1 /* 2 * SPDX-License-Identifier: MIT 3 * 4 * Copyright © 2014-2016 Intel Corporation 5 */ 6 7 #include <linux/dma-fence-array.h> 8 9 #include "gt/intel_engine.h" 10 11 #include "i915_gem_ioctls.h" 12 #include "i915_gem_object.h" 13 14 static __always_inline u32 __busy_read_flag(u16 id) 15 { 16 if (id == (u16)I915_ENGINE_CLASS_INVALID) 17 return 0xffff0000u; 18 19 GEM_BUG_ON(id >= 16); 20 return 0x10000u << id; 21 } 22 23 static __always_inline u32 __busy_write_id(u16 id) 24 { 25 /* 26 * The uABI guarantees an active writer is also amongst the read 27 * engines. This would be true if we accessed the activity tracking 28 * under the lock, but as we perform the lookup of the object and 29 * its activity locklessly we can not guarantee that the last_write 30 * being active implies that we have set the same engine flag from 31 * last_read - hence we always set both read and write busy for 32 * last_write. 33 */ 34 if (id == (u16)I915_ENGINE_CLASS_INVALID) 35 return 0xffffffffu; 36 37 return (id + 1) | __busy_read_flag(id); 38 } 39 40 static __always_inline unsigned int 41 __busy_set_if_active(struct dma_fence *fence, u32 (*flag)(u16 id)) 42 { 43 const struct i915_request *rq; 44 45 /* 46 * We have to check the current hw status of the fence as the uABI 47 * guarantees forward progress. We could rely on the idle worker 48 * to eventually flush us, but to minimise latency just ask the 49 * hardware. 50 * 51 * Note we only report on the status of native fences and we currently 52 * have two native fences: 53 * 54 * 1. A composite fence (dma_fence_array) constructed of i915 requests 55 * created during a parallel submission. In this case we deconstruct the 56 * composite fence into individual i915 requests and check the status of 57 * each request. 58 * 59 * 2. A single i915 request. 60 */ 61 if (dma_fence_is_array(fence)) { 62 struct dma_fence_array *array = to_dma_fence_array(fence); 63 struct dma_fence **child = array->fences; 64 unsigned int nchild = array->num_fences; 65 66 do { 67 struct dma_fence *current_fence = *child++; 68 69 /* Not an i915 fence, can't be busy per above */ 70 if (!dma_fence_is_i915(current_fence) || 71 !test_bit(I915_FENCE_FLAG_COMPOSITE, 72 ¤t_fence->flags)) { 73 return 0; 74 } 75 76 rq = to_request(current_fence); 77 if (!i915_request_completed(rq)) 78 return flag(rq->engine->uabi_class); 79 } while (--nchild); 80 81 /* All requests in array complete, not busy */ 82 return 0; 83 } else { 84 if (!dma_fence_is_i915(fence)) 85 return 0; 86 87 rq = to_request(fence); 88 if (i915_request_completed(rq)) 89 return 0; 90 91 /* Beware type-expansion follies! */ 92 BUILD_BUG_ON(!typecheck(u16, rq->engine->uabi_class)); 93 return flag(rq->engine->uabi_class); 94 } 95 } 96 97 static __always_inline unsigned int 98 busy_check_reader(struct dma_fence *fence) 99 { 100 return __busy_set_if_active(fence, __busy_read_flag); 101 } 102 103 static __always_inline unsigned int 104 busy_check_writer(struct dma_fence *fence) 105 { 106 if (!fence) 107 return 0; 108 109 return __busy_set_if_active(fence, __busy_write_id); 110 } 111 112 int 113 i915_gem_busy_ioctl(struct drm_device *dev, void *data, 114 struct drm_file *file) 115 { 116 struct drm_i915_gem_busy *args = data; 117 struct drm_i915_gem_object *obj; 118 struct dma_resv_list *list; 119 unsigned int seq; 120 int err; 121 122 err = -ENOENT; 123 rcu_read_lock(); 124 obj = i915_gem_object_lookup_rcu(file, args->handle); 125 if (!obj) 126 goto out; 127 128 /* 129 * A discrepancy here is that we do not report the status of 130 * non-i915 fences, i.e. even though we may report the object as idle, 131 * a call to set-domain may still stall waiting for foreign rendering. 132 * This also means that wait-ioctl may report an object as busy, 133 * where busy-ioctl considers it idle. 134 * 135 * We trade the ability to warn of foreign fences to report on which 136 * i915 engines are active for the object. 137 * 138 * Alternatively, we can trade that extra information on read/write 139 * activity with 140 * args->busy = 141 * !dma_resv_test_signaled(obj->resv, true); 142 * to report the overall busyness. This is what the wait-ioctl does. 143 * 144 */ 145 retry: 146 seq = raw_read_seqcount(&obj->base.resv->seq); 147 148 /* Translate the exclusive fence to the READ *and* WRITE engine */ 149 args->busy = busy_check_writer(dma_resv_excl_fence(obj->base.resv)); 150 151 /* Translate shared fences to READ set of engines */ 152 list = dma_resv_shared_list(obj->base.resv); 153 if (list) { 154 unsigned int shared_count = list->shared_count, i; 155 156 for (i = 0; i < shared_count; ++i) { 157 struct dma_fence *fence = 158 rcu_dereference(list->shared[i]); 159 160 args->busy |= busy_check_reader(fence); 161 } 162 } 163 164 if (args->busy && read_seqcount_retry(&obj->base.resv->seq, seq)) 165 goto retry; 166 167 err = 0; 168 out: 169 rcu_read_unlock(); 170 return err; 171 } 172