_DGFX(xe)) { + u16 pat_index = xe->pat.idx[XE_CACHE_WB]; u32 ptes, ofs; ppgtt_ofs = NUM_KERNEL_PDE - 1; @@ -1409,7 +1410,7 @@ __xe_migrate_update_pgtables(struct xe_migrate *m, pt_bo->update_index = current_update; addr = vm->pt_ops->pte_encode_bo(pt_bo, 0, - XE_CACHE_WB, 0); + pat_index, 0); bb->cs[bb->len++] = lower_32_bits(addr); bb->cs[bb->len++] = upper_32_bits(addr); } -- 2.51.0 From febc689b27d28973cd02f667548a5dca383d859a Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Tue, 26 Nov 2024 18:13:01 +0000 Subject: [PATCH 09/16] drm/xe/migrate: use XE_BO_FLAG_PAGETABLE On some HW we want to avoid the host caching PTEs, since access from GPU side can be incoherent. However here the special migrate object is mapping PTEs which are written from the host and potentially cached. Use XE_BO_FLAG_PAGETABLE to ensure that non-cached mapping is used, on platforms where this matters. Fixes: 7a060d786cc1 ("drm/xe/mtl: Map PPGTT as CPU:WC") Signed-off-by: Matthew Auld Cc: Matthew Brost Cc: Nirmoy Das Cc: # v6.8+ Reviewed-by: Nirmoy Das Reviewed-by: Matthew Brost Link: https://patchwork.freedesktop.org/patch/msgid/20241126181259.159713-4-matthew.auld@intel.com --- drivers/gpu/drm/xe/xe_migrate.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c index 48e205a40fd2..1b97d90aadda 100644 --- a/drivers/gpu/drm/xe/xe_migrate.c +++ b/drivers/gpu/drm/xe/xe_migrate.c @@ -209,7 +209,8 @@ static int xe_migrate_prepare_vm(struct xe_tile *tile, struct xe_migrate *m, num_entries * XE_PAGE_SIZE, ttm_bo_type_kernel, XE_BO_FLAG_VRAM_IF_DGFX(tile) | - XE_BO_FLAG_PINNED); + XE_BO_FLAG_PINNED | + XE_BO_FLAG_PAGETABLE); if (IS_ERR(bo)) return PTR_ERR(bo); -- 2.51.0 From f85dc3c5abddaa6ef3674755f28cf49774d27ea5 Mon Sep 17 00:00:00 2001 From: Matthew Brost Date: Tue, 26 Nov 2024 09:46:08 -0800 Subject: [PATCH 10/16] drm/xe: Add xe_bo_vm_access MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Add xe_bo_vm_access which is wrapper around ttm_bo_vm_access which takes rpm refs for device access. Suggested-by: Thomas Hellström Signed-off-by: Matthew Brost Reviewed-by: Matthew Auld Link: https://patchwork.freedesktop.org/patch/msgid/20241126174615.2665852-2-matthew.brost@intel.com --- drivers/gpu/drm/xe/xe_bo.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c index ec070af12662..aaf54131c89e 100644 --- a/drivers/gpu/drm/xe/xe_bo.c +++ b/drivers/gpu/drm/xe/xe_bo.c @@ -1249,11 +1249,26 @@ out: return ret; } +static int xe_bo_vm_access(struct vm_area_struct *vma, unsigned long addr, + void *buf, int len, int write) +{ + struct ttm_buffer_object *ttm_bo = vma->vm_private_data; + struct xe_bo *bo = ttm_to_xe_bo(ttm_bo); + struct xe_device *xe = xe_bo_device(bo); + int ret; + + xe_pm_runtime_get(xe); + ret = ttm_bo_vm_access(vma, addr, buf, len, write); + xe_pm_runtime_put(xe); + + return ret; +} + static const struct vm_operations_struct xe_gem_vm_ops = { .fault = xe_gem_fault, .open = ttm_bo_vm_open, .close = ttm_bo_vm_close, - .access = ttm_bo_vm_access + .access = xe_bo_vm_access, }; static const struct drm_gem_object_funcs xe_gem_object_funcs = { -- 2.51.0 From 7d08df5d0bd3d12d14dcec773fcddbe3eed3a8e8 Mon Sep 17 00:00:00 2001 From: Matthew Brost Date: Tue, 26 Nov 2024 09:46:09 -0800 Subject: [PATCH 11/16] drm/ttm: Add ttm_bo_access MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Non-contiguous VRAM cannot easily be mapped in TTM nor can non-visible VRAM easily be accessed. Add ttm_bo_access, which is similar to ttm_bo_vm_access, to access such memory. v4: - Fix checkpatch warnings (CI) v5: - Fix checkpatch warnings (CI) v6: - Fix kernel doc (Auld) v7: - Move ttm_bo_access to ttm_bo_vm.c (Christian) Cc: Christian König Reported-by: Christoph Manszewski Suggested-by: Thomas Hellström Signed-off-by: Matthew Brost Tested-by: Mika Kuoppala Reviewed-by: Matthew Auld Reviewed-by: Christian König Link: https://patchwork.freedesktop.org/patch/msgid/20241126174615.2665852-3-matthew.brost@intel.com --- drivers/gpu/drm/ttm/ttm_bo_vm.c | 40 ++++++++++++++++++++++++++------- include/drm/ttm/ttm_bo.h | 2 ++ 2 files changed, 34 insertions(+), 8 deletions(-) diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c index 2c699ed1963a..f02d3966cc84 100644 --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c @@ -405,13 +405,25 @@ static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo, return len; } -int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr, - void *buf, int len, int write) +/** + * ttm_bo_access - Helper to access a buffer object + * + * @bo: ttm buffer object + * @offset: access offset into buffer object + * @buf: pointer to caller memory to read into or write from + * @len: length of access + * @write: write access + * + * Utility function to access a buffer object. Useful when buffer object cannot + * be easily mapped (non-contiguous, non-visible, etc...). Should not directly + * be exported to user space via a peak / poke interface. + * + * Returns: + * @len if successful, negative error code on failure. + */ +int ttm_bo_access(struct ttm_buffer_object *bo, unsigned long offset, + void *buf, int len, int write) { - struct ttm_buffer_object *bo = vma->vm_private_data; - unsigned long offset = (addr) - vma->vm_start + - ((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node)) - << PAGE_SHIFT); int ret; if (len < 1 || (offset + len) > bo->base.size) @@ -429,8 +441,8 @@ int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr, break; default: if (bo->bdev->funcs->access_memory) - ret = bo->bdev->funcs->access_memory( - bo, offset, buf, len, write); + ret = bo->bdev->funcs->access_memory + (bo, offset, buf, len, write); else ret = -EIO; } @@ -439,6 +451,18 @@ int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr, return ret; } +EXPORT_SYMBOL(ttm_bo_access); + +int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr, + void *buf, int len, int write) +{ + struct ttm_buffer_object *bo = vma->vm_private_data; + unsigned long offset = (addr) - vma->vm_start + + ((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node)) + << PAGE_SHIFT); + + return ttm_bo_access(bo, offset, buf, len, write); +} EXPORT_SYMBOL(ttm_bo_vm_access); static const struct vm_operations_struct ttm_bo_vm_ops = { diff --git a/include/drm/ttm/ttm_bo.h b/include/drm/ttm/ttm_bo.h index 5804408815be..8ea11cd8df39 100644 --- a/include/drm/ttm/ttm_bo.h +++ b/include/drm/ttm/ttm_bo.h @@ -421,6 +421,8 @@ void ttm_bo_unpin(struct ttm_buffer_object *bo); int ttm_bo_evict_first(struct ttm_device *bdev, struct ttm_resource_manager *man, struct ttm_operation_ctx *ctx); +int ttm_bo_access(struct ttm_buffer_object *bo, unsigned long offset, + void *buf, int len, int write); vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo, struct vm_fault *vmf); vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, -- 2.51.0 From 97e8cecb6a0dde7ebb8e8e973146b5a131f52b9f Mon Sep 17 00:00:00 2001 From: Matthew Brost Date: Tue, 26 Nov 2024 09:46:10 -0800 Subject: [PATCH 12/16] drm/xe: Add xe_ttm_access_memory MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Non-contiguous VRAM cannot easily be mapped in TTM nor can non-visible VRAM easily be accessed. Add xe_ttm_access_memory which hooks into ttm_bo_access to access such memory. v4: - Assert memory access rather than taking RPM ref (Thomas / Auld) - Fix warning on xe_res_cursor.h for non-zero offset (Mika) Reported-by: Christoph Manszewski Suggested-by: Thomas Hellström Signed-off-by: Matthew Brost Reviewed-by: Matthew Auld Link: https://patchwork.freedesktop.org/patch/msgid/20241126174615.2665852-4-matthew.brost@intel.com --- drivers/gpu/drm/xe/xe_bo.c | 59 ++++++++++++++++++++++++++++++++++++-- 1 file changed, 56 insertions(+), 3 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c index aaf54131c89e..9de28798ad1c 100644 --- a/drivers/gpu/drm/xe/xe_bo.c +++ b/drivers/gpu/drm/xe/xe_bo.c @@ -442,6 +442,14 @@ static void xe_ttm_tt_destroy(struct ttm_device *ttm_dev, struct ttm_tt *tt) kfree(tt); } +static bool xe_ttm_resource_visible(struct ttm_resource *mem) +{ + struct xe_ttm_vram_mgr_resource *vres = + to_xe_ttm_vram_mgr_resource(mem); + + return vres->used_visible_size == mem->size; +} + static int xe_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem) { @@ -453,11 +461,9 @@ static int xe_ttm_io_mem_reserve(struct ttm_device *bdev, return 0; case XE_PL_VRAM0: case XE_PL_VRAM1: { - struct xe_ttm_vram_mgr_resource *vres = - to_xe_ttm_vram_mgr_resource(mem); struct xe_mem_region *vram = res_to_mem_region(mem); - if (vres->used_visible_size < mem->size) + if (!xe_ttm_resource_visible(mem)) return -EINVAL; mem->bus.offset = mem->start << PAGE_SHIFT; @@ -1121,6 +1127,52 @@ static void xe_ttm_bo_swap_notify(struct ttm_buffer_object *ttm_bo) } } +static int xe_ttm_access_memory(struct ttm_buffer_object *ttm_bo, + unsigned long offset, void *buf, int len, + int write) +{ + struct xe_bo *bo = ttm_to_xe_bo(ttm_bo); + struct xe_device *xe = ttm_to_xe_device(ttm_bo->bdev); + struct iosys_map vmap; + struct xe_res_cursor cursor; + struct xe_mem_region *vram; + int bytes_left = len; + + xe_bo_assert_held(bo); + xe_device_assert_mem_access(xe); + + if (!mem_type_is_vram(ttm_bo->resource->mem_type)) + return -EIO; + + /* FIXME: Use GPU for non-visible VRAM */ + if (!xe_ttm_resource_visible(ttm_bo->resource)) + return -EIO; + + vram = res_to_mem_region(ttm_bo->resource); + xe_res_first(ttm_bo->resource, offset & PAGE_MASK, + bo->size - (offset & PAGE_MASK), &cursor); + + do { + unsigned long page_offset = (offset & ~PAGE_MASK); + int byte_count = min((int)(PAGE_SIZE - page_offset), bytes_left); + + iosys_map_set_vaddr_iomem(&vmap, (u8 __iomem *)vram->mapping + + cursor.start); + if (write) + xe_map_memcpy_to(xe, &vmap, page_offset, buf, byte_count); + else + xe_map_memcpy_from(xe, buf, &vmap, page_offset, byte_count); + + buf += byte_count; + offset += byte_count; + bytes_left -= byte_count; + if (bytes_left) + xe_res_next(&cursor, PAGE_SIZE); + } while (bytes_left); + + return len; +} + const struct ttm_device_funcs xe_ttm_funcs = { .ttm_tt_create = xe_ttm_tt_create, .ttm_tt_populate = xe_ttm_tt_populate, @@ -1130,6 +1182,7 @@ const struct ttm_device_funcs xe_ttm_funcs = { .move = xe_bo_move, .io_mem_reserve = xe_ttm_io_mem_reserve, .io_mem_pfn = xe_ttm_io_mem_pfn, + .access_memory = xe_ttm_access_memory, .release_notify = xe_ttm_bo_release_notify, .eviction_valuable = ttm_bo_eviction_valuable, .delete_mem_notify = xe_ttm_bo_delete_mem_notify, -- 2.51.0 From 1c6878af115a4586a40d6c14d530fa9f93e0bd83 Mon Sep 17 00:00:00 2001 From: Matthew Brost Date: Tue, 26 Nov 2024 09:46:11 -0800 Subject: [PATCH 13/16] drm/xe: Take PM ref in delayed snapshot capture worker The delayed snapshot capture worker can access the GPU or VRAM both of which require a PM reference. Take a reference in this worker. Cc: Rodrigo Vivi Cc: Maarten Lankhorst Fixes: 4f04d07c0a94 ("drm/xe: Faster devcoredump") Signed-off-by: Matthew Brost Reviewed-by: Matthew Auld Link: https://patchwork.freedesktop.org/patch/msgid/20241126174615.2665852-5-matthew.brost@intel.com --- drivers/gpu/drm/xe/xe_devcoredump.c | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/drivers/gpu/drm/xe/xe_devcoredump.c b/drivers/gpu/drm/xe/xe_devcoredump.c index 0e5edf14a241..0d3a4dc87a7f 100644 --- a/drivers/gpu/drm/xe/xe_devcoredump.c +++ b/drivers/gpu/drm/xe/xe_devcoredump.c @@ -23,6 +23,7 @@ #include "xe_guc_submit.h" #include "xe_hw_engine.h" #include "xe_module.h" +#include "xe_pm.h" #include "xe_sched_job.h" #include "xe_vm.h" @@ -167,8 +168,11 @@ static void xe_devcoredump_deferred_snap_work(struct work_struct *work) { struct xe_devcoredump_snapshot *ss = container_of(work, typeof(*ss), work); struct xe_devcoredump *coredump = container_of(ss, typeof(*coredump), snapshot); + struct xe_device *xe = coredump_to_xe(coredump); unsigned int fw_ref; + xe_pm_runtime_get(xe); + /* keep going if fw fails as we still want to save the memory and SW data */ fw_ref = xe_force_wake_get(gt_to_fw(ss->gt), XE_FORCEWAKE_ALL); if (!xe_force_wake_ref_has_domain(fw_ref, XE_FORCEWAKE_ALL)) @@ -177,6 +181,8 @@ static void xe_devcoredump_deferred_snap_work(struct work_struct *work) xe_guc_exec_queue_snapshot_capture_delayed(ss->ge); xe_force_wake_put(gt_to_fw(ss->gt), fw_ref); + xe_pm_runtime_put(xe); + /* Calculate devcoredump size */ ss->read.size = __xe_devcoredump_read(NULL, INT_MAX, coredump); -- 2.51.0 From b6308aaa24a7ad3dfc6157b6afc550b9ab7e8945 Mon Sep 17 00:00:00 2001 From: Matthew Brost Date: Tue, 26 Nov 2024 09:46:12 -0800 Subject: [PATCH 14/16] drm/xe/display: Update intel_bo_read_from_page to use ttm_bo_access Don't open code vmap of a BO, use ttm_bo_access helper which is safe for non-contiguous BOs and non-visible BOs. Suggested-by: Matthew Auld Signed-off-by: Matthew Brost Reviewed-by: Matthew Auld Link: https://patchwork.freedesktop.org/patch/msgid/20241126174615.2665852-6-matthew.brost@intel.com --- drivers/gpu/drm/xe/display/intel_bo.c | 25 +------------------------ 1 file changed, 1 insertion(+), 24 deletions(-) diff --git a/drivers/gpu/drm/xe/display/intel_bo.c b/drivers/gpu/drm/xe/display/intel_bo.c index 9f54fad0f1c0..43141964f6f2 100644 --- a/drivers/gpu/drm/xe/display/intel_bo.c +++ b/drivers/gpu/drm/xe/display/intel_bo.c @@ -40,31 +40,8 @@ int intel_bo_fb_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma) int intel_bo_read_from_page(struct drm_gem_object *obj, u64 offset, void *dst, int size) { struct xe_bo *bo = gem_to_xe_bo(obj); - struct ttm_bo_kmap_obj map; - void *src; - bool is_iomem; - int ret; - ret = xe_bo_lock(bo, true); - if (ret) - return ret; - - ret = ttm_bo_kmap(&bo->ttm, offset >> PAGE_SHIFT, 1, &map); - if (ret) - goto out_unlock; - - offset &= ~PAGE_MASK; - src = ttm_kmap_obj_virtual(&map, &is_iomem); - src += offset; - if (is_iomem) - memcpy_fromio(dst, (void __iomem *)src, size); - else - memcpy(dst, src, size); - - ttm_bo_kunmap(&map); -out_unlock: - xe_bo_unlock(bo); - return ret; + return ttm_bo_access(&bo->ttm, offset, dst, size, 0); } struct intel_frontbuffer *intel_bo_get_frontbuffer(struct drm_gem_object *obj) -- 2.51.0 From 5f7bec831f1f17c354e4307a12cf79b018296975 Mon Sep 17 00:00:00 2001 From: Matthew Brost Date: Tue, 26 Nov 2024 09:46:13 -0800 Subject: [PATCH 15/16] drm/xe: Use ttm_bo_access in xe_vm_snapshot_capture_delayed Non-contiguous mapping of BO in VRAM doesn't work, use ttm_bo_access instead. v2: - Fix error handling Fixes: 0eb2a18a8fad ("drm/xe: Implement VM snapshot support for BO's and userptr") Suggested-by: Matthew Auld Signed-off-by: Matthew Brost Reviewed-by: Matthew Auld Link: https://patchwork.freedesktop.org/patch/msgid/20241126174615.2665852-7-matthew.brost@intel.com --- drivers/gpu/drm/xe/xe_vm.c | 17 ++++++----------- 1 file changed, 6 insertions(+), 11 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c index 2e67648ed512..610226c7c1ce 100644 --- a/drivers/gpu/drm/xe/xe_vm.c +++ b/drivers/gpu/drm/xe/xe_vm.c @@ -3310,7 +3310,6 @@ void xe_vm_snapshot_capture_delayed(struct xe_vm_snapshot *snap) for (int i = 0; i < snap->num_snaps; i++) { struct xe_bo *bo = snap->snap[i].bo; - struct iosys_map src; int err; if (IS_ERR(snap->snap[i].data)) @@ -3323,16 +3322,12 @@ void xe_vm_snapshot_capture_delayed(struct xe_vm_snapshot *snap) } if (bo) { - xe_bo_lock(bo, false); - err = ttm_bo_vmap(&bo->ttm, &src); - if (!err) { - xe_map_memcpy_from(xe_bo_device(bo), - snap->snap[i].data, - &src, snap->snap[i].bo_ofs, - snap->snap[i].len); - ttm_bo_vunmap(&bo->ttm, &src); - } - xe_bo_unlock(bo); + err = ttm_bo_access(&bo->ttm, snap->snap[i].bo_ofs, + snap->snap[i].data, snap->snap[i].len, 0); + if (!(err < 0) && err != snap->snap[i].len) + err = -EIO; + else if (!(err < 0)) + err = 0; } else { void __user *userptr = (void __user *)(size_t)snap->snap[i].bo_ofs; -- 2.51.0 From e03b0aa67ac0106d8961581a426649fabab50827 Mon Sep 17 00:00:00 2001 From: Matthew Brost Date: Tue, 26 Nov 2024 09:46:14 -0800 Subject: [PATCH 16/16] drm/xe: Set XE_BO_FLAG_PINNED in migrate selftest BOs We only allow continguous BOs to be vmapped, set XE_BO_FLAG_PINNED on BOs in migrate selftest as this forces continguous BOs and selftest uses vmaps. Signed-off-by: Matthew Brost Reviewed-by: Matthew Auld Link: https://patchwork.freedesktop.org/patch/msgid/20241126174615.2665852-8-matthew.brost@intel.com --- drivers/gpu/drm/xe/tests/xe_migrate.c | 13 +++++++++---- 1 file changed, 9 insertions(+), 4 deletions(-) diff --git a/drivers/gpu/drm/xe/tests/xe_migrate.c b/drivers/gpu/drm/xe/tests/xe_migrate.c index 1a192a2a941b..4cef3b20bd17 100644 --- a/drivers/gpu/drm/xe/tests/xe_migrate.c +++ b/drivers/gpu/drm/xe/tests/xe_migrate.c @@ -83,7 +83,8 @@ static void test_copy(struct xe_migrate *m, struct xe_bo *bo, bo->size, ttm_bo_type_kernel, region | - XE_BO_FLAG_NEEDS_CPU_ACCESS); + XE_BO_FLAG_NEEDS_CPU_ACCESS | + XE_BO_FLAG_PINNED); if (IS_ERR(remote)) { KUNIT_FAIL(test, "Failed to allocate remote bo for %s: %pe\n", str, remote); @@ -642,7 +643,9 @@ static void validate_ccs_test_run_tile(struct xe_device *xe, struct xe_tile *til sys_bo = xe_bo_create_user(xe, NULL, NULL, SZ_4M, DRM_XE_GEM_CPU_CACHING_WC, - XE_BO_FLAG_SYSTEM | XE_BO_FLAG_NEEDS_CPU_ACCESS); + XE_BO_FLAG_SYSTEM | + XE_BO_FLAG_NEEDS_CPU_ACCESS | + XE_BO_FLAG_PINNED); if (IS_ERR(sys_bo)) { KUNIT_FAIL(test, "xe_bo_create() failed with err=%ld\n", @@ -666,7 +669,8 @@ static void validate_ccs_test_run_tile(struct xe_device *xe, struct xe_tile *til ccs_bo = xe_bo_create_user(xe, NULL, NULL, SZ_4M, DRM_XE_GEM_CPU_CACHING_WC, - bo_flags | XE_BO_FLAG_NEEDS_CPU_ACCESS); + bo_flags | XE_BO_FLAG_NEEDS_CPU_ACCESS | + XE_BO_FLAG_PINNED); if (IS_ERR(ccs_bo)) { KUNIT_FAIL(test, "xe_bo_create() failed with err=%ld\n", @@ -690,7 +694,8 @@ static void validate_ccs_test_run_tile(struct xe_device *xe, struct xe_tile *til vram_bo = xe_bo_create_user(xe, NULL, NULL, SZ_4M, DRM_XE_GEM_CPU_CACHING_WC, - bo_flags | XE_BO_FLAG_NEEDS_CPU_ACCESS); + bo_flags | XE_BO_FLAG_NEEDS_CPU_ACCESS | + XE_BO_FLAG_PINNED); if (IS_ERR(vram_bo)) { KUNIT_FAIL(test, "xe_bo_create() failed with err=%ld\n", PTR_ERR(vram_bo)); -- 2.51.0