From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 4075FC369A3 for ; Thu, 3 Apr 2025 12:45:35 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id F301E10E9BD; Thu, 3 Apr 2025 12:45:34 +0000 (UTC) Received: from mblankhorst.nl (lankhorst.se [141.105.120.124]) by gabe.freedesktop.org (Postfix) with ESMTPS id 4B6D010E23C for ; Thu, 3 Apr 2025 12:45:32 +0000 (UTC) From: Maarten Lankhorst To: intel-xe@lists.freedesktop.org Cc: Maarten Lankhorst Subject: [PATCH 5/6] drm/xe/ggtt: Seperate flags and address in PTE encoding Date: Thu, 3 Apr 2025 14:45:19 +0200 Message-ID: <20250403124520.470916-6-dev@lankhorst.se> X-Mailer: git-send-email 2.45.2 In-Reply-To: <20250403124520.470916-1-dev@lankhorst.se> References: <20250403124520.470916-1-dev@lankhorst.se> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-BeenThere: intel-xe@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Intel Xe graphics driver List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: intel-xe-bounces@lists.freedesktop.org Sender: "Intel-xe" Pinning large linear display framebuffers is becoming a bottleneck. My plan of attack is doing a custom walk over the BO, this allows for easier optimization of consecutive entries. Signed-off-by: Maarten Lankhorst --- drivers/gpu/drm/xe/display/xe_fb_pin.c | 37 +++++------- drivers/gpu/drm/xe/xe_ggtt.c | 78 +++++++++++++++++--------- drivers/gpu/drm/xe/xe_ggtt.h | 2 + drivers/gpu/drm/xe/xe_ggtt_types.h | 4 +- 4 files changed, 70 insertions(+), 51 deletions(-) diff --git a/drivers/gpu/drm/xe/display/xe_fb_pin.c b/drivers/gpu/drm/xe/display/xe_fb_pin.c index d509def82b134..f1b938001100a 100644 --- a/drivers/gpu/drm/xe/display/xe_fb_pin.c +++ b/drivers/gpu/drm/xe/display/xe_fb_pin.c @@ -23,6 +23,7 @@ write_dpt_rotated(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs, u32 bo_ struct xe_device *xe = xe_bo_device(bo); struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt; u32 column, row; + u64 pte = ggtt->pt_ops->pte_encode_flags(bo, xe->pat.idx[XE_CACHE_NONE]); /* TODO: Maybe rewrite so we can traverse the bo addresses sequentially, * by writing dpt/ggtt in a different order? @@ -32,10 +33,9 @@ write_dpt_rotated(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs, u32 bo_ u32 src_idx = src_stride * (height - 1) + column + bo_ofs; for (row = 0; row < height; row++) { - u64 pte = ggtt->pt_ops->pte_encode_bo(bo, src_idx * XE_PAGE_SIZE, - xe->pat.idx[XE_CACHE_NONE]); + u64 addr = xe_bo_addr(bo, src_idx * XE_PAGE_SIZE, XE_PAGE_SIZE); - iosys_map_wr(map, *dpt_ofs, u64, pte); + iosys_map_wr(map, *dpt_ofs, u64, pte | addr); *dpt_ofs += 8; src_idx -= src_stride; } @@ -55,17 +55,15 @@ write_dpt_remapped(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs, { struct xe_device *xe = xe_bo_device(bo); struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt; - u64 (*pte_encode_bo)(struct xe_bo *bo, u64 bo_offset, u16 pat_index) - = ggtt->pt_ops->pte_encode_bo; u32 column, row; + u64 pte = ggtt->pt_ops->pte_encode_flags(bo, xe->pat.idx[XE_CACHE_NONE]); for (row = 0; row < height; row++) { u32 src_idx = src_stride * row + bo_ofs; for (column = 0; column < width; column++) { - iosys_map_wr(map, *dpt_ofs, u64, - pte_encode_bo(bo, src_idx * XE_PAGE_SIZE, - xe->pat.idx[XE_CACHE_NONE])); + u64 addr = xe_bo_addr(bo, src_idx * XE_PAGE_SIZE, XE_PAGE_SIZE); + iosys_map_wr(map, *dpt_ofs, u64, pte | addr); *dpt_ofs += 8; src_idx++; @@ -129,13 +127,13 @@ static int __xe_pin_fb_vma_dpt(const struct intel_framebuffer *fb, return PTR_ERR(dpt); if (view->type == I915_GTT_VIEW_NORMAL) { + u64 pte = ggtt->pt_ops->pte_encode_flags(bo, xe->pat.idx[XE_CACHE_NONE]); u32 x; for (x = 0; x < size / XE_PAGE_SIZE; x++) { - u64 pte = ggtt->pt_ops->pte_encode_bo(bo, x * XE_PAGE_SIZE, - xe->pat.idx[XE_CACHE_NONE]); + u64 addr = xe_bo_addr(bo, x * XE_PAGE_SIZE, XE_PAGE_SIZE); - iosys_map_wr(&dpt->vmap, x * 8, u64, pte); + iosys_map_wr(&dpt->vmap, x * 8, u64, pte | addr); } } else if (view->type == I915_GTT_VIEW_REMAPPED) { const struct intel_remapped_info *remap_info = &view->remapped; @@ -173,15 +171,15 @@ write_ggtt_rotated(struct xe_bo *bo, struct xe_ggtt *ggtt, u32 *ggtt_ofs, u32 bo { struct xe_device *xe = xe_bo_device(bo); u32 column, row; + u64 pte = ggtt->pt_ops->pte_encode_flags(bo, xe->pat.idx[XE_CACHE_NONE]); for (column = 0; column < width; column++) { u32 src_idx = src_stride * (height - 1) + column + bo_ofs; for (row = 0; row < height; row++) { - u64 pte = ggtt->pt_ops->pte_encode_bo(bo, src_idx * XE_PAGE_SIZE, - xe->pat.idx[XE_CACHE_NONE]); + u64 addr = xe_bo_addr(bo, src_idx * XE_PAGE_SIZE, XE_PAGE_SIZE); - ggtt->pt_ops->ggtt_set_pte(ggtt, *ggtt_ofs, pte); + ggtt->pt_ops->ggtt_set_pte(ggtt, *ggtt_ofs, pte | addr); *ggtt_ofs += XE_PAGE_SIZE; src_idx -= src_stride; } @@ -219,26 +217,19 @@ static int __xe_pin_fb_vma_ggtt(const struct intel_framebuffer *fb, if (bo->ggtt_node[tile0->id] && view->type == I915_GTT_VIEW_NORMAL) { vma->node = bo->ggtt_node[tile0->id]; } else if (view->type == I915_GTT_VIEW_NORMAL) { - u32 x, size = bo->ttm.base.size; - vma->node = xe_ggtt_node_init(ggtt); if (IS_ERR(vma->node)) { ret = PTR_ERR(vma->node); goto out_unlock; } - ret = xe_ggtt_node_insert_locked(vma->node, size, align, 0); + ret = xe_ggtt_node_insert_locked(vma->node, bo->size, align, 0); if (ret) { xe_ggtt_node_fini(vma->node); goto out_unlock; } - for (x = 0; x < size; x += XE_PAGE_SIZE) { - u64 pte = ggtt->pt_ops->pte_encode_bo(bo, x, - xe->pat.idx[XE_CACHE_NONE]); - - ggtt->pt_ops->ggtt_set_pte(ggtt, vma->node->base.start + x, pte); - } + xe_ggtt_map_bo(ggtt, vma->node, bo, xe->pat.idx[XE_CACHE_NONE]); } else { u32 i, ggtt_ofs; const struct intel_rotation_info *rot_info = &view->rotated; diff --git a/drivers/gpu/drm/xe/xe_ggtt.c b/drivers/gpu/drm/xe/xe_ggtt.c index b197b835ed442..e37dc2149f9a3 100644 --- a/drivers/gpu/drm/xe/xe_ggtt.c +++ b/drivers/gpu/drm/xe/xe_ggtt.c @@ -27,6 +27,7 @@ #include "xe_map.h" #include "xe_mmio.h" #include "xe_pm.h" +#include "xe_res_cursor.h" #include "xe_sriov.h" #include "xe_wa.h" #include "xe_wopcm.h" @@ -64,13 +65,9 @@ * give us the correct placement for free. */ -static u64 xelp_ggtt_pte_encode_bo(struct xe_bo *bo, u64 bo_offset, - u16 pat_index) +static u64 xelp_ggtt_pte_flags(struct xe_bo *bo, u16 pat_index) { - u64 pte; - - pte = xe_bo_addr(bo, bo_offset, XE_PAGE_SIZE); - pte |= XE_PAGE_PRESENT; + u64 pte = XE_PAGE_PRESENT; if (xe_bo_is_vram(bo) || xe_bo_is_stolen_devmem(bo)) pte |= XE_GGTT_PTE_DM; @@ -78,13 +75,12 @@ static u64 xelp_ggtt_pte_encode_bo(struct xe_bo *bo, u64 bo_offset, return pte; } -static u64 xelpg_ggtt_pte_encode_bo(struct xe_bo *bo, u64 bo_offset, - u16 pat_index) +static u64 xelpg_ggtt_pte_flags(struct xe_bo *bo, u16 pat_index) { struct xe_device *xe = xe_bo_device(bo); u64 pte; - pte = xelp_ggtt_pte_encode_bo(bo, bo_offset, pat_index); + pte = xelp_ggtt_pte_flags(bo, pat_index); xe_assert(xe, pat_index <= 3); @@ -149,8 +145,9 @@ static void xe_ggtt_clear(struct xe_ggtt *ggtt, u64 start, u64 size) xe_tile_assert(ggtt->tile, start < end); if (ggtt->scratch) - scratch_pte = ggtt->pt_ops->pte_encode_bo(ggtt->scratch, 0, - pat_index); + scratch_pte = xe_bo_addr(ggtt->scratch, 0, XE_PAGE_SIZE) | + ggtt->pt_ops->pte_encode_flags(ggtt->scratch, + pat_index); else scratch_pte = 0; @@ -210,17 +207,17 @@ static void primelockdep(struct xe_ggtt *ggtt) } static const struct xe_ggtt_pt_ops xelp_pt_ops = { - .pte_encode_bo = xelp_ggtt_pte_encode_bo, + .pte_encode_flags = xelp_ggtt_pte_flags, .ggtt_set_pte = xe_ggtt_set_pte, }; static const struct xe_ggtt_pt_ops xelpg_pt_ops = { - .pte_encode_bo = xelpg_ggtt_pte_encode_bo, + .pte_encode_flags = xelpg_ggtt_pte_flags, .ggtt_set_pte = xe_ggtt_set_pte, }; static const struct xe_ggtt_pt_ops xelpg_pt_wa_ops = { - .pte_encode_bo = xelpg_ggtt_pte_encode_bo, + .pte_encode_flags = xelpg_ggtt_pte_flags, .ggtt_set_pte = xe_ggtt_set_pte_and_flush, }; @@ -612,23 +609,46 @@ bool xe_ggtt_node_allocated(const struct xe_ggtt_node *node) /** * xe_ggtt_map_bo - Map the BO into GGTT * @ggtt: the &xe_ggtt where node will be mapped + * @node: the &xe_ggtt_node where this BO is mapped * @bo: the &xe_bo to be mapped + * @pat_index: Which pat_index to use. */ -static void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_bo *bo) +void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_ggtt_node *node, + struct xe_bo *bo, u16 pat_index) { - u16 cache_mode = bo->flags & XE_BO_FLAG_NEEDS_UC ? XE_CACHE_NONE : XE_CACHE_WB; - u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[cache_mode]; - u64 start; - u64 offset, pte; - if (XE_WARN_ON(!bo->ggtt_node[ggtt->tile->id])) + u64 start, pte; + struct xe_res_cursor cur; + + if (XE_WARN_ON(!node)) return; - start = bo->ggtt_node[ggtt->tile->id]->base.start; + start = node->base.start; + + pte = ggtt->pt_ops->pte_encode_flags(bo, pat_index); + if (!xe_bo_is_vram(bo) && !xe_bo_is_stolen(bo)) { + xe_assert(xe_bo_device(bo), bo->ttm.ttm); - for (offset = 0; offset < bo->size; offset += XE_PAGE_SIZE) { - pte = ggtt->pt_ops->pte_encode_bo(bo, offset, pat_index); - ggtt->pt_ops->ggtt_set_pte(ggtt, start + offset, pte); + for (xe_res_first_sg(xe_bo_sg(bo), 0, bo->size, &cur); + cur.remaining; + xe_res_next(&cur, XE_PAGE_SIZE)) { + u64 addr = xe_res_dma(&cur); + + ggtt->pt_ops->ggtt_set_pte(ggtt, start + cur.start, addr | pte); + } + } else { + u64 end = start + bo->size; + + /* Prepend GPU offset */ + pte |= vram_region_gpu_offset(bo->ttm.resource); + + for (xe_res_first(bo->ttm.resource, 0, bo->size, &cur); + cur.remaining; + xe_res_next(&cur, XE_PAGE_SIZE)) { + + ggtt->pt_ops->ggtt_set_pte(ggtt, end - cur.remaining, + pte + cur.start); + } } } @@ -641,8 +661,11 @@ static void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_bo *bo) */ void xe_ggtt_map_bo_unlocked(struct xe_ggtt *ggtt, struct xe_bo *bo) { + u16 cache_mode = bo->flags & XE_BO_FLAG_NEEDS_UC ? XE_CACHE_NONE : XE_CACHE_WB; + u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[cache_mode]; + mutex_lock(&ggtt->lock); - xe_ggtt_map_bo(ggtt, bo); + xe_ggtt_map_bo(ggtt, bo->ggtt_node[ggtt->tile->id], bo, pat_index); mutex_unlock(&ggtt->lock); } @@ -682,7 +705,10 @@ static int __xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo, xe_ggtt_node_fini(bo->ggtt_node[tile_id]); bo->ggtt_node[tile_id] = NULL; } else { - xe_ggtt_map_bo(ggtt, bo); + u16 cache_mode = bo->flags & XE_BO_FLAG_NEEDS_UC ? XE_CACHE_NONE : XE_CACHE_WB; + u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[cache_mode]; + + xe_ggtt_map_bo(ggtt, bo->ggtt_node[tile_id], bo, pat_index); } mutex_unlock(&ggtt->lock); diff --git a/drivers/gpu/drm/xe/xe_ggtt.h b/drivers/gpu/drm/xe/xe_ggtt.h index 0bab1fd7cc817..c48da99908848 100644 --- a/drivers/gpu/drm/xe/xe_ggtt.h +++ b/drivers/gpu/drm/xe/xe_ggtt.h @@ -26,6 +26,8 @@ int xe_ggtt_node_insert_locked(struct xe_ggtt_node *node, u32 size, u32 align, u32 mm_flags); void xe_ggtt_node_remove(struct xe_ggtt_node *node, bool invalidate); bool xe_ggtt_node_allocated(const struct xe_ggtt_node *node); +void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_ggtt_node *node, + struct xe_bo *bo, u16 pat_index); void xe_ggtt_map_bo_unlocked(struct xe_ggtt *ggtt, struct xe_bo *bo); int xe_ggtt_insert_bo(struct xe_ggtt *ggtt, struct xe_bo *bo); int xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo, diff --git a/drivers/gpu/drm/xe/xe_ggtt_types.h b/drivers/gpu/drm/xe/xe_ggtt_types.h index cb02b7994a9ac..c5e999d58ff2a 100644 --- a/drivers/gpu/drm/xe/xe_ggtt_types.h +++ b/drivers/gpu/drm/xe/xe_ggtt_types.h @@ -74,8 +74,8 @@ struct xe_ggtt_node { * Which can vary from platform to platform. */ struct xe_ggtt_pt_ops { - /** @pte_encode_bo: Encode PTE address for a given BO */ - u64 (*pte_encode_bo)(struct xe_bo *bo, u64 bo_offset, u16 pat_index); + /** @pte_encode_flags: Encode PTE flags for a given BO */ + u64 (*pte_encode_flags)(struct xe_bo *bo, u16 pat_index); /** @ggtt_set_pte: Directly write into GGTT's PTE */ void (*ggtt_set_pte)(struct xe_ggtt *ggtt, u64 addr, u64 pte); }; -- 2.45.2