From 6405f5b70b1c240ffddef01c7a140498f47d4fe7 Mon Sep 17 00:00:00 2001 From: Lucas De Marchi Date: Wed, 9 Apr 2025 21:59:34 -0700 Subject: drm/xe: Set LRC addresses before guc load The metadata saved in the ADS is read by GuC when it's initialized. Saving the addresses to the LRCs when they are populated is too late as GuC will keep using the old ones. This was causing GuC to use the RCS LRC for any engine class. It's not a big problem on a Linux-only scenario since the they are used by GuC only on media engines when the watchdog is triggered. However, in a virtualization scenario with Windows as the VF, it causes the wrong LRCs to be loaded as the watchdog is used for all engines. Fix it by letting guc_golden_lrc_init() initialize the metadata, like other *_init() functions, and later guc_golden_lrc_populate() to copy the LRCs to the right places. The former is called before the second GuC load, while the latter is called after LRCs have been recorded. Cc: Chee Yin Wong Cc: John Harrison Cc: Matt Roper Cc: Matthew Brost Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs") Cc: # v6.11+ Reviewed-by: Matthew Brost Tested-by: Chee Yin Wong Link: https://lore.kernel.org/r/20250409-fix-guc-ads-v1-1-494135f7a5d0@intel.com Signed-off-by: Lucas De Marchi (cherry picked from commit c31a0b6402d15b530514eee9925adfcb8cfbb1c9) Signed-off-by: Lucas De Marchi --- drivers/gpu/drm/xe/xe_guc_ads.c | 75 ++++++++++++++++++++++++----------------- 1 file changed, 45 insertions(+), 30 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_guc_ads.c b/drivers/gpu/drm/xe/xe_guc_ads.c index e7c9e095a19f..7031542a70ce 100644 --- a/drivers/gpu/drm/xe/xe_guc_ads.c +++ b/drivers/gpu/drm/xe/xe_guc_ads.c @@ -490,24 +490,52 @@ static void fill_engine_enable_masks(struct xe_gt *gt, engine_enable_mask(gt, XE_ENGINE_CLASS_OTHER)); } -static void guc_prep_golden_lrc_null(struct xe_guc_ads *ads) +/* + * Write the offsets corresponding to the golden LRCs. The actual data is + * populated later by guc_golden_lrc_populate() + */ +static void guc_golden_lrc_init(struct xe_guc_ads *ads) { struct xe_device *xe = ads_to_xe(ads); + struct xe_gt *gt = ads_to_gt(ads); struct iosys_map info_map = IOSYS_MAP_INIT_OFFSET(ads_to_map(ads), offsetof(struct __guc_ads_blob, system_info)); - u8 guc_class; + size_t alloc_size, real_size; + u32 addr_ggtt, offset; + int class; + + offset = guc_ads_golden_lrc_offset(ads); + addr_ggtt = xe_bo_ggtt_addr(ads->bo) + offset; + + for (class = 0; class < XE_ENGINE_CLASS_MAX; ++class) { + u8 guc_class; + + guc_class = xe_engine_class_to_guc_class(class); - for (guc_class = 0; guc_class <= GUC_MAX_ENGINE_CLASSES; ++guc_class) { if (!info_map_read(xe, &info_map, engine_enabled_masks[guc_class])) continue; + real_size = xe_gt_lrc_size(gt, class); + alloc_size = PAGE_ALIGN(real_size); + + /* + * This interface is slightly confusing. We need to pass the + * base address of the full golden context and the size of just + * the engine state, which is the section of the context image + * that starts after the execlists LRC registers. This is + * required to allow the GuC to restore just the engine state + * when a watchdog reset occurs. + * We calculate the engine state size by removing the size of + * what comes before it in the context image (which is identical + * on all engines). + */ ads_blob_write(ads, ads.eng_state_size[guc_class], - guc_ads_golden_lrc_size(ads) - - xe_lrc_skip_size(xe)); + real_size - xe_lrc_skip_size(xe)); ads_blob_write(ads, ads.golden_context_lrca[guc_class], - xe_bo_ggtt_addr(ads->bo) + - guc_ads_golden_lrc_offset(ads)); + addr_ggtt); + + addr_ggtt += alloc_size; } } @@ -857,7 +885,7 @@ void xe_guc_ads_populate_minimal(struct xe_guc_ads *ads) xe_map_memset(ads_to_xe(ads), ads_to_map(ads), 0, 0, ads->bo->size); guc_policies_init(ads); - guc_prep_golden_lrc_null(ads); + guc_golden_lrc_init(ads); guc_mapping_table_init_invalid(gt, &info_map); guc_doorbell_init(ads); @@ -883,7 +911,7 @@ void xe_guc_ads_populate(struct xe_guc_ads *ads) guc_policies_init(ads); fill_engine_enable_masks(gt, &info_map); guc_mmio_reg_state_init(ads); - guc_prep_golden_lrc_null(ads); + guc_golden_lrc_init(ads); guc_mapping_table_init(gt, &info_map); guc_capture_prep_lists(ads); guc_doorbell_init(ads); @@ -903,18 +931,22 @@ void xe_guc_ads_populate(struct xe_guc_ads *ads) guc_ads_private_data_offset(ads)); } -static void guc_populate_golden_lrc(struct xe_guc_ads *ads) +/* + * After the golden LRC's are recorded for each engine class by the first + * submission, copy them to the ADS, as initialized earlier by + * guc_golden_lrc_init(). + */ +static void guc_golden_lrc_populate(struct xe_guc_ads *ads) { struct xe_device *xe = ads_to_xe(ads); struct xe_gt *gt = ads_to_gt(ads); struct iosys_map info_map = IOSYS_MAP_INIT_OFFSET(ads_to_map(ads), offsetof(struct __guc_ads_blob, system_info)); size_t total_size = 0, alloc_size, real_size; - u32 addr_ggtt, offset; + u32 offset; int class; offset = guc_ads_golden_lrc_offset(ads); - addr_ggtt = xe_bo_ggtt_addr(ads->bo) + offset; for (class = 0; class < XE_ENGINE_CLASS_MAX; ++class) { u8 guc_class; @@ -931,26 +963,9 @@ static void guc_populate_golden_lrc(struct xe_guc_ads *ads) alloc_size = PAGE_ALIGN(real_size); total_size += alloc_size; - /* - * This interface is slightly confusing. We need to pass the - * base address of the full golden context and the size of just - * the engine state, which is the section of the context image - * that starts after the execlists LRC registers. This is - * required to allow the GuC to restore just the engine state - * when a watchdog reset occurs. - * We calculate the engine state size by removing the size of - * what comes before it in the context image (which is identical - * on all engines). - */ - ads_blob_write(ads, ads.eng_state_size[guc_class], - real_size - xe_lrc_skip_size(xe)); - ads_blob_write(ads, ads.golden_context_lrca[guc_class], - addr_ggtt); - xe_map_memcpy_to(xe, ads_to_map(ads), offset, gt->default_lrc[class], real_size); - addr_ggtt += alloc_size; offset += alloc_size; } @@ -959,7 +974,7 @@ static void guc_populate_golden_lrc(struct xe_guc_ads *ads) void xe_guc_ads_populate_post_load(struct xe_guc_ads *ads) { - guc_populate_golden_lrc(ads); + guc_golden_lrc_populate(ads); } static int guc_ads_action_update_policies(struct xe_guc_ads *ads, u32 policy_offset) -- cgit v1.2.3 From 2577b202458cddff85cc154b1fe7f313e0d1f418 Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Mon, 14 Apr 2025 14:25:40 +0100 Subject: drm/xe/userptr: fix notifier vs folio deadlock MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit User is reporting what smells like notifier vs folio deadlock, where migrate_pages_batch() on core kernel side is holding folio lock(s) and then interacting with the mappings of it, however those mappings are tied to some userptr, which means calling into the notifier callback and grabbing the notifier lock. With perfect timing it looks possible that the pages we pulled from the hmm fault can get sniped by migrate_pages_batch() at the same time that we are holding the notifier lock to mark the pages as accessed/dirty, but at this point we also want to grab the folio locks(s) to mark them as dirty, but if they are contended from notifier/migrate_pages_batch side then we deadlock since folio lock won't be dropped until we drop the notifier lock. Fortunately the mark_page_accessed/dirty is not really needed in the first place it seems and should have already been done by hmm fault, so just remove it. Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4765 Fixes: 0a98219bcc96 ("drm/xe/hmm: Don't dereference struct page pointers without notifier lock") Signed-off-by: Matthew Auld Cc: Thomas Hellström Cc: Matthew Brost Cc: # v6.10+ Reviewed-by: Thomas Hellström Reviewed-by: Matthew Brost Link: https://lore.kernel.org/r/20250414132539.26654-2-matthew.auld@intel.com (cherry picked from commit bd7c0cb695e87c0e43247be8196b4919edbe0e85) Signed-off-by: Lucas De Marchi --- drivers/gpu/drm/xe/xe_hmm.c | 24 ------------------------ 1 file changed, 24 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_hmm.c b/drivers/gpu/drm/xe/xe_hmm.c index c3cc0fa105e8..57b71956ddf4 100644 --- a/drivers/gpu/drm/xe/xe_hmm.c +++ b/drivers/gpu/drm/xe/xe_hmm.c @@ -19,29 +19,6 @@ static u64 xe_npages_in_range(unsigned long start, unsigned long end) return (end - start) >> PAGE_SHIFT; } -/** - * xe_mark_range_accessed() - mark a range is accessed, so core mm - * have such information for memory eviction or write back to - * hard disk - * @range: the range to mark - * @write: if write to this range, we mark pages in this range - * as dirty - */ -static void xe_mark_range_accessed(struct hmm_range *range, bool write) -{ - struct page *page; - u64 i, npages; - - npages = xe_npages_in_range(range->start, range->end); - for (i = 0; i < npages; i++) { - page = hmm_pfn_to_page(range->hmm_pfns[i]); - if (write) - set_page_dirty_lock(page); - - mark_page_accessed(page); - } -} - static int xe_alloc_sg(struct xe_device *xe, struct sg_table *st, struct hmm_range *range, struct rw_semaphore *notifier_sem) { @@ -331,7 +308,6 @@ int xe_hmm_userptr_populate_range(struct xe_userptr_vma *uvma, if (ret) goto out_unlock; - xe_mark_range_accessed(&hmm_range, write); userptr->sg = &userptr->sgt; xe_hmm_userptr_set_mapped(uvma); userptr->notifier_seq = hmm_range.notifier_seq; -- cgit v1.2.3 From 25583ad42d091819157832e894179200ba8b54ee Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Thu, 10 Apr 2025 17:27:17 +0100 Subject: drm/xe/dma_buf: stop relying on placement in unmap MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The is_vram() is checking the current placement, however if we consider exported VRAM with dynamic dma-buf, it looks possible for the xe driver to async evict the memory, notifying the importer, however importer does not have to call unmap_attachment() immediately, but rather just as "soon as possible", like when the dma-resv idles. Following from this we would then pipeline the move, attaching the fence to the manager, and then update the current placement. But when the unmap_attachment() runs at some later point we might see that is_vram() is now false, and take the complete wrong path when dma-unmapping the sg, leading to explosions. To fix this check if the sgl was mapping a struct page. v2: - The attachment can be mapped multiple times it seems, so we can't really rely on encoding something in the attachment->priv. Instead see if the page_link has an encoded struct page. For vram we expect this to be NULL. Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4563 Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs") Signed-off-by: Matthew Auld Cc: Thomas Hellström Cc: Matthew Brost Cc: # v6.8+ Acked-by: Christian König Link: https://lore.kernel.org/r/20250410162716.159403-2-matthew.auld@intel.com (cherry picked from commit d755887f8e5a2a18e15e6632a5193e5feea18499) Signed-off-by: Lucas De Marchi --- drivers/gpu/drm/xe/xe_dma_buf.c | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c index f67803e15a0e..f7a20264ea33 100644 --- a/drivers/gpu/drm/xe/xe_dma_buf.c +++ b/drivers/gpu/drm/xe/xe_dma_buf.c @@ -145,10 +145,7 @@ static void xe_dma_buf_unmap(struct dma_buf_attachment *attach, struct sg_table *sgt, enum dma_data_direction dir) { - struct dma_buf *dma_buf = attach->dmabuf; - struct xe_bo *bo = gem_to_xe_bo(dma_buf->priv); - - if (!xe_bo_is_vram(bo)) { + if (sg_page(sgt->sgl)) { dma_unmap_sgtable(attach->dev, sgt, dir, 0); sg_free_table(sgt); kfree(sgt); -- cgit v1.2.3 From 78600df8f593407a3df2d6c48c35d0ad203d7fb4 Mon Sep 17 00:00:00 2001 From: Daniele Ceraolo Spurio Date: Wed, 16 Apr 2025 13:16:22 -0700 Subject: drm/xe/pxp: do not queue unneeded terminations from debugfs The PXP terminate debugfs currently unconditionally simulates a termination, no matter what the HW status is. This is unneeded if PXP is not in use and can cause errors if the HW init hasn't completed yet. To solve these issues, we can simply limit the terminations to the cases where PXP is fully initialized and in use. v2: s/pxp_status/ready/ to avoid confusion with pxp->status (John) Fixes: 385a8015b214 ("drm/xe/pxp: Add PXP debugfs support") Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4749 Signed-off-by: Daniele Ceraolo Spurio Cc: John Harrison Reviewed-by: John Harrison Link: https://lore.kernel.org/r/20250416201622.1295369-1-daniele.ceraolospurio@intel.com (cherry picked from commit ba1f62a0cac84757ca35f4217e3cd3a2654233ae) Signed-off-by: Lucas De Marchi --- drivers/gpu/drm/xe/xe_pxp_debugfs.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_pxp_debugfs.c b/drivers/gpu/drm/xe/xe_pxp_debugfs.c index ccfbacf08efc..525a2f6bb076 100644 --- a/drivers/gpu/drm/xe/xe_pxp_debugfs.c +++ b/drivers/gpu/drm/xe/xe_pxp_debugfs.c @@ -66,9 +66,18 @@ static int pxp_terminate(struct seq_file *m, void *data) { struct xe_pxp *pxp = node_to_pxp(m->private); struct drm_printer p = drm_seq_file_printer(m); + int ready = xe_pxp_get_readiness_status(pxp); - if (!xe_pxp_is_enabled(pxp)) - return -ENODEV; + if (ready < 0) + return ready; /* disabled or error occurred */ + else if (!ready) + return -EBUSY; /* init still in progress */ + + /* no need for a termination if PXP is not active */ + if (pxp->status != XE_PXP_ACTIVE) { + drm_printf(&p, "PXP not active\n"); + return 0; + } /* simulate a termination interrupt */ spin_lock_irq(&pxp->xe->irq.lock); -- cgit v1.2.3