diff options
Diffstat (limited to 'drivers/gpu/drm/xe/xe_ggtt.c')
| -rw-r--r-- | drivers/gpu/drm/xe/xe_ggtt.c | 947 |
1 files changed, 777 insertions, 170 deletions
diff --git a/drivers/gpu/drm/xe/xe_ggtt.c b/drivers/gpu/drm/xe/xe_ggtt.c index 0d541f55b4fc..ef481b334af4 100644 --- a/drivers/gpu/drm/xe/xe_ggtt.c +++ b/drivers/gpu/drm/xe/xe_ggtt.c @@ -5,11 +5,15 @@ #include "xe_ggtt.h" +#include <kunit/visibility.h> +#include <linux/fault-inject.h> #include <linux/io-64-nonatomic-lo-hi.h> #include <linux/sizes.h> +#include <drm/drm_drv.h> #include <drm/drm_managed.h> -#include <drm/i915_drm.h> +#include <drm/intel/i915_drm.h> +#include <generated/xe_wa_oob.h> #include "regs/xe_gt_regs.h" #include "regs/xe_gtt_defs.h" @@ -19,19 +23,53 @@ #include "xe_device.h" #include "xe_gt.h" #include "xe_gt_printk.h" -#include "xe_gt_tlb_invalidation.h" #include "xe_map.h" +#include "xe_mmio.h" #include "xe_pm.h" +#include "xe_res_cursor.h" #include "xe_sriov.h" +#include "xe_tile_printk.h" +#include "xe_tile_sriov_vf.h" +#include "xe_tlb_inval.h" +#include "xe_wa.h" #include "xe_wopcm.h" -static u64 xelp_ggtt_pte_encode_bo(struct xe_bo *bo, u64 bo_offset, - u16 pat_index) -{ - u64 pte; +/** + * DOC: Global Graphics Translation Table (GGTT) + * + * Xe GGTT implements the support for a Global Virtual Address space that is used + * for resources that are accessible to privileged (i.e. kernel-mode) processes, + * and not tied to a specific user-level process. For example, the Graphics + * micro-Controller (GuC) and Display Engine (if present) utilize this Global + * address space. + * + * The Global GTT (GGTT) translates from the Global virtual address to a physical + * address that can be accessed by HW. The GGTT is a flat, single-level table. + * + * Xe implements a simplified version of the GGTT specifically managing only a + * certain range of it that goes from the Write Once Protected Content Memory (WOPCM) + * Layout to a predefined GUC_GGTT_TOP. This approach avoids complications related to + * the GuC (Graphics Microcontroller) hardware limitations. The GuC address space + * is limited on both ends of the GGTT, because the GuC shim HW redirects + * accesses to those addresses to other HW areas instead of going through the + * GGTT. On the bottom end, the GuC can't access offsets below the WOPCM size, + * while on the top side the limit is fixed at GUC_GGTT_TOP. To keep things + * simple, instead of checking each object to see if they are accessed by GuC or + * not, we just exclude those areas from the allocator. Additionally, to simplify + * the driver load, we use the maximum WOPCM size in this logic instead of the + * programmed one, so we don't need to wait until the actual size to be + * programmed is determined (which requires FW fetch) before initializing the + * GGTT. These simplifications might waste space in the GGTT (about 20-25 MBs + * depending on the platform) but we can live with this. Another benefit of this + * is the GuC bootrom can't access anything below the WOPCM max size so anything + * the bootrom needs to access (e.g. a RSA key) needs to be placed in the GGTT + * above the WOPCM max size. Starting the GGTT allocations above the WOPCM max + * give us the correct placement for free. + */ - pte = xe_bo_addr(bo, bo_offset, XE_PAGE_SIZE); - pte |= XE_PAGE_PRESENT; +static u64 xelp_ggtt_pte_flags(struct xe_bo *bo, u16 pat_index) +{ + u64 pte = XE_PAGE_PRESENT; if (xe_bo_is_vram(bo) || xe_bo_is_stolen_devmem(bo)) pte |= XE_GGTT_PTE_DM; @@ -39,13 +77,12 @@ 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); @@ -67,7 +104,40 @@ static unsigned int probe_gsm_size(struct pci_dev *pdev) return ggms ? SZ_1M << ggms : 0; } -void xe_ggtt_set_pte(struct xe_ggtt *ggtt, u64 addr, u64 pte) +static void ggtt_update_access_counter(struct xe_ggtt *ggtt) +{ + struct xe_tile *tile = ggtt->tile; + struct xe_gt *affected_gt; + u32 max_gtt_writes; + + if (tile->primary_gt && XE_GT_WA(tile->primary_gt, 22019338487)) { + affected_gt = tile->primary_gt; + max_gtt_writes = 1100; + + /* Only expected to apply to primary GT on dgpu platforms */ + xe_tile_assert(tile, IS_DGFX(tile_to_xe(tile))); + } else { + affected_gt = tile->media_gt; + max_gtt_writes = 63; + + /* Only expected to apply to media GT on igpu platforms */ + xe_tile_assert(tile, !IS_DGFX(tile_to_xe(tile))); + } + + /* + * Wa_22019338487: GMD_ID is a RO register, a dummy write forces gunit + * to wait for completion of prior GTT writes before letting this through. + * This needs to be done for all GGTT writes originating from the CPU. + */ + lockdep_assert_held(&ggtt->lock); + + if ((++ggtt->access_count % max_gtt_writes) == 0) { + xe_mmio_write32(&affected_gt->mmio, GMD_ID, 0x0); + ggtt->access_count = 0; + } +} + +static void xe_ggtt_set_pte(struct xe_ggtt *ggtt, u64 addr, u64 pte) { xe_tile_assert(ggtt->tile, !(addr & XE_PTE_MASK)); xe_tile_assert(ggtt->tile, addr < ggtt->size); @@ -75,6 +145,20 @@ void xe_ggtt_set_pte(struct xe_ggtt *ggtt, u64 addr, u64 pte) writeq(pte, &ggtt->gsm[addr >> XE_PTE_SHIFT]); } +static void xe_ggtt_set_pte_and_flush(struct xe_ggtt *ggtt, u64 addr, u64 pte) +{ + xe_ggtt_set_pte(ggtt, addr, pte); + ggtt_update_access_counter(ggtt); +} + +static u64 xe_ggtt_get_pte(struct xe_ggtt *ggtt, u64 addr) +{ + xe_tile_assert(ggtt->tile, !(addr & XE_PTE_MASK)); + xe_tile_assert(ggtt->tile, addr < ggtt->size); + + return readq(&ggtt->gsm[addr >> XE_PTE_SHIFT]); +} + static void xe_ggtt_clear(struct xe_ggtt *ggtt, u64 start, u64 size) { u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[XE_CACHE_WB]; @@ -84,70 +168,140 @@ 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; while (start < end) { - xe_ggtt_set_pte(ggtt, start, scratch_pte); + ggtt->pt_ops->ggtt_set_pte(ggtt, start, scratch_pte); start += XE_PAGE_SIZE; } } +static void primelockdep(struct xe_ggtt *ggtt) +{ + if (!IS_ENABLED(CONFIG_LOCKDEP)) + return; + + fs_reclaim_acquire(GFP_KERNEL); + might_lock(&ggtt->lock); + fs_reclaim_release(GFP_KERNEL); +} + +/** + * xe_ggtt_alloc - Allocate a GGTT for a given &xe_tile + * @tile: &xe_tile + * + * Allocates a &xe_ggtt for a given tile. + * + * Return: &xe_ggtt on success, or NULL when out of memory. + */ +struct xe_ggtt *xe_ggtt_alloc(struct xe_tile *tile) +{ + struct xe_device *xe = tile_to_xe(tile); + struct xe_ggtt *ggtt; + + ggtt = drmm_kzalloc(&xe->drm, sizeof(*ggtt), GFP_KERNEL); + if (!ggtt) + return NULL; + + if (drmm_mutex_init(&xe->drm, &ggtt->lock)) + return NULL; + + primelockdep(ggtt); + ggtt->tile = tile; + + return ggtt; +} + static void ggtt_fini_early(struct drm_device *drm, void *arg) { struct xe_ggtt *ggtt = arg; - mutex_destroy(&ggtt->lock); + destroy_workqueue(ggtt->wq); drm_mm_takedown(&ggtt->mm); } -static void ggtt_fini(struct drm_device *drm, void *arg) +static void ggtt_fini(void *arg) { struct xe_ggtt *ggtt = arg; ggtt->scratch = NULL; } -static void primelockdep(struct xe_ggtt *ggtt) +#ifdef CONFIG_LOCKDEP +void xe_ggtt_might_lock(struct xe_ggtt *ggtt) { - if (!IS_ENABLED(CONFIG_LOCKDEP)) - return; - - fs_reclaim_acquire(GFP_KERNEL); might_lock(&ggtt->lock); - fs_reclaim_release(GFP_KERNEL); } +#endif 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, + .ggtt_get_pte = xe_ggtt_get_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, + .ggtt_get_pte = xe_ggtt_get_pte, }; -/* - * Early GGTT initialization, which allows to create new mappings usable by the - * GuC. - * Mappings are not usable by the HW engines, as it doesn't have scratch / +static const struct xe_ggtt_pt_ops xelpg_pt_wa_ops = { + .pte_encode_flags = xelpg_ggtt_pte_flags, + .ggtt_set_pte = xe_ggtt_set_pte_and_flush, + .ggtt_get_pte = xe_ggtt_get_pte, +}; + +static void __xe_ggtt_init_early(struct xe_ggtt *ggtt, u32 reserved) +{ + drm_mm_init(&ggtt->mm, reserved, + ggtt->size - reserved); +} + +int xe_ggtt_init_kunit(struct xe_ggtt *ggtt, u32 reserved, u32 size) +{ + ggtt->size = size; + __xe_ggtt_init_early(ggtt, reserved); + return 0; +} +EXPORT_SYMBOL_IF_KUNIT(xe_ggtt_init_kunit); + +static void dev_fini_ggtt(void *arg) +{ + struct xe_ggtt *ggtt = arg; + + drain_workqueue(ggtt->wq); +} + +/** + * xe_ggtt_init_early - Early GGTT initialization + * @ggtt: the &xe_ggtt to be initialized + * + * It allows to create new mappings usable by the GuC. + * Mappings are not usable by the HW engines, as it doesn't have scratch nor * initial clear done to it yet. That will happen in the regular, non-early - * GGTT init. + * GGTT initialization. + * + * Return: 0 on success or a negative error code on failure. */ int xe_ggtt_init_early(struct xe_ggtt *ggtt) { struct xe_device *xe = tile_to_xe(ggtt->tile); struct pci_dev *pdev = to_pci_dev(xe->drm.dev); unsigned int gsm_size; + int err; - if (IS_SRIOV_VF(xe)) + if (IS_SRIOV_VF(xe) || GRAPHICS_VERx100(xe) >= 1250) gsm_size = SZ_8M; /* GGTT is expected to be 4GiB */ else gsm_size = probe_gsm_size(pdev); if (gsm_size == 0) { - drm_err(&xe->drm, "Hardware reported no preallocated GSM\n"); + xe_tile_err(ggtt->tile, "Hardware reported no preallocated GSM\n"); return -ENOMEM; } @@ -157,44 +311,40 @@ int xe_ggtt_init_early(struct xe_ggtt *ggtt) if (IS_DGFX(xe) && xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K) ggtt->flags |= XE_GGTT_FLAGS_64K; - /* - * 8B per entry, each points to a 4KB page. - * - * The GuC address space is limited on both ends of the GGTT, because - * the GuC shim HW redirects accesses to those addresses to other HW - * areas instead of going through the GGTT. On the bottom end, the GuC - * can't access offsets below the WOPCM size, while on the top side the - * limit is fixed at GUC_GGTT_TOP. To keep things simple, instead of - * checking each object to see if they are accessed by GuC or not, we - * just exclude those areas from the allocator. Additionally, to - * simplify the driver load, we use the maximum WOPCM size in this logic - * instead of the programmed one, so we don't need to wait until the - * actual size to be programmed is determined (which requires FW fetch) - * before initializing the GGTT. These simplifications might waste space - * in the GGTT (about 20-25 MBs depending on the platform) but we can - * live with this. - * - * Another benifit of this is the GuC bootrom can't access anything - * below the WOPCM max size so anything the bootom needs to access (e.g. - * a RSA key) needs to be placed in the GGTT above the WOPCM max size. - * Starting the GGTT allocations above the WOPCM max give us the correct - * placement for free. - */ if (ggtt->size > GUC_GGTT_TOP) ggtt->size = GUC_GGTT_TOP; if (GRAPHICS_VERx100(xe) >= 1270) - ggtt->pt_ops = &xelpg_pt_ops; + ggtt->pt_ops = + (ggtt->tile->media_gt && XE_GT_WA(ggtt->tile->media_gt, 22019338487)) || + (ggtt->tile->primary_gt && XE_GT_WA(ggtt->tile->primary_gt, 22019338487)) ? + &xelpg_pt_wa_ops : &xelpg_pt_ops; else ggtt->pt_ops = &xelp_pt_ops; - drm_mm_init(&ggtt->mm, xe_wopcm_size(xe), - ggtt->size - xe_wopcm_size(xe)); - mutex_init(&ggtt->lock); - primelockdep(ggtt); + ggtt->wq = alloc_workqueue("xe-ggtt-wq", 0, WQ_MEM_RECLAIM); + if (!ggtt->wq) + return -ENOMEM; + + __xe_ggtt_init_early(ggtt, xe_wopcm_size(xe)); + + err = drmm_add_action_or_reset(&xe->drm, ggtt_fini_early, ggtt); + if (err) + return err; - return drmm_add_action_or_reset(&xe->drm, ggtt_fini_early, ggtt); + err = devm_add_action_or_reset(xe->drm.dev, dev_fini_ggtt, ggtt); + if (err) + return err; + + if (IS_SRIOV_VF(xe)) { + err = xe_tile_sriov_vf_prepare_ggtt(ggtt->tile); + if (err) + return err; + } + + return 0; } +ALLOW_ERROR_INJECTION(xe_ggtt_init_early, ERRNO); /* See xe_pci_probe() */ static void xe_ggtt_invalidate(struct xe_ggtt *ggtt); @@ -212,6 +362,77 @@ static void xe_ggtt_initial_clear(struct xe_ggtt *ggtt) mutex_unlock(&ggtt->lock); } +static void ggtt_node_remove(struct xe_ggtt_node *node) +{ + struct xe_ggtt *ggtt = node->ggtt; + struct xe_device *xe = tile_to_xe(ggtt->tile); + bool bound; + int idx; + + bound = drm_dev_enter(&xe->drm, &idx); + + mutex_lock(&ggtt->lock); + if (bound) + xe_ggtt_clear(ggtt, node->base.start, node->base.size); + drm_mm_remove_node(&node->base); + node->base.size = 0; + mutex_unlock(&ggtt->lock); + + if (!bound) + goto free_node; + + if (node->invalidate_on_remove) + xe_ggtt_invalidate(ggtt); + + drm_dev_exit(idx); + +free_node: + xe_ggtt_node_fini(node); +} + +static void ggtt_node_remove_work_func(struct work_struct *work) +{ + struct xe_ggtt_node *node = container_of(work, typeof(*node), + delayed_removal_work); + struct xe_device *xe = tile_to_xe(node->ggtt->tile); + + xe_pm_runtime_get(xe); + ggtt_node_remove(node); + xe_pm_runtime_put(xe); +} + +/** + * xe_ggtt_node_remove - Remove a &xe_ggtt_node from the GGTT + * @node: the &xe_ggtt_node to be removed + * @invalidate: if node needs invalidation upon removal + */ +void xe_ggtt_node_remove(struct xe_ggtt_node *node, bool invalidate) +{ + struct xe_ggtt *ggtt; + struct xe_device *xe; + + if (!node || !node->ggtt) + return; + + ggtt = node->ggtt; + xe = tile_to_xe(ggtt->tile); + + node->invalidate_on_remove = invalidate; + + if (xe_pm_runtime_get_if_active(xe)) { + ggtt_node_remove(node); + xe_pm_runtime_put(xe); + } else { + queue_work(ggtt->wq, &node->delayed_removal_work); + } +} + +/** + * xe_ggtt_init - Regular non-early GGTT initialization + * @ggtt: the &xe_ggtt to be initialized + * + * Return: 0 on success or a negative error code on failure. + */ int xe_ggtt_init(struct xe_ggtt *ggtt) { struct xe_device *xe = tile_to_xe(ggtt->tile); @@ -220,10 +441,10 @@ int xe_ggtt_init(struct xe_ggtt *ggtt) /* * So we don't need to worry about 64K GGTT layout when dealing with - * scratch entires, rather keep the scratch page in system memory on + * scratch entries, rather keep the scratch page in system memory on * platforms where 64K pages are needed for VRAM. */ - flags = XE_BO_FLAG_PINNED; + flags = 0; if (ggtt->flags & XE_GGTT_FLAGS_64K) flags |= XE_BO_FLAG_SYSTEM; else @@ -235,11 +456,11 @@ int xe_ggtt_init(struct xe_ggtt *ggtt) goto err; } - xe_map_memset(xe, &ggtt->scratch->vmap, 0, 0, ggtt->scratch->size); + xe_map_memset(xe, &ggtt->scratch->vmap, 0, 0, xe_bo_size(ggtt->scratch)); xe_ggtt_initial_clear(ggtt); - return drmm_add_action_or_reset(&xe->drm, ggtt_fini, ggtt); + return devm_add_action_or_reset(xe->drm.dev, ggtt_fini, ggtt); err: ggtt->scratch = NULL; return err; @@ -252,38 +473,27 @@ static void ggtt_invalidate_gt_tlb(struct xe_gt *gt) if (!gt) return; - err = xe_gt_tlb_invalidation_ggtt(gt); - if (err) - drm_warn(>_to_xe(gt)->drm, "xe_gt_tlb_invalidation_ggtt error=%d", err); + err = xe_tlb_inval_ggtt(>->tlb_inval); + xe_gt_WARN(gt, err, "Failed to invalidate GGTT (%pe)", ERR_PTR(err)); } static void xe_ggtt_invalidate(struct xe_ggtt *ggtt) { + struct xe_device *xe = tile_to_xe(ggtt->tile); + + /* + * XXX: Barrier for GGTT pages. Unsure exactly why this required but + * without this LNL is having issues with the GuC reading scratch page + * vs. correct GGTT page. Not particularly a hot code path so blindly + * do a mmio read here which results in GuC reading correct GGTT page. + */ + xe_mmio_read32(xe_root_tile_mmio(xe), VF_CAP_REG); + /* Each GT in a tile has its own TLB to cache GGTT lookups */ ggtt_invalidate_gt_tlb(ggtt->tile->primary_gt); ggtt_invalidate_gt_tlb(ggtt->tile->media_gt); } -void xe_ggtt_printk(struct xe_ggtt *ggtt, const char *prefix) -{ - u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[XE_CACHE_WB]; - u64 addr, scratch_pte; - - scratch_pte = ggtt->pt_ops->pte_encode_bo(ggtt->scratch, 0, pat_index); - - printk("%sGlobal GTT:", prefix); - for (addr = 0; addr < ggtt->size; addr += XE_PAGE_SIZE) { - unsigned int i = addr / XE_PAGE_SIZE; - - xe_tile_assert(ggtt->tile, addr <= U32_MAX); - if (ggtt->gsm[i] == scratch_pte) - continue; - - printk("%s ggtt[0x%08x] = 0x%016llx", - prefix, (u32)addr, ggtt->gsm[i]); - } -} - static void xe_ggtt_dump_node(struct xe_ggtt *ggtt, const struct drm_mm_node *node, const char *description) { @@ -291,172 +501,417 @@ static void xe_ggtt_dump_node(struct xe_ggtt *ggtt, if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) { string_get_size(node->size, 1, STRING_UNITS_2, buf, sizeof(buf)); - xe_gt_dbg(ggtt->tile->primary_gt, "GGTT %#llx-%#llx (%s) %s\n", - node->start, node->start + node->size, buf, description); + xe_tile_dbg(ggtt->tile, "GGTT %#llx-%#llx (%s) %s\n", + node->start, node->start + node->size, buf, description); } } /** - * xe_ggtt_balloon - prevent allocation of specified GGTT addresses - * @ggtt: the &xe_ggtt where we want to make reservation + * xe_ggtt_node_insert_balloon_locked - prevent allocation of specified GGTT addresses + * @node: the &xe_ggtt_node to hold reserved GGTT node * @start: the starting GGTT address of the reserved region * @end: then end GGTT address of the reserved region - * @node: the &drm_mm_node to hold reserved GGTT node * - * Use xe_ggtt_deballoon() to release a reserved GGTT node. + * To be used in cases where ggtt->lock is already taken. + * Use xe_ggtt_node_remove_balloon_locked() to release a reserved GGTT node. * * Return: 0 on success or a negative error code on failure. */ -int xe_ggtt_balloon(struct xe_ggtt *ggtt, u64 start, u64 end, struct drm_mm_node *node) +int xe_ggtt_node_insert_balloon_locked(struct xe_ggtt_node *node, u64 start, u64 end) { + struct xe_ggtt *ggtt = node->ggtt; int err; xe_tile_assert(ggtt->tile, start < end); xe_tile_assert(ggtt->tile, IS_ALIGNED(start, XE_PAGE_SIZE)); xe_tile_assert(ggtt->tile, IS_ALIGNED(end, XE_PAGE_SIZE)); - xe_tile_assert(ggtt->tile, !drm_mm_node_allocated(node)); + xe_tile_assert(ggtt->tile, !drm_mm_node_allocated(&node->base)); + lockdep_assert_held(&ggtt->lock); - node->color = 0; - node->start = start; - node->size = end - start; + node->base.color = 0; + node->base.start = start; + node->base.size = end - start; - mutex_lock(&ggtt->lock); - err = drm_mm_reserve_node(&ggtt->mm, node); - mutex_unlock(&ggtt->lock); + err = drm_mm_reserve_node(&ggtt->mm, &node->base); - if (xe_gt_WARN(ggtt->tile->primary_gt, err, - "Failed to balloon GGTT %#llx-%#llx (%pe)\n", - node->start, node->start + node->size, ERR_PTR(err))) + if (xe_tile_WARN(ggtt->tile, err, "Failed to balloon GGTT %#llx-%#llx (%pe)\n", + node->base.start, node->base.start + node->base.size, ERR_PTR(err))) return err; - xe_ggtt_dump_node(ggtt, node, "balloon"); + xe_ggtt_dump_node(ggtt, &node->base, "balloon"); return 0; } /** - * xe_ggtt_deballoon - release a reserved GGTT region - * @ggtt: the &xe_ggtt where reserved node belongs - * @node: the &drm_mm_node with reserved GGTT region + * xe_ggtt_node_remove_balloon_locked - release a reserved GGTT region + * @node: the &xe_ggtt_node with reserved GGTT region * - * See xe_ggtt_balloon() for details. + * To be used in cases where ggtt->lock is already taken. + * See xe_ggtt_node_insert_balloon_locked() for details. */ -void xe_ggtt_deballoon(struct xe_ggtt *ggtt, struct drm_mm_node *node) +void xe_ggtt_node_remove_balloon_locked(struct xe_ggtt_node *node) { - if (!drm_mm_node_allocated(node)) + if (!xe_ggtt_node_allocated(node)) return; - xe_ggtt_dump_node(ggtt, node, "deballoon"); + lockdep_assert_held(&node->ggtt->lock); - mutex_lock(&ggtt->lock); - drm_mm_remove_node(node); - mutex_unlock(&ggtt->lock); + xe_ggtt_dump_node(node->ggtt, &node->base, "remove-balloon"); + + drm_mm_remove_node(&node->base); } -int xe_ggtt_insert_special_node_locked(struct xe_ggtt *ggtt, struct drm_mm_node *node, - u32 size, u32 align, u32 mm_flags) +static void xe_ggtt_assert_fit(struct xe_ggtt *ggtt, u64 start, u64 size) { - return drm_mm_insert_node_generic(&ggtt->mm, node, size, align, 0, + struct xe_tile *tile = ggtt->tile; + struct xe_device *xe = tile_to_xe(tile); + u64 __maybe_unused wopcm = xe_wopcm_size(xe); + + xe_tile_assert(tile, start >= wopcm); + xe_tile_assert(tile, start + size < ggtt->size - wopcm); +} + +/** + * xe_ggtt_shift_nodes_locked - Shift GGTT nodes to adjust for a change in usable address range. + * @ggtt: the &xe_ggtt struct instance + * @shift: change to the location of area provisioned for current VF + * + * This function moves all nodes from the GGTT VM, to a temp list. These nodes are expected + * to represent allocations in range formerly assigned to current VF, before the range changed. + * When the GGTT VM is completely clear of any nodes, they are re-added with shifted offsets. + * + * The function has no ability of failing - because it shifts existing nodes, without + * any additional processing. If the nodes were successfully existing at the old address, + * they will do the same at the new one. A fail inside this function would indicate that + * the list of nodes was either already damaged, or that the shift brings the address range + * outside of valid bounds. Both cases justify an assert rather than error code. + */ +void xe_ggtt_shift_nodes_locked(struct xe_ggtt *ggtt, s64 shift) +{ + struct xe_tile *tile __maybe_unused = ggtt->tile; + struct drm_mm_node *node, *tmpn; + LIST_HEAD(temp_list_head); + + lockdep_assert_held(&ggtt->lock); + + if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) + drm_mm_for_each_node_safe(node, tmpn, &ggtt->mm) + xe_ggtt_assert_fit(ggtt, node->start + shift, node->size); + + drm_mm_for_each_node_safe(node, tmpn, &ggtt->mm) { + drm_mm_remove_node(node); + list_add(&node->node_list, &temp_list_head); + } + + list_for_each_entry_safe(node, tmpn, &temp_list_head, node_list) { + list_del(&node->node_list); + node->start += shift; + drm_mm_reserve_node(&ggtt->mm, node); + xe_tile_assert(tile, drm_mm_node_allocated(node)); + } +} + +/** + * xe_ggtt_node_insert_locked - Locked version to insert a &xe_ggtt_node into the GGTT + * @node: the &xe_ggtt_node to be inserted + * @size: size of the node + * @align: alignment constrain of the node + * @mm_flags: flags to control the node behavior + * + * It cannot be called without first having called xe_ggtt_init() once. + * To be used in cases where ggtt->lock is already taken. + * + * Return: 0 on success or a negative error code on failure. + */ +int xe_ggtt_node_insert_locked(struct xe_ggtt_node *node, + u32 size, u32 align, u32 mm_flags) +{ + return drm_mm_insert_node_generic(&node->ggtt->mm, &node->base, size, align, 0, mm_flags); } -int xe_ggtt_insert_special_node(struct xe_ggtt *ggtt, struct drm_mm_node *node, - u32 size, u32 align) +/** + * xe_ggtt_node_insert - Insert a &xe_ggtt_node into the GGTT + * @node: the &xe_ggtt_node to be inserted + * @size: size of the node + * @align: alignment constrain of the node + * + * It cannot be called without first having called xe_ggtt_init() once. + * + * Return: 0 on success or a negative error code on failure. + */ +int xe_ggtt_node_insert(struct xe_ggtt_node *node, u32 size, u32 align) { int ret; - mutex_lock(&ggtt->lock); - ret = xe_ggtt_insert_special_node_locked(ggtt, node, size, - align, DRM_MM_INSERT_HIGH); - mutex_unlock(&ggtt->lock); + if (!node || !node->ggtt) + return -ENOENT; + + mutex_lock(&node->ggtt->lock); + ret = xe_ggtt_node_insert_locked(node, size, align, + DRM_MM_INSERT_HIGH); + mutex_unlock(&node->ggtt->lock); return ret; } -void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_bo *bo) +/** + * xe_ggtt_node_init - Initialize %xe_ggtt_node struct + * @ggtt: the &xe_ggtt where the new node will later be inserted/reserved. + * + * This function will allocate the struct %xe_ggtt_node and return its pointer. + * This struct will then be freed after the node removal upon xe_ggtt_node_remove() + * or xe_ggtt_node_remove_balloon_locked(). + * Having %xe_ggtt_node struct allocated doesn't mean that the node is already allocated + * in GGTT. Only the xe_ggtt_node_insert(), xe_ggtt_node_insert_locked(), + * xe_ggtt_node_insert_balloon_locked() will ensure the node is inserted or reserved in GGTT. + * + * Return: A pointer to %xe_ggtt_node struct on success. An ERR_PTR otherwise. + **/ +struct xe_ggtt_node *xe_ggtt_node_init(struct xe_ggtt *ggtt) +{ + struct xe_ggtt_node *node = kzalloc(sizeof(*node), GFP_NOFS); + + if (!node) + return ERR_PTR(-ENOMEM); + + INIT_WORK(&node->delayed_removal_work, ggtt_node_remove_work_func); + node->ggtt = ggtt; + + return node; +} + +/** + * xe_ggtt_node_fini - Forcebly finalize %xe_ggtt_node struct + * @node: the &xe_ggtt_node to be freed + * + * If anything went wrong with either xe_ggtt_node_insert(), xe_ggtt_node_insert_locked(), + * or xe_ggtt_node_insert_balloon_locked(); and this @node is not going to be reused, then, + * this function needs to be called to free the %xe_ggtt_node struct + **/ +void xe_ggtt_node_fini(struct xe_ggtt_node *node) +{ + kfree(node); +} + +/** + * xe_ggtt_node_allocated - Check if node is allocated in GGTT + * @node: the &xe_ggtt_node to be inspected + * + * Return: True if allocated, False otherwise. + */ +bool xe_ggtt_node_allocated(const struct xe_ggtt_node *node) +{ + if (!node || !node->ggtt) + return false; + + return drm_mm_node_allocated(&node->base); +} + +/** + * xe_ggtt_node_pt_size() - Get the size of page table entries needed to map a GGTT node. + * @node: the &xe_ggtt_node + * + * Return: GGTT node page table entries size in bytes. + */ +size_t xe_ggtt_node_pt_size(const struct xe_ggtt_node *node) +{ + if (!node) + return 0; + + return node->base.size / XE_PAGE_SIZE * sizeof(u64); +} + +/** + * 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. + */ +void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_ggtt_node *node, + struct xe_bo *bo, u16 pat_index) +{ + + u64 start, pte, end; + struct xe_res_cursor cur; + + if (XE_WARN_ON(!node)) + return; + + start = node->base.start; + end = start + xe_bo_size(bo); + + 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 (xe_res_first_sg(xe_bo_sg(bo), 0, xe_bo_size(bo), &cur); + cur.remaining; xe_res_next(&cur, XE_PAGE_SIZE)) + ggtt->pt_ops->ggtt_set_pte(ggtt, end - cur.remaining, + pte | xe_res_dma(&cur)); + } else { + /* Prepend GPU offset */ + pte |= vram_region_gpu_offset(bo->ttm.resource); + + for (xe_res_first(bo->ttm.resource, 0, xe_bo_size(bo), &cur); + cur.remaining; xe_res_next(&cur, XE_PAGE_SIZE)) + ggtt->pt_ops->ggtt_set_pte(ggtt, end - cur.remaining, + pte + cur.start); + } +} + +/** + * xe_ggtt_map_bo_unlocked - Restore a mapping of a BO into GGTT + * @ggtt: the &xe_ggtt where node will be mapped + * @bo: the &xe_bo to be mapped + * + * This is used to restore a GGTT mapping after suspend. + */ +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]; - u64 start = bo->ggtt_node.start; - u64 offset, pte; - for (offset = 0; offset < bo->size; offset += XE_PAGE_SIZE) { - pte = ggtt->pt_ops->pte_encode_bo(bo, offset, pat_index); - xe_ggtt_set_pte(ggtt, start + offset, pte); - } + mutex_lock(&ggtt->lock); + xe_ggtt_map_bo(ggtt, bo->ggtt_node[ggtt->tile->id], bo, pat_index); + mutex_unlock(&ggtt->lock); } static int __xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo, - u64 start, u64 end) + u64 start, u64 end, struct drm_exec *exec) { + u64 alignment = bo->min_align > 0 ? bo->min_align : XE_PAGE_SIZE; + u8 tile_id = ggtt->tile->id; int err; - u64 alignment = XE_PAGE_SIZE; if (xe_bo_is_vram(bo) && ggtt->flags & XE_GGTT_FLAGS_64K) alignment = SZ_64K; - if (XE_WARN_ON(bo->ggtt_node.size)) { + if (XE_WARN_ON(bo->ggtt_node[tile_id])) { /* Someone's already inserted this BO in the GGTT */ - xe_tile_assert(ggtt->tile, bo->ggtt_node.size == bo->size); + xe_tile_assert(ggtt->tile, bo->ggtt_node[tile_id]->base.size == xe_bo_size(bo)); return 0; } - err = xe_bo_validate(bo, NULL, false); + err = xe_bo_validate(bo, NULL, false, exec); if (err) return err; xe_pm_runtime_get_noresume(tile_to_xe(ggtt->tile)); + + bo->ggtt_node[tile_id] = xe_ggtt_node_init(ggtt); + if (IS_ERR(bo->ggtt_node[tile_id])) { + err = PTR_ERR(bo->ggtt_node[tile_id]); + bo->ggtt_node[tile_id] = NULL; + goto out; + } + mutex_lock(&ggtt->lock); - err = drm_mm_insert_node_in_range(&ggtt->mm, &bo->ggtt_node, bo->size, - alignment, 0, start, end, 0); - if (!err) - xe_ggtt_map_bo(ggtt, bo); + err = drm_mm_insert_node_in_range(&ggtt->mm, &bo->ggtt_node[tile_id]->base, + xe_bo_size(bo), alignment, 0, start, end, 0); + if (err) { + xe_ggtt_node_fini(bo->ggtt_node[tile_id]); + bo->ggtt_node[tile_id] = NULL; + } else { + 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); if (!err && bo->flags & XE_BO_FLAG_GGTT_INVALIDATE) xe_ggtt_invalidate(ggtt); + +out: xe_pm_runtime_put(tile_to_xe(ggtt->tile)); return err; } +/** + * xe_ggtt_insert_bo_at - Insert BO at a specific GGTT space + * @ggtt: the &xe_ggtt where bo will be inserted + * @bo: the &xe_bo to be inserted + * @start: address where it will be inserted + * @end: end of the range where it will be inserted + * @exec: The drm_exec transaction to use for exhaustive eviction. + * + * Return: 0 on success or a negative error code on failure. + */ int xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo, - u64 start, u64 end) + u64 start, u64 end, struct drm_exec *exec) { - return __xe_ggtt_insert_bo_at(ggtt, bo, start, end); + return __xe_ggtt_insert_bo_at(ggtt, bo, start, end, exec); } -int xe_ggtt_insert_bo(struct xe_ggtt *ggtt, struct xe_bo *bo) +/** + * xe_ggtt_insert_bo - Insert BO into GGTT + * @ggtt: the &xe_ggtt where bo will be inserted + * @bo: the &xe_bo to be inserted + * @exec: The drm_exec transaction to use for exhaustive eviction. + * + * Return: 0 on success or a negative error code on failure. + */ +int xe_ggtt_insert_bo(struct xe_ggtt *ggtt, struct xe_bo *bo, + struct drm_exec *exec) { - return __xe_ggtt_insert_bo_at(ggtt, bo, 0, U64_MAX); + return __xe_ggtt_insert_bo_at(ggtt, bo, 0, U64_MAX, exec); } -void xe_ggtt_remove_node(struct xe_ggtt *ggtt, struct drm_mm_node *node, - bool invalidate) +/** + * xe_ggtt_remove_bo - Remove a BO from the GGTT + * @ggtt: the &xe_ggtt where node will be removed + * @bo: the &xe_bo to be removed + */ +void xe_ggtt_remove_bo(struct xe_ggtt *ggtt, struct xe_bo *bo) { - xe_pm_runtime_get_noresume(tile_to_xe(ggtt->tile)); + u8 tile_id = ggtt->tile->id; - mutex_lock(&ggtt->lock); - xe_ggtt_clear(ggtt, node->start, node->size); - drm_mm_remove_node(node); - node->size = 0; - mutex_unlock(&ggtt->lock); + if (XE_WARN_ON(!bo->ggtt_node[tile_id])) + return; - if (invalidate) - xe_ggtt_invalidate(ggtt); + /* This BO is not currently in the GGTT */ + xe_tile_assert(ggtt->tile, bo->ggtt_node[tile_id]->base.size == xe_bo_size(bo)); - xe_pm_runtime_put(tile_to_xe(ggtt->tile)); + xe_ggtt_node_remove(bo->ggtt_node[tile_id], + bo->flags & XE_BO_FLAG_GGTT_INVALIDATE); } -void xe_ggtt_remove_bo(struct xe_ggtt *ggtt, struct xe_bo *bo) +/** + * xe_ggtt_largest_hole - Largest GGTT hole + * @ggtt: the &xe_ggtt that will be inspected + * @alignment: minimum alignment + * @spare: If not NULL: in: desired memory size to be spared / out: Adjusted possible spare + * + * Return: size of the largest continuous GGTT region + */ +u64 xe_ggtt_largest_hole(struct xe_ggtt *ggtt, u64 alignment, u64 *spare) { - if (XE_WARN_ON(!bo->ggtt_node.size)) - return; + const struct drm_mm *mm = &ggtt->mm; + const struct drm_mm_node *entry; + u64 hole_min_start = xe_wopcm_size(tile_to_xe(ggtt->tile)); + u64 hole_start, hole_end, hole_size; + u64 max_hole = 0; - /* This BO is not currently in the GGTT */ - xe_tile_assert(ggtt->tile, bo->ggtt_node.size == bo->size); + mutex_lock(&ggtt->lock); - xe_ggtt_remove_node(ggtt, &bo->ggtt_node, - bo->flags & XE_BO_FLAG_GGTT_INVALIDATE); + drm_mm_for_each_hole(entry, mm, hole_start, hole_end) { + hole_start = max(hole_start, hole_min_start); + hole_start = ALIGN(hole_start, alignment); + hole_end = ALIGN_DOWN(hole_end, alignment); + if (hole_start >= hole_end) + continue; + hole_size = hole_end - hole_start; + if (spare) + *spare -= min3(*spare, hole_size, max_hole); + max_hole = max(max_hole, hole_size); + } + + mutex_unlock(&ggtt->lock); + + return max_hole; } #ifdef CONFIG_PCI_IOV @@ -478,7 +933,7 @@ static void xe_ggtt_assign_locked(struct xe_ggtt *ggtt, const struct drm_mm_node return; while (start < end) { - xe_ggtt_set_pte(ggtt, start, pte); + ggtt->pt_ops->ggtt_set_pte(ggtt, start, pte); start += XE_PAGE_SIZE; } @@ -487,22 +942,107 @@ static void xe_ggtt_assign_locked(struct xe_ggtt *ggtt, const struct drm_mm_node /** * xe_ggtt_assign - assign a GGTT region to the VF - * @ggtt: the &xe_ggtt where the node belongs - * @node: the &drm_mm_node to update + * @node: the &xe_ggtt_node to update * @vfid: the VF identifier * * This function is used by the PF driver to assign a GGTT region to the VF. * In addition to PTE's VFID bits 11:2 also PRESENT bit 0 is set as on some * platforms VFs can't modify that either. */ -void xe_ggtt_assign(struct xe_ggtt *ggtt, const struct drm_mm_node *node, u16 vfid) +void xe_ggtt_assign(const struct xe_ggtt_node *node, u16 vfid) { - mutex_lock(&ggtt->lock); - xe_ggtt_assign_locked(ggtt, node, vfid); - mutex_unlock(&ggtt->lock); + mutex_lock(&node->ggtt->lock); + xe_ggtt_assign_locked(node->ggtt, &node->base, vfid); + mutex_unlock(&node->ggtt->lock); +} + +/** + * xe_ggtt_node_save() - Save a &xe_ggtt_node to a buffer. + * @node: the &xe_ggtt_node to be saved + * @dst: destination buffer + * @size: destination buffer size in bytes + * @vfid: VF identifier + * + * Return: 0 on success or a negative error code on failure. + */ +int xe_ggtt_node_save(struct xe_ggtt_node *node, void *dst, size_t size, u16 vfid) +{ + struct xe_ggtt *ggtt; + u64 start, end; + u64 *buf = dst; + u64 pte; + + if (!node) + return -ENOENT; + + guard(mutex)(&node->ggtt->lock); + + if (xe_ggtt_node_pt_size(node) != size) + return -EINVAL; + + ggtt = node->ggtt; + start = node->base.start; + end = start + node->base.size - 1; + + while (start < end) { + pte = ggtt->pt_ops->ggtt_get_pte(ggtt, start); + if (vfid != u64_get_bits(pte, GGTT_PTE_VFID)) + return -EPERM; + + *buf++ = u64_replace_bits(pte, 0, GGTT_PTE_VFID); + start += XE_PAGE_SIZE; + } + + return 0; } + +/** + * xe_ggtt_node_load() - Load a &xe_ggtt_node from a buffer. + * @node: the &xe_ggtt_node to be loaded + * @src: source buffer + * @size: source buffer size in bytes + * @vfid: VF identifier + * + * Return: 0 on success or a negative error code on failure. + */ +int xe_ggtt_node_load(struct xe_ggtt_node *node, const void *src, size_t size, u16 vfid) +{ + u64 vfid_pte = xe_encode_vfid_pte(vfid); + const u64 *buf = src; + struct xe_ggtt *ggtt; + u64 start, end; + + if (!node) + return -ENOENT; + + guard(mutex)(&node->ggtt->lock); + + if (xe_ggtt_node_pt_size(node) != size) + return -EINVAL; + + ggtt = node->ggtt; + start = node->base.start; + end = start + node->base.size - 1; + + while (start < end) { + vfid_pte = u64_replace_bits(*buf++, vfid, GGTT_PTE_VFID); + ggtt->pt_ops->ggtt_set_pte(ggtt, start, vfid_pte); + start += XE_PAGE_SIZE; + } + xe_ggtt_invalidate(ggtt); + + return 0; +} + #endif +/** + * xe_ggtt_dump - Dump GGTT for debug + * @ggtt: the &xe_ggtt to be dumped + * @p: the &drm_mm_printer helper handle to be used to dump the information + * + * Return: 0 on success or a negative error code on failure. + */ int xe_ggtt_dump(struct xe_ggtt *ggtt, struct drm_printer *p) { int err; @@ -515,3 +1055,70 @@ int xe_ggtt_dump(struct xe_ggtt *ggtt, struct drm_printer *p) mutex_unlock(&ggtt->lock); return err; } + +/** + * xe_ggtt_print_holes - Print holes + * @ggtt: the &xe_ggtt to be inspected + * @alignment: min alignment + * @p: the &drm_printer + * + * Print GGTT ranges that are available and return total size available. + * + * Return: Total available size. + */ +u64 xe_ggtt_print_holes(struct xe_ggtt *ggtt, u64 alignment, struct drm_printer *p) +{ + const struct drm_mm *mm = &ggtt->mm; + const struct drm_mm_node *entry; + u64 hole_min_start = xe_wopcm_size(tile_to_xe(ggtt->tile)); + u64 hole_start, hole_end, hole_size; + u64 total = 0; + char buf[10]; + + mutex_lock(&ggtt->lock); + + drm_mm_for_each_hole(entry, mm, hole_start, hole_end) { + hole_start = max(hole_start, hole_min_start); + hole_start = ALIGN(hole_start, alignment); + hole_end = ALIGN_DOWN(hole_end, alignment); + if (hole_start >= hole_end) + continue; + hole_size = hole_end - hole_start; + total += hole_size; + + string_get_size(hole_size, 1, STRING_UNITS_2, buf, sizeof(buf)); + drm_printf(p, "range:\t%#llx-%#llx\t(%s)\n", + hole_start, hole_end - 1, buf); + } + + mutex_unlock(&ggtt->lock); + + return total; +} + +/** + * xe_ggtt_encode_pte_flags - Get PTE encoding flags for BO + * @ggtt: &xe_ggtt + * @bo: &xe_bo + * @pat_index: The pat_index for the PTE. + * + * This function returns the pte_flags for a given BO, without address. + * It's used for DPT to fill a GGTT mapped BO with a linear lookup table. + */ +u64 xe_ggtt_encode_pte_flags(struct xe_ggtt *ggtt, + struct xe_bo *bo, u16 pat_index) +{ + return ggtt->pt_ops->pte_encode_flags(bo, pat_index); +} + +/** + * xe_ggtt_read_pte - Read a PTE from the GGTT + * @ggtt: &xe_ggtt + * @offset: the offset for which the mapping should be read. + * + * Used by testcases, and by display reading out an inherited bios FB. + */ +u64 xe_ggtt_read_pte(struct xe_ggtt *ggtt, u64 offset) +{ + return ioread64(ggtt->gsm + (offset / XE_PAGE_SIZE)); +} |