From: Chris Wilson Date: Wed, 3 May 2017 09:39:21 +0000 (+0100) Subject: drm/i915: Squash repeated awaits on the same fence X-Git-Url: https://git.stricted.de/?a=commitdiff_plain;h=4797948071f6;p=GitHub%2Fmoto-9609%2Fandroid_kernel_motorola_exynos9610.git drm/i915: Squash repeated awaits on the same fence Track the latest fence waited upon on each context, and only add a new asynchronous wait if the new fence is more recent than the recorded fence for that context. This requires us to filter out unordered timelines, which are noted by DMA_FENCE_NO_CONTEXT. However, in the absence of a universal identifier, we have to use our own i915->mm.unordered_timeline token. v2: Throw around the debug crutches v3: Inline the likely case of the pre-allocation cache being full. v4: Drop the pre-allocation support, we can lose the most recent fence in case of allocation failure -- it just means we may emit more awaits than strictly necessary but will not break. v5: Trim allocation size for leaf nodes, they only need an array of u32 not pointers. v6: Create mock_timeline to tidy selftest writing v7: s/intel_timeline_sync_get/intel_timeline_sync_is_later/ (Tvrtko) v8: Prune the stale sync points when we idle. v9: Include a small benchmark in the kselftests v10: Separate the idr implementation into its own compartment. (Tvrkto) v11: Refactor igt_sync kselftests to avoid deep nesting (Tvrkto) v12: __sync_leaf_idx() to assert that p->height is 0 when checking leaves v13: kselftests to investigate struct i915_syncmap itself (Tvrtko) v14: Foray into ascii art graphs v15: Take into account that the random lookup/insert does 2 prng calls, not 1, when benchmarking, and use for_each_set_bit() (Tvrtko) v16: Improved ascii art Signed-off-by: Chris Wilson Cc: Tvrtko Ursulin Cc: Joonas Lahtinen Reviewed-by: Tvrtko Ursulin Link: http://patchwork.freedesktop.org/patch/msgid/20170503093924.5320-4-chris@chris-wilson.co.uk --- diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile index 2cf04504e494..7b05fb802f4c 100644 --- a/drivers/gpu/drm/i915/Makefile +++ b/drivers/gpu/drm/i915/Makefile @@ -16,6 +16,7 @@ i915-y := i915_drv.o \ i915_params.o \ i915_pci.o \ i915_suspend.o \ + i915_syncmap.o \ i915_sw_fence.o \ i915_sysfs.o \ intel_csr.o \ diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index e91590dae083..f9c6b9b5002c 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -3196,6 +3196,7 @@ i915_gem_idle_work_handler(struct work_struct *work) intel_engine_disarm_breadcrumbs(engine); i915_gem_batch_pool_fini(&engine->batch_pool); } + i915_gem_timelines_mark_idle(dev_priv); GEM_BUG_ON(!dev_priv->gt.awake); dev_priv->gt.awake = false; diff --git a/drivers/gpu/drm/i915/i915_gem.h b/drivers/gpu/drm/i915/i915_gem.h index 5a49487368ca..ee54597465b6 100644 --- a/drivers/gpu/drm/i915/i915_gem.h +++ b/drivers/gpu/drm/i915/i915_gem.h @@ -25,6 +25,8 @@ #ifndef __I915_GEM_H__ #define __I915_GEM_H__ +#include + #ifdef CONFIG_DRM_I915_DEBUG_GEM #define GEM_BUG_ON(expr) BUG_ON(expr) #define GEM_WARN_ON(expr) WARN_ON(expr) diff --git a/drivers/gpu/drm/i915/i915_gem_request.c b/drivers/gpu/drm/i915/i915_gem_request.c index 022f5588d906..637b8cddf988 100644 --- a/drivers/gpu/drm/i915/i915_gem_request.c +++ b/drivers/gpu/drm/i915/i915_gem_request.c @@ -773,6 +773,11 @@ i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req, if (fence->context == req->fence.context) continue; + /* Squash repeated waits to the same timelines */ + if (fence->context != req->i915->mm.unordered_timeline && + intel_timeline_sync_is_later(req->timeline, fence)) + continue; + if (dma_fence_is_i915(fence)) ret = i915_gem_request_await_request(req, to_request(fence)); @@ -782,6 +787,10 @@ i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req, GFP_KERNEL); if (ret < 0) return ret; + + /* Record the latest fence used against each timeline */ + if (fence->context != req->i915->mm.unordered_timeline) + intel_timeline_sync_set(req->timeline, fence); } while (--nchild); return 0; diff --git a/drivers/gpu/drm/i915/i915_gem_timeline.c b/drivers/gpu/drm/i915/i915_gem_timeline.c index b596ca7ee058..c597ce277a04 100644 --- a/drivers/gpu/drm/i915/i915_gem_timeline.c +++ b/drivers/gpu/drm/i915/i915_gem_timeline.c @@ -23,6 +23,32 @@ */ #include "i915_drv.h" +#include "i915_syncmap.h" + +static void __intel_timeline_init(struct intel_timeline *tl, + struct i915_gem_timeline *parent, + u64 context, + struct lock_class_key *lockclass, + const char *lockname) +{ + tl->fence_context = context; + tl->common = parent; +#ifdef CONFIG_DEBUG_SPINLOCK + __raw_spin_lock_init(&tl->lock.rlock, lockname, lockclass); +#else + spin_lock_init(&tl->lock); +#endif + init_request_active(&tl->last_request, NULL); + INIT_LIST_HEAD(&tl->requests); + i915_syncmap_init(&tl->sync); +} + +static void __intel_timeline_fini(struct intel_timeline *tl) +{ + GEM_BUG_ON(!list_empty(&tl->requests)); + + i915_syncmap_free(&tl->sync); +} static int __i915_gem_timeline_init(struct drm_i915_private *i915, struct i915_gem_timeline *timeline, @@ -35,6 +61,14 @@ static int __i915_gem_timeline_init(struct drm_i915_private *i915, lockdep_assert_held(&i915->drm.struct_mutex); + /* + * Ideally we want a set of engines on a single leaf as we expect + * to mostly be tracking synchronisation between engines. It is not + * a huge issue if this is not the case, but we may want to mitigate + * any page crossing penalties if they become an issue. + */ + BUILD_BUG_ON(KSYNCMAP < I915_NUM_ENGINES); + timeline->i915 = i915; timeline->name = kstrdup(name ?: "[kernel]", GFP_KERNEL); if (!timeline->name) @@ -44,19 +78,10 @@ static int __i915_gem_timeline_init(struct drm_i915_private *i915, /* Called during early_init before we know how many engines there are */ fences = dma_fence_context_alloc(ARRAY_SIZE(timeline->engine)); - for (i = 0; i < ARRAY_SIZE(timeline->engine); i++) { - struct intel_timeline *tl = &timeline->engine[i]; - - tl->fence_context = fences++; - tl->common = timeline; -#ifdef CONFIG_DEBUG_SPINLOCK - __raw_spin_lock_init(&tl->lock.rlock, lockname, lockclass); -#else - spin_lock_init(&tl->lock); -#endif - init_request_active(&tl->last_request, NULL); - INIT_LIST_HEAD(&tl->requests); - } + for (i = 0; i < ARRAY_SIZE(timeline->engine); i++) + __intel_timeline_init(&timeline->engine[i], + timeline, fences++, + lockclass, lockname); return 0; } @@ -81,18 +106,52 @@ int i915_gem_timeline_init__global(struct drm_i915_private *i915) &class, "&global_timeline->lock"); } +/** + * i915_gem_timelines_mark_idle -- called when the driver idles + * @i915 - the drm_i915_private device + * + * When the driver is completely idle, we know that all of our sync points + * have been signaled and our tracking is then entirely redundant. Any request + * to wait upon an older sync point will be completed instantly as we know + * the fence is signaled and therefore we will not even look them up in the + * sync point map. + */ +void i915_gem_timelines_mark_idle(struct drm_i915_private *i915) +{ + struct i915_gem_timeline *timeline; + int i; + + lockdep_assert_held(&i915->drm.struct_mutex); + + list_for_each_entry(timeline, &i915->gt.timelines, link) { + for (i = 0; i < ARRAY_SIZE(timeline->engine); i++) { + struct intel_timeline *tl = &timeline->engine[i]; + + /* + * All known fences are completed so we can scrap + * the current sync point tracking and start afresh, + * any attempt to wait upon a previous sync point + * will be skipped as the fence was signaled. + */ + i915_syncmap_free(&tl->sync); + } + } +} + void i915_gem_timeline_fini(struct i915_gem_timeline *timeline) { int i; lockdep_assert_held(&timeline->i915->drm.struct_mutex); - for (i = 0; i < ARRAY_SIZE(timeline->engine); i++) { - struct intel_timeline *tl = &timeline->engine[i]; - - GEM_BUG_ON(!list_empty(&tl->requests)); - } + for (i = 0; i < ARRAY_SIZE(timeline->engine); i++) + __intel_timeline_fini(&timeline->engine[i]); list_del(&timeline->link); kfree(timeline->name); } + +#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) +#include "selftests/mock_timeline.c" +#include "selftests/i915_gem_timeline.c" +#endif diff --git a/drivers/gpu/drm/i915/i915_gem_timeline.h b/drivers/gpu/drm/i915/i915_gem_timeline.h index 6c53e14cab2a..ff65c648407f 100644 --- a/drivers/gpu/drm/i915/i915_gem_timeline.h +++ b/drivers/gpu/drm/i915/i915_gem_timeline.h @@ -27,7 +27,9 @@ #include +#include "i915_utils.h" #include "i915_gem_request.h" +#include "i915_syncmap.h" struct i915_gem_timeline; @@ -55,6 +57,17 @@ struct intel_timeline { * struct_mutex. */ struct i915_gem_active last_request; + + /** + * We track the most recent seqno that we wait on in every context so + * that we only have to emit a new await and dependency on a more + * recent sync point. As the contexts may be executed out-of-order, we + * have to track each individually and can not rely on an absolute + * global_seqno. When we know that all tracked fences are completed + * (i.e. when the driver is idle), we know that the syncmap is + * redundant and we can discard it without loss of generality. + */ + struct i915_syncmap *sync; u32 sync_seqno[I915_NUM_ENGINES]; struct i915_gem_timeline *common; @@ -73,6 +86,31 @@ int i915_gem_timeline_init(struct drm_i915_private *i915, struct i915_gem_timeline *tl, const char *name); int i915_gem_timeline_init__global(struct drm_i915_private *i915); +void i915_gem_timelines_mark_idle(struct drm_i915_private *i915); void i915_gem_timeline_fini(struct i915_gem_timeline *tl); +static inline int __intel_timeline_sync_set(struct intel_timeline *tl, + u64 context, u32 seqno) +{ + return i915_syncmap_set(&tl->sync, context, seqno); +} + +static inline int intel_timeline_sync_set(struct intel_timeline *tl, + const struct dma_fence *fence) +{ + return __intel_timeline_sync_set(tl, fence->context, fence->seqno); +} + +static inline bool __intel_timeline_sync_is_later(struct intel_timeline *tl, + u64 context, u32 seqno) +{ + return i915_syncmap_is_later(&tl->sync, context, seqno); +} + +static inline bool intel_timeline_sync_is_later(struct intel_timeline *tl, + const struct dma_fence *fence) +{ + return __intel_timeline_sync_is_later(tl, fence->context, fence->seqno); +} + #endif diff --git a/drivers/gpu/drm/i915/i915_syncmap.c b/drivers/gpu/drm/i915/i915_syncmap.c new file mode 100644 index 000000000000..0087acf731a8 --- /dev/null +++ b/drivers/gpu/drm/i915/i915_syncmap.c @@ -0,0 +1,412 @@ +/* + * Copyright © 2017 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + */ + +#include + +#include "i915_syncmap.h" + +#include "i915_gem.h" /* GEM_BUG_ON() */ +#include "i915_selftest.h" + +#define SHIFT ilog2(KSYNCMAP) +#define MASK (KSYNCMAP - 1) + +/* + * struct i915_syncmap is a layer of a radixtree that maps a u64 fence + * context id to the last u32 fence seqno waited upon from that context. + * Unlike lib/radixtree it uses a parent pointer that allows traversal back to + * the root. This allows us to access the whole tree via a single pointer + * to the most recently used layer. We expect fence contexts to be dense + * and most reuse to be on the same i915_gem_context but on neighbouring + * engines (i.e. on adjacent contexts) and reuse the same leaf, a very + * effective lookup cache. If the new lookup is not on the same leaf, we + * expect it to be on the neighbouring branch. + * + * A leaf holds an array of u32 seqno, and has height 0. The bitmap field + * allows us to store whether a particular seqno is valid (i.e. allows us + * to distinguish unset from 0). + * + * A branch holds an array of layer pointers, and has height > 0, and always + * has at least 2 layers (either branches or leaves) below it. + * + * For example, + * for x in + * 0 1 2 0x10 0x11 0x200 0x201 + * 0x500000 0x500001 0x503000 0x503001 + * 0xE<<60: + * i915_syncmap_set(&sync, x, lower_32_bits(x)); + * will build a tree like: + * 0xXXXXXXXXXXXXXXXX + * 0-> 0x0000000000XXXXXX + * | 0-> 0x0000000000000XXX + * | | 0-> 0x00000000000000XX + * | | | 0-> 0x000000000000000X 0:0, 1:1, 2:2 + * | | | 1-> 0x000000000000001X 0:10, 1:11 + * | | 2-> 0x000000000000020X 0:200, 1:201 + * | 5-> 0x000000000050XXXX + * | 0-> 0x000000000050000X 0:500000, 1:500001 + * | 3-> 0x000000000050300X 0:503000, 1:503001 + * e-> 0xe00000000000000X e:e + */ + +struct i915_syncmap { + u64 prefix; + unsigned int height; + unsigned int bitmap; + struct i915_syncmap *parent; + /* + * Following this header is an array of either seqno or child pointers: + * union { + * u32 seqno[KSYNCMAP]; + * struct i915_syncmap *child[KSYNCMAP]; + * }; + */ +}; + +/** + * i915_syncmap_init -- initialise the #i915_syncmap + * @root - pointer to the #i915_syncmap + */ +void i915_syncmap_init(struct i915_syncmap **root) +{ + BUILD_BUG_ON_NOT_POWER_OF_2(KSYNCMAP); + BUILD_BUG_ON_NOT_POWER_OF_2(SHIFT); + BUILD_BUG_ON(KSYNCMAP > BITS_PER_BYTE * sizeof((*root)->bitmap)); + *root = NULL; +} + +static inline u32 *__sync_seqno(struct i915_syncmap *p) +{ + GEM_BUG_ON(p->height); + return (u32 *)(p + 1); +} + +static inline struct i915_syncmap **__sync_child(struct i915_syncmap *p) +{ + GEM_BUG_ON(!p->height); + return (struct i915_syncmap **)(p + 1); +} + +static inline unsigned int +__sync_branch_idx(const struct i915_syncmap *p, u64 id) +{ + return (id >> p->height) & MASK; +} + +static inline unsigned int +__sync_leaf_idx(const struct i915_syncmap *p, u64 id) +{ + GEM_BUG_ON(p->height); + return id & MASK; +} + +static inline u64 __sync_branch_prefix(const struct i915_syncmap *p, u64 id) +{ + return id >> p->height >> SHIFT; +} + +static inline u64 __sync_leaf_prefix(const struct i915_syncmap *p, u64 id) +{ + GEM_BUG_ON(p->height); + return id >> SHIFT; +} + +static inline bool seqno_later(u32 a, u32 b) +{ + return (s32)(a - b) >= 0; +} + +/** + * i915_syncmap_is_later -- compare against the last know sync point + * @root - pointer to the #i915_syncmap + * @id - the context id (other timeline) we are synchronising to + * @seqno - the sequence number along the other timeline + * + * If we have already synchronised this @root timeline with another (@id) then + * we can omit any repeated or earlier synchronisation requests. If the two + * timelines are already coupled, we can also omit the dependency between the + * two as that is already known via the timeline. + * + * Returns true if the two timelines are already synchronised wrt to @seqno, + * false if not and the synchronisation must be emitted. + */ +bool i915_syncmap_is_later(struct i915_syncmap **root, u64 id, u32 seqno) +{ + struct i915_syncmap *p; + unsigned int idx; + + p = *root; + if (!p) + return false; + + if (likely(__sync_leaf_prefix(p, id) == p->prefix)) + goto found; + + /* First climb the tree back to a parent branch */ + do { + p = p->parent; + if (!p) + return false; + + if (__sync_branch_prefix(p, id) == p->prefix) + break; + } while (1); + + /* And then descend again until we find our leaf */ + do { + if (!p->height) + break; + + p = __sync_child(p)[__sync_branch_idx(p, id)]; + if (!p) + return false; + + if (__sync_branch_prefix(p, id) != p->prefix) + return false; + } while (1); + + *root = p; +found: + idx = __sync_leaf_idx(p, id); + if (!(p->bitmap & BIT(idx))) + return false; + + return seqno_later(__sync_seqno(p)[idx], seqno); +} + +static struct i915_syncmap * +__sync_alloc_leaf(struct i915_syncmap *parent, u64 id) +{ + struct i915_syncmap *p; + + p = kmalloc(sizeof(*p) + KSYNCMAP * sizeof(u32), GFP_KERNEL); + if (unlikely(!p)) + return NULL; + + p->parent = parent; + p->height = 0; + p->bitmap = 0; + p->prefix = __sync_leaf_prefix(p, id); + return p; +} + +static inline void __sync_set_seqno(struct i915_syncmap *p, u64 id, u32 seqno) +{ + unsigned int idx = __sync_leaf_idx(p, id); + + p->bitmap |= BIT(idx); + __sync_seqno(p)[idx] = seqno; +} + +static inline void __sync_set_child(struct i915_syncmap *p, + unsigned int idx, + struct i915_syncmap *child) +{ + p->bitmap |= BIT(idx); + __sync_child(p)[idx] = child; +} + +static noinline int __sync_set(struct i915_syncmap **root, u64 id, u32 seqno) +{ + struct i915_syncmap *p = *root; + unsigned int idx; + + if (!p) { + p = __sync_alloc_leaf(NULL, id); + if (unlikely(!p)) + return -ENOMEM; + + goto found; + } + + /* Caller handled the likely cached case */ + GEM_BUG_ON(__sync_leaf_prefix(p, id) == p->prefix); + + /* Climb back up the tree until we find a common prefix */ + do { + if (!p->parent) + break; + + p = p->parent; + + if (__sync_branch_prefix(p, id) == p->prefix) + break; + } while (1); + + /* + * No shortcut, we have to descend the tree to find the right layer + * containing this fence. + * + * Each layer in the tree holds 16 (KSYNCMAP) pointers, either fences + * or lower layers. Leaf nodes (height = 0) contain the fences, all + * other nodes (height > 0) are internal layers that point to a lower + * node. Each internal layer has at least 2 descendents. + * + * Starting at the top, we check whether the current prefix matches. If + * it doesn't, we have gone past our target and need to insert a join + * into the tree, and a new leaf node for the target as a descendent + * of the join, as well as the original layer. + * + * The matching prefix means we are still following the right branch + * of the tree. If it has height 0, we have found our leaf and just + * need to replace the fence slot with ourselves. If the height is + * not zero, our slot contains the next layer in the tree (unless + * it is empty, in which case we can add ourselves as a new leaf). + * As descend the tree the prefix grows (and height decreases). + */ + do { + struct i915_syncmap *next; + + if (__sync_branch_prefix(p, id) != p->prefix) { + unsigned int above; + + /* Insert a join above the current layer */ + next = kzalloc(sizeof(*next) + KSYNCMAP * sizeof(next), + GFP_KERNEL); + if (unlikely(!next)) + return -ENOMEM; + + /* Compute the height at which these two diverge */ + above = fls64(__sync_branch_prefix(p, id) ^ p->prefix); + above = round_up(above, SHIFT); + next->height = above + p->height; + next->prefix = __sync_branch_prefix(next, id); + + /* Insert the join into the parent */ + if (p->parent) { + idx = __sync_branch_idx(p->parent, id); + __sync_child(p->parent)[idx] = next; + GEM_BUG_ON(!(p->parent->bitmap & BIT(idx))); + } + next->parent = p->parent; + + /* Compute the idx of the other branch, not our id! */ + idx = p->prefix >> (above - SHIFT) & MASK; + __sync_set_child(next, idx, p); + p->parent = next; + + /* Ascend to the join */ + p = next; + } else { + if (!p->height) + break; + } + + /* Descend into the next layer */ + GEM_BUG_ON(!p->height); + idx = __sync_branch_idx(p, id); + next = __sync_child(p)[idx]; + if (!next) { + next = __sync_alloc_leaf(p, id); + if (unlikely(!next)) + return -ENOMEM; + + __sync_set_child(p, idx, next); + p = next; + break; + } + + p = next; + } while (1); + +found: + GEM_BUG_ON(p->prefix != __sync_leaf_prefix(p, id)); + __sync_set_seqno(p, id, seqno); + *root = p; + return 0; +} + +/** + * i915_syncmap_set -- mark the most recent syncpoint between contexts + * @root - pointer to the #i915_syncmap + * @id - the context id (other timeline) we have synchronised to + * @seqno - the sequence number along the other timeline + * + * When we synchronise this @root timeline with another (@id), we also know + * that we have synchronized with all previous seqno along that timeline. If + * we then have a request to synchronise with the same seqno or older, we can + * omit it, see i915_syncmap_is_later() + * + * Returns 0 on success, or a negative error code. + */ +int i915_syncmap_set(struct i915_syncmap **root, u64 id, u32 seqno) +{ + struct i915_syncmap *p = *root; + + /* + * We expect to be called in sequence following is_later(id), which + * should have preloaded the root for us. + */ + if (likely(p && __sync_leaf_prefix(p, id) == p->prefix)) { + __sync_set_seqno(p, id, seqno); + return 0; + } + + return __sync_set(root, id, seqno); +} + +static void __sync_free(struct i915_syncmap *p) +{ + if (p->height) { + unsigned int i; + + while ((i = ffs(p->bitmap))) { + p->bitmap &= ~0u << i; + __sync_free(__sync_child(p)[i - 1]); + } + } + + kfree(p); +} + +/** + * i915_syncmap_free -- free all memory associated with the syncmap + * @root - pointer to the #i915_syncmap + * + * Either when the timeline is to be freed and we no longer need the sync + * point tracking, or when the fences are all known to be signaled and the + * sync point tracking is redundant, we can free the #i915_syncmap to recover + * its allocations. + * + * Will reinitialise the @root pointer so that the #i915_syncmap is ready for + * reuse. + */ +void i915_syncmap_free(struct i915_syncmap **root) +{ + struct i915_syncmap *p; + + p = *root; + if (!p) + return; + + while (p->parent) + p = p->parent; + + __sync_free(p); + *root = NULL; +} + +#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) +#include "selftests/i915_syncmap.c" +#endif diff --git a/drivers/gpu/drm/i915/i915_syncmap.h b/drivers/gpu/drm/i915/i915_syncmap.h new file mode 100644 index 000000000000..0653f70bee82 --- /dev/null +++ b/drivers/gpu/drm/i915/i915_syncmap.h @@ -0,0 +1,38 @@ +/* + * Copyright © 2017 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + */ + +#ifndef __I915_SYNCMAP_H__ +#define __I915_SYNCMAP_H__ + +#include + +struct i915_syncmap; +#define KSYNCMAP 16 /* radix of the tree, how many slots in each layer */ + +void i915_syncmap_init(struct i915_syncmap **root); +int i915_syncmap_set(struct i915_syncmap **root, u64 id, u32 seqno); +bool i915_syncmap_is_later(struct i915_syncmap **root, u64 id, u32 seqno); +void i915_syncmap_free(struct i915_syncmap **root); + +#endif /* __I915_SYNCMAP_H__ */ diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_timeline.c b/drivers/gpu/drm/i915/selftests/i915_gem_timeline.c new file mode 100644 index 000000000000..6df00cc02c12 --- /dev/null +++ b/drivers/gpu/drm/i915/selftests/i915_gem_timeline.c @@ -0,0 +1,301 @@ +/* + * Copyright © 2017 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + */ + +#include "../i915_selftest.h" +#include "i915_random.h" + +#include "mock_gem_device.h" +#include "mock_timeline.h" + +struct __igt_sync { + const char *name; + u32 seqno; + bool expected; + bool set; +}; + +static int __igt_sync(struct intel_timeline *tl, + u64 ctx, + const struct __igt_sync *p, + const char *name) +{ + int ret; + + if (__intel_timeline_sync_is_later(tl, ctx, p->seqno) != p->expected) { + pr_err("%s: %s(ctx=%llu, seqno=%u) expected passed %s but failed\n", + name, p->name, ctx, p->seqno, yesno(p->expected)); + return -EINVAL; + } + + if (p->set) { + ret = __intel_timeline_sync_set(tl, ctx, p->seqno); + if (ret) + return ret; + } + + return 0; +} + +static int igt_sync(void *arg) +{ + const struct __igt_sync pass[] = { + { "unset", 0, false, false }, + { "new", 0, false, true }, + { "0a", 0, true, true }, + { "1a", 1, false, true }, + { "1b", 1, true, true }, + { "0b", 0, true, false }, + { "2a", 2, false, true }, + { "4", 4, false, true }, + { "INT_MAX", INT_MAX, false, true }, + { "INT_MAX-1", INT_MAX-1, true, false }, + { "INT_MAX+1", (u32)INT_MAX+1, false, true }, + { "INT_MAX", INT_MAX, true, false }, + { "UINT_MAX", UINT_MAX, false, true }, + { "wrap", 0, false, true }, + { "unwrap", UINT_MAX, true, false }, + {}, + }, *p; + struct intel_timeline *tl; + int order, offset; + int ret; + + tl = mock_timeline(0); + if (!tl) + return -ENOMEM; + + for (p = pass; p->name; p++) { + for (order = 1; order < 64; order++) { + for (offset = -1; offset <= (order > 1); offset++) { + u64 ctx = BIT_ULL(order) + offset; + + ret = __igt_sync(tl, ctx, p, "1"); + if (ret) + goto out; + } + } + } + mock_timeline_destroy(tl); + + tl = mock_timeline(0); + if (!tl) + return -ENOMEM; + + for (order = 1; order < 64; order++) { + for (offset = -1; offset <= (order > 1); offset++) { + u64 ctx = BIT_ULL(order) + offset; + + for (p = pass; p->name; p++) { + ret = __igt_sync(tl, ctx, p, "2"); + if (ret) + goto out; + } + } + } + +out: + mock_timeline_destroy(tl); + return ret; +} + +static unsigned int random_engine(struct rnd_state *rnd) +{ + return ((u64)prandom_u32_state(rnd) * I915_NUM_ENGINES) >> 32; +} + +static int bench_sync(void *arg) +{ +#define M (1 << 20) + struct rnd_state prng; + struct intel_timeline *tl; + unsigned long end_time, count; + u64 prng32_1M; + ktime_t kt; + int order, last_order; + + tl = mock_timeline(0); + if (!tl) + return -ENOMEM; + + /* Lookups from cache are very fast and so the random number generation + * and the loop itself becomes a significant factor in the per-iteration + * timings. We try to compensate the results by measuring the overhead + * of the prng and subtract it from the reported results. + */ + prandom_seed_state(&prng, i915_selftest.random_seed); + count = 0; + kt = ktime_get(); + end_time = jiffies + HZ/10; + do { + u32 x; + + /* Make sure the compiler doesn't optimise away the prng call */ + WRITE_ONCE(x, prandom_u32_state(&prng)); + + count++; + } while (!time_after(jiffies, end_time)); + kt = ktime_sub(ktime_get(), kt); + pr_debug("%s: %lu random evaluations, %lluns/prng\n", + __func__, count, (long long)div64_ul(ktime_to_ns(kt), count)); + prng32_1M = ktime_to_ns(kt) * M / count; + + /* Benchmark (only) setting random context ids */ + prandom_seed_state(&prng, i915_selftest.random_seed); + count = 0; + kt = ktime_get(); + end_time = jiffies + HZ/10; + do { + u64 id = i915_prandom_u64_state(&prng); + + __intel_timeline_sync_set(tl, id, 0); + count++; + } while (!time_after(jiffies, end_time)); + kt = ktime_sub(ktime_get(), kt); + kt = ktime_sub_ns(kt, count * prng32_1M * 2 / M); + pr_info("%s: %lu random insertions, %lluns/insert\n", + __func__, count, (long long)div64_ul(ktime_to_ns(kt), count)); + + /* Benchmark looking up the exact same context ids as we just set */ + prandom_seed_state(&prng, i915_selftest.random_seed); + end_time = count; + kt = ktime_get(); + while (end_time--) { + u64 id = i915_prandom_u64_state(&prng); + + if (!__intel_timeline_sync_is_later(tl, id, 0)) { + mock_timeline_destroy(tl); + pr_err("Lookup of %llu failed\n", id); + return -EINVAL; + } + } + kt = ktime_sub(ktime_get(), kt); + kt = ktime_sub_ns(kt, count * prng32_1M * 2 / M); + pr_info("%s: %lu random lookups, %lluns/lookup\n", + __func__, count, (long long)div64_ul(ktime_to_ns(kt), count)); + + mock_timeline_destroy(tl); + cond_resched(); + + tl = mock_timeline(0); + if (!tl) + return -ENOMEM; + + /* Benchmark setting the first N (in order) contexts */ + count = 0; + kt = ktime_get(); + end_time = jiffies + HZ/10; + do { + __intel_timeline_sync_set(tl, count++, 0); + } while (!time_after(jiffies, end_time)); + kt = ktime_sub(ktime_get(), kt); + pr_info("%s: %lu in-order insertions, %lluns/insert\n", + __func__, count, (long long)div64_ul(ktime_to_ns(kt), count)); + + /* Benchmark looking up the exact same context ids as we just set */ + end_time = count; + kt = ktime_get(); + while (end_time--) { + if (!__intel_timeline_sync_is_later(tl, end_time, 0)) { + pr_err("Lookup of %lu failed\n", end_time); + mock_timeline_destroy(tl); + return -EINVAL; + } + } + kt = ktime_sub(ktime_get(), kt); + pr_info("%s: %lu in-order lookups, %lluns/lookup\n", + __func__, count, (long long)div64_ul(ktime_to_ns(kt), count)); + + mock_timeline_destroy(tl); + cond_resched(); + + tl = mock_timeline(0); + if (!tl) + return -ENOMEM; + + /* Benchmark searching for a random context id and maybe changing it */ + prandom_seed_state(&prng, i915_selftest.random_seed); + count = 0; + kt = ktime_get(); + end_time = jiffies + HZ/10; + do { + u32 id = random_engine(&prng); + u32 seqno = prandom_u32_state(&prng); + + if (!__intel_timeline_sync_is_later(tl, id, seqno)) + __intel_timeline_sync_set(tl, id, seqno); + + count++; + } while (!time_after(jiffies, end_time)); + kt = ktime_sub(ktime_get(), kt); + kt = ktime_sub_ns(kt, count * prng32_1M * 2 / M); + pr_info("%s: %lu repeated insert/lookups, %lluns/op\n", + __func__, count, (long long)div64_ul(ktime_to_ns(kt), count)); + mock_timeline_destroy(tl); + cond_resched(); + + /* Benchmark searching for a known context id and changing the seqno */ + for (last_order = 1, order = 1; order < 32; + ({ int tmp = last_order; last_order = order; order += tmp; })) { + unsigned int mask = BIT(order) - 1; + + tl = mock_timeline(0); + if (!tl) + return -ENOMEM; + + count = 0; + kt = ktime_get(); + end_time = jiffies + HZ/10; + do { + /* Without assuming too many details of the underlying + * implementation, try to identify its phase-changes + * (if any)! + */ + u64 id = (u64)(count & mask) << order; + + __intel_timeline_sync_is_later(tl, id, 0); + __intel_timeline_sync_set(tl, id, 0); + + count++; + } while (!time_after(jiffies, end_time)); + kt = ktime_sub(ktime_get(), kt); + pr_info("%s: %lu cyclic/%d insert/lookups, %lluns/op\n", + __func__, count, order, + (long long)div64_ul(ktime_to_ns(kt), count)); + mock_timeline_destroy(tl); + cond_resched(); + } + + return 0; +#undef M +} + +int i915_gem_timeline_mock_selftests(void) +{ + static const struct i915_subtest tests[] = { + SUBTEST(igt_sync), + SUBTEST(bench_sync), + }; + + return i915_subtests(tests, NULL); +} diff --git a/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h b/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h index be9a9ebf5692..76c1f149a0a0 100644 --- a/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h +++ b/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h @@ -10,8 +10,10 @@ */ selftest(sanitycheck, i915_mock_sanitycheck) /* keep first (igt selfcheck) */ selftest(scatterlist, scatterlist_mock_selftests) +selftest(syncmap, i915_syncmap_mock_selftests) selftest(uncore, intel_uncore_mock_selftests) selftest(breadcrumbs, intel_breadcrumbs_mock_selftests) +selftest(timelines, i915_gem_timeline_mock_selftests) selftest(requests, i915_gem_request_mock_selftests) selftest(objects, i915_gem_object_mock_selftests) selftest(dmabuf, i915_gem_dmabuf_mock_selftests) diff --git a/drivers/gpu/drm/i915/selftests/i915_random.c b/drivers/gpu/drm/i915/selftests/i915_random.c index c17c83c30637..d044bf9a6feb 100644 --- a/drivers/gpu/drm/i915/selftests/i915_random.c +++ b/drivers/gpu/drm/i915/selftests/i915_random.c @@ -30,6 +30,17 @@ #include "i915_random.h" +u64 i915_prandom_u64_state(struct rnd_state *rnd) +{ + u64 x; + + x = prandom_u32_state(rnd); + x <<= 32; + x |= prandom_u32_state(rnd); + + return x; +} + static inline u32 i915_prandom_u32_max_state(u32 ep_ro, struct rnd_state *state) { return upper_32_bits((u64)prandom_u32_state(state) * ep_ro); diff --git a/drivers/gpu/drm/i915/selftests/i915_random.h b/drivers/gpu/drm/i915/selftests/i915_random.h index b9c334ce6cd9..6c9379871384 100644 --- a/drivers/gpu/drm/i915/selftests/i915_random.h +++ b/drivers/gpu/drm/i915/selftests/i915_random.h @@ -41,6 +41,8 @@ #define I915_RND_SUBSTATE(name__, parent__) \ struct rnd_state name__ = I915_RND_STATE_INITIALIZER(prandom_u32_state(&(parent__))) +u64 i915_prandom_u64_state(struct rnd_state *rnd); + unsigned int *i915_random_order(unsigned int count, struct rnd_state *state); void i915_random_reorder(unsigned int *order, diff --git a/drivers/gpu/drm/i915/selftests/i915_syncmap.c b/drivers/gpu/drm/i915/selftests/i915_syncmap.c new file mode 100644 index 000000000000..bcab3d00a785 --- /dev/null +++ b/drivers/gpu/drm/i915/selftests/i915_syncmap.c @@ -0,0 +1,616 @@ +/* + * Copyright © 2017 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + */ + +#include "../i915_selftest.h" +#include "i915_random.h" + +static char * +__sync_print(struct i915_syncmap *p, + char *buf, unsigned long *sz, + unsigned int depth, + unsigned int last, + unsigned int idx) +{ + unsigned long len; + unsigned int i, X; + + if (depth) { + unsigned int d; + + for (d = 0; d < depth - 1; d++) { + if (last & BIT(depth - d - 1)) + len = scnprintf(buf, *sz, "| "); + else + len = scnprintf(buf, *sz, " "); + buf += len; + *sz -= len; + } + len = scnprintf(buf, *sz, "%x-> ", idx); + buf += len; + *sz -= len; + } + + /* We mark bits after the prefix as "X" */ + len = scnprintf(buf, *sz, "0x%016llx", p->prefix << p->height << SHIFT); + buf += len; + *sz -= len; + X = (p->height + SHIFT) / 4; + scnprintf(buf - X, *sz + X, "%*s", X, "XXXXXXXXXXXXXXXXX"); + + if (!p->height) { + for_each_set_bit(i, (unsigned long *)&p->bitmap, KSYNCMAP) { + len = scnprintf(buf, *sz, " %x:%x,", + i, __sync_seqno(p)[i]); + buf += len; + *sz -= len; + } + buf -= 1; + *sz += 1; + } + + len = scnprintf(buf, *sz, "\n"); + buf += len; + *sz -= len; + + if (p->height) { + for_each_set_bit(i, (unsigned long *)&p->bitmap, KSYNCMAP) { + buf = __sync_print(__sync_child(p)[i], buf, sz, + depth + 1, + last << 1 | !!(p->bitmap >> (i + 1)), + i); + } + } + + return buf; +} + +static bool +i915_syncmap_print_to_buf(struct i915_syncmap *p, char *buf, unsigned long sz) +{ + if (!p) + return false; + + while (p->parent) + p = p->parent; + + __sync_print(p, buf, &sz, 0, 1, 0); + return true; +} + +static int check_syncmap_free(struct i915_syncmap **sync) +{ + i915_syncmap_free(sync); + if (*sync) { + pr_err("sync not cleared after free\n"); + return -EINVAL; + } + + return 0; +} + +static int dump_syncmap(struct i915_syncmap *sync, int err) +{ + char *buf; + + if (!err) + return check_syncmap_free(&sync); + + buf = kmalloc(PAGE_SIZE, GFP_KERNEL); + if (!buf) + goto skip; + + if (i915_syncmap_print_to_buf(sync, buf, PAGE_SIZE)) + pr_err("%s", buf); + + kfree(buf); + +skip: + i915_syncmap_free(&sync); + return err; +} + +static int igt_syncmap_init(void *arg) +{ + struct i915_syncmap *sync = (void *)~0ul; + + /* + * Cursory check that we can initialise a random pointer and transform + * it into the root pointer of a syncmap. + */ + + i915_syncmap_init(&sync); + return check_syncmap_free(&sync); +} + +static int check_seqno(struct i915_syncmap *leaf, unsigned int idx, u32 seqno) +{ + if (leaf->height) { + pr_err("%s: not a leaf, height is %d\n", + __func__, leaf->height); + return -EINVAL; + } + + if (__sync_seqno(leaf)[idx] != seqno) { + pr_err("%s: seqno[%d], found %x, expected %x\n", + __func__, idx, __sync_seqno(leaf)[idx], seqno); + return -EINVAL; + } + + return 0; +} + +static int check_one(struct i915_syncmap **sync, u64 context, u32 seqno) +{ + int err; + + err = i915_syncmap_set(sync, context, seqno); + if (err) + return err; + + if ((*sync)->height) { + pr_err("Inserting first context=%llx did not return leaf (height=%d, prefix=%llx\n", + context, (*sync)->height, (*sync)->prefix); + return -EINVAL; + } + + if ((*sync)->parent) { + pr_err("Inserting first context=%llx created branches!\n", + context); + return -EINVAL; + } + + if (hweight32((*sync)->bitmap) != 1) { + pr_err("First bitmap does not contain a single entry, found %x (count=%d)!\n", + (*sync)->bitmap, hweight32((*sync)->bitmap)); + return -EINVAL; + } + + err = check_seqno((*sync), ilog2((*sync)->bitmap), seqno); + if (err) + return err; + + if (!i915_syncmap_is_later(sync, context, seqno)) { + pr_err("Lookup of first context=%llx/seqno=%x failed!\n", + context, seqno); + return -EINVAL; + } + + return 0; +} + +static int igt_syncmap_one(void *arg) +{ + I915_RND_STATE(prng); + IGT_TIMEOUT(end_time); + struct i915_syncmap *sync; + unsigned long max = 1; + int err; + + /* + * Check that inserting a new id, creates a leaf and only that leaf. + */ + + i915_syncmap_init(&sync); + + do { + u64 context = i915_prandom_u64_state(&prng); + unsigned long loop; + + err = check_syncmap_free(&sync); + if (err) + goto out; + + for (loop = 0; loop <= max; loop++) { + err = check_one(&sync, context, + prandom_u32_state(&prng)); + if (err) + goto out; + } + max++; + } while (!__igt_timeout(end_time, NULL)); + pr_debug("%s: Completed %lu single insertions\n", + __func__, max * (max - 1) / 2); +out: + return dump_syncmap(sync, err); +} + +static int check_leaf(struct i915_syncmap **sync, u64 context, u32 seqno) +{ + int err; + + err = i915_syncmap_set(sync, context, seqno); + if (err) + return err; + + if ((*sync)->height) { + pr_err("Inserting context=%llx did not return leaf (height=%d, prefix=%llx\n", + context, (*sync)->height, (*sync)->prefix); + return -EINVAL; + } + + if (hweight32((*sync)->bitmap) != 1) { + pr_err("First entry into leaf (context=%llx) does not contain a single entry, found %x (count=%d)!\n", + context, (*sync)->bitmap, hweight32((*sync)->bitmap)); + return -EINVAL; + } + + err = check_seqno((*sync), ilog2((*sync)->bitmap), seqno); + if (err) + return err; + + if (!i915_syncmap_is_later(sync, context, seqno)) { + pr_err("Lookup of first entry context=%llx/seqno=%x failed!\n", + context, seqno); + return -EINVAL; + } + + return 0; +} + +static int igt_syncmap_join_above(void *arg) +{ + struct i915_syncmap *sync; + unsigned int pass, order; + int err; + + i915_syncmap_init(&sync); + + /* + * When we have a new id that doesn't fit inside the existing tree, + * we need to add a new layer above. + * + * 1: 0x00000001 + * 2: 0x00000010 + * 3: 0x00000100 + * 4: 0x00001000 + * ... + * Each pass the common prefix shrinks and we have to insert a join. + * Each join will only contain two branches, the latest of which + * is always a leaf. + * + * If we then reuse the same set of contexts, we expect to build an + * identical tree. + */ + for (pass = 0; pass < 3; pass++) { + for (order = 0; order < 64; order += SHIFT) { + u64 context = BIT_ULL(order); + struct i915_syncmap *join; + + err = check_leaf(&sync, context, 0); + if (err) + goto out; + + join = sync->parent; + if (!join) /* very first insert will have no parents */ + continue; + + if (!join->height) { + pr_err("Parent with no height!\n"); + err = -EINVAL; + goto out; + } + + if (hweight32(join->bitmap) != 2) { + pr_err("Join does not have 2 children: %x (%d)\n", + join->bitmap, hweight32(join->bitmap)); + err = -EINVAL; + goto out; + } + + if (__sync_child(join)[__sync_branch_idx(join, context)] != sync) { + pr_err("Leaf misplaced in parent!\n"); + err = -EINVAL; + goto out; + } + } + } +out: + return dump_syncmap(sync, err); +} + +static int igt_syncmap_join_below(void *arg) +{ + struct i915_syncmap *sync; + unsigned int step, order, idx; + int err; + + i915_syncmap_init(&sync); + + /* + * Check that we can split a compacted branch by replacing it with + * a join. + */ + for (step = 0; step < KSYNCMAP; step++) { + for (order = 64 - SHIFT; order > 0; order -= SHIFT) { + u64 context = step * BIT_ULL(order); + + err = i915_syncmap_set(&sync, context, 0); + if (err) + goto out; + + if (sync->height) { + pr_err("Inserting context=%llx (order=%d, step=%d) did not return leaf (height=%d, prefix=%llx\n", + context, order, step, sync->height, sync->prefix); + err = -EINVAL; + goto out; + } + } + } + + for (step = 0; step < KSYNCMAP; step++) { + for (order = SHIFT; order < 64; order += SHIFT) { + u64 context = step * BIT_ULL(order); + + if (!i915_syncmap_is_later(&sync, context, 0)) { + pr_err("1: context %llx (order=%d, step=%d) not found\n", + context, order, step); + err = -EINVAL; + goto out; + } + + for (idx = 1; idx < KSYNCMAP; idx++) { + if (i915_syncmap_is_later(&sync, context + idx, 0)) { + pr_err("1: context %llx (order=%d, step=%d) should not exist\n", + context + idx, order, step); + err = -EINVAL; + goto out; + } + } + } + } + + for (order = SHIFT; order < 64; order += SHIFT) { + for (step = 0; step < KSYNCMAP; step++) { + u64 context = step * BIT_ULL(order); + + if (!i915_syncmap_is_later(&sync, context, 0)) { + pr_err("2: context %llx (order=%d, step=%d) not found\n", + context, order, step); + err = -EINVAL; + goto out; + } + } + } + +out: + return dump_syncmap(sync, err); +} + +static int igt_syncmap_neighbours(void *arg) +{ + I915_RND_STATE(prng); + IGT_TIMEOUT(end_time); + struct i915_syncmap *sync; + int err; + + /* + * Each leaf holds KSYNCMAP seqno. Check that when we create KSYNCMAP + * neighbouring ids, they all fit into the same leaf. + */ + + i915_syncmap_init(&sync); + do { + u64 context = i915_prandom_u64_state(&prng) & ~MASK; + unsigned int idx; + + if (i915_syncmap_is_later(&sync, context, 0)) /* Skip repeats */ + continue; + + for (idx = 0; idx < KSYNCMAP; idx++) { + err = i915_syncmap_set(&sync, context + idx, 0); + if (err) + goto out; + + if (sync->height) { + pr_err("Inserting context=%llx did not return leaf (height=%d, prefix=%llx\n", + context, sync->height, sync->prefix); + err = -EINVAL; + goto out; + } + + if (sync->bitmap != BIT(idx + 1) - 1) { + pr_err("Inserting neighbouring context=0x%llx+%d, did not fit into the same leaf bitmap=%x (%d), expected %lx (%d)\n", + context, idx, + sync->bitmap, hweight32(sync->bitmap), + BIT(idx + 1) - 1, idx + 1); + err = -EINVAL; + goto out; + } + } + } while (!__igt_timeout(end_time, NULL)); +out: + return dump_syncmap(sync, err); +} + +static int igt_syncmap_compact(void *arg) +{ + struct i915_syncmap *sync; + unsigned int idx, order; + int err; + + i915_syncmap_init(&sync); + + /* + * The syncmap are "space efficient" compressed radix trees - any + * branch with only one child is skipped and replaced by the child. + * + * If we construct a tree with ids that are neighbouring at a non-zero + * height, we form a join but each child of that join is directly a + * leaf holding the single id. + */ + for (order = SHIFT; order < 64; order += SHIFT) { + err = check_syncmap_free(&sync); + if (err) + goto out; + + /* Create neighbours in the parent */ + for (idx = 0; idx < KSYNCMAP; idx++) { + u64 context = idx * BIT_ULL(order) + idx; + + err = i915_syncmap_set(&sync, context, 0); + if (err) + goto out; + + if (sync->height) { + pr_err("Inserting context=%llx (order=%d, idx=%d) did not return leaf (height=%d, prefix=%llx\n", + context, order, idx, + sync->height, sync->prefix); + err = -EINVAL; + goto out; + } + } + + sync = sync->parent; + if (sync->parent) { + pr_err("Parent (join) of last leaf was not the sync!\n"); + err = -EINVAL; + goto out; + } + + if (sync->height != order) { + pr_err("Join does not have the expected height, found %d, expected %d\n", + sync->height, order); + err = -EINVAL; + goto out; + } + + if (sync->bitmap != BIT(KSYNCMAP) - 1) { + pr_err("Join is not full!, found %x (%d) expected %lx (%d)\n", + sync->bitmap, hweight32(sync->bitmap), + BIT(KSYNCMAP) - 1, KSYNCMAP); + err = -EINVAL; + goto out; + } + + /* Each of our children should be a leaf */ + for (idx = 0; idx < KSYNCMAP; idx++) { + struct i915_syncmap *leaf = __sync_child(sync)[idx]; + + if (leaf->height) { + pr_err("Child %d is a not leaf!\n", idx); + err = -EINVAL; + goto out; + } + + if (leaf->parent != sync) { + pr_err("Child %d is not attached to us!\n", + idx); + err = -EINVAL; + goto out; + } + + if (!is_power_of_2(leaf->bitmap)) { + pr_err("Child %d holds more than one id, found %x (%d)\n", + idx, leaf->bitmap, hweight32(leaf->bitmap)); + err = -EINVAL; + goto out; + } + + if (leaf->bitmap != BIT(idx)) { + pr_err("Child %d has wrong seqno idx, found %d, expected %d\n", + idx, ilog2(leaf->bitmap), idx); + err = -EINVAL; + goto out; + } + } + } +out: + return dump_syncmap(sync, err); +} + +static int igt_syncmap_random(void *arg) +{ + I915_RND_STATE(prng); + IGT_TIMEOUT(end_time); + struct i915_syncmap *sync; + unsigned long count, phase, i; + u32 seqno; + int err; + + i915_syncmap_init(&sync); + + /* + * Having tried to test the individual operations within i915_syncmap, + * run a smoketest exploring the entire u64 space with random + * insertions. + */ + + count = 0; + phase = jiffies + HZ/100 + 1; + do { + u64 context = i915_prandom_u64_state(&prng); + + err = i915_syncmap_set(&sync, context, 0); + if (err) + goto out; + + count++; + } while (!time_after(jiffies, phase)); + seqno = 0; + + phase = 0; + do { + I915_RND_STATE(ctx); + u32 last_seqno = seqno; + bool expect; + + seqno = prandom_u32_state(&prng); + expect = seqno_later(last_seqno, seqno); + + for (i = 0; i < count; i++) { + u64 context = i915_prandom_u64_state(&ctx); + + if (i915_syncmap_is_later(&sync, context, seqno) != expect) { + pr_err("context=%llu, last=%u this=%u did not match expectation (%d)\n", + context, last_seqno, seqno, expect); + err = -EINVAL; + goto out; + } + + err = i915_syncmap_set(&sync, context, seqno); + if (err) + goto out; + } + + phase++; + } while (!__igt_timeout(end_time, NULL)); + pr_debug("Completed %lu passes, each of %lu contexts\n", phase, count); +out: + return dump_syncmap(sync, err); +} + +int i915_syncmap_mock_selftests(void) +{ + static const struct i915_subtest tests[] = { + SUBTEST(igt_syncmap_init), + SUBTEST(igt_syncmap_one), + SUBTEST(igt_syncmap_join_above), + SUBTEST(igt_syncmap_join_below), + SUBTEST(igt_syncmap_neighbours), + SUBTEST(igt_syncmap_compact), + SUBTEST(igt_syncmap_random), + }; + + return i915_subtests(tests, NULL); +} diff --git a/drivers/gpu/drm/i915/selftests/mock_timeline.c b/drivers/gpu/drm/i915/selftests/mock_timeline.c new file mode 100644 index 000000000000..47b1f47c5812 --- /dev/null +++ b/drivers/gpu/drm/i915/selftests/mock_timeline.c @@ -0,0 +1,45 @@ +/* + * Copyright © 2017 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + */ + +#include "mock_timeline.h" + +struct intel_timeline *mock_timeline(u64 context) +{ + static struct lock_class_key class; + struct intel_timeline *tl; + + tl = kzalloc(sizeof(*tl), GFP_KERNEL); + if (!tl) + return NULL; + + __intel_timeline_init(tl, NULL, context, &class, "mock"); + + return tl; +} + +void mock_timeline_destroy(struct intel_timeline *tl) +{ + __intel_timeline_fini(tl); + kfree(tl); +} diff --git a/drivers/gpu/drm/i915/selftests/mock_timeline.h b/drivers/gpu/drm/i915/selftests/mock_timeline.h new file mode 100644 index 000000000000..c27ff4639b8b --- /dev/null +++ b/drivers/gpu/drm/i915/selftests/mock_timeline.h @@ -0,0 +1,33 @@ +/* + * Copyright © 2017 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + */ + +#ifndef __MOCK_TIMELINE__ +#define __MOCK_TIMELINE__ + +#include "../i915_gem_timeline.h" + +struct intel_timeline *mock_timeline(u64 context); +void mock_timeline_destroy(struct intel_timeline *tl); + +#endif /* !__MOCK_TIMELINE__ */