xref: /openbmc/linux/drivers/gpu/drm/i915/gt/intel_ring.c (revision 8938c48f)
1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2019 Intel Corporation
5  */
6 
7 #include "gem/i915_gem_object.h"
8 
9 #include "i915_drv.h"
10 #include "i915_vma.h"
11 #include "intel_engine.h"
12 #include "intel_gpu_commands.h"
13 #include "intel_ring.h"
14 #include "intel_timeline.h"
15 
16 unsigned int intel_ring_update_space(struct intel_ring *ring)
17 {
18 	unsigned int space;
19 
20 	space = __intel_ring_space(ring->head, ring->emit, ring->size);
21 
22 	ring->space = space;
23 	return space;
24 }
25 
26 void __intel_ring_pin(struct intel_ring *ring)
27 {
28 	GEM_BUG_ON(!atomic_read(&ring->pin_count));
29 	atomic_inc(&ring->pin_count);
30 }
31 
32 int intel_ring_pin(struct intel_ring *ring, struct i915_gem_ww_ctx *ww)
33 {
34 	struct i915_vma *vma = ring->vma;
35 	unsigned int flags;
36 	void *addr;
37 	int ret;
38 
39 	if (atomic_fetch_inc(&ring->pin_count))
40 		return 0;
41 
42 	/* Ring wraparound at offset 0 sometimes hangs. No idea why. */
43 	flags = PIN_OFFSET_BIAS | i915_ggtt_pin_bias(vma);
44 
45 	if (i915_gem_object_is_stolen(vma->obj))
46 		flags |= PIN_MAPPABLE;
47 	else
48 		flags |= PIN_HIGH;
49 
50 	ret = i915_ggtt_pin(vma, ww, 0, flags);
51 	if (unlikely(ret))
52 		goto err_unpin;
53 
54 	if (i915_vma_is_map_and_fenceable(vma))
55 		addr = (void __force *)i915_vma_pin_iomap(vma);
56 	else
57 		addr = i915_gem_object_pin_map(vma->obj,
58 					       i915_coherent_map_type(vma->vm->i915));
59 	if (IS_ERR(addr)) {
60 		ret = PTR_ERR(addr);
61 		goto err_ring;
62 	}
63 
64 	i915_vma_make_unshrinkable(vma);
65 
66 	/* Discard any unused bytes beyond that submitted to hw. */
67 	intel_ring_reset(ring, ring->emit);
68 
69 	ring->vaddr = addr;
70 	return 0;
71 
72 err_ring:
73 	i915_vma_unpin(vma);
74 err_unpin:
75 	atomic_dec(&ring->pin_count);
76 	return ret;
77 }
78 
79 void intel_ring_reset(struct intel_ring *ring, u32 tail)
80 {
81 	tail = intel_ring_wrap(ring, tail);
82 	ring->tail = tail;
83 	ring->head = tail;
84 	ring->emit = tail;
85 	intel_ring_update_space(ring);
86 }
87 
88 void intel_ring_unpin(struct intel_ring *ring)
89 {
90 	struct i915_vma *vma = ring->vma;
91 
92 	if (!atomic_dec_and_test(&ring->pin_count))
93 		return;
94 
95 	i915_vma_unset_ggtt_write(vma);
96 	if (i915_vma_is_map_and_fenceable(vma))
97 		i915_vma_unpin_iomap(vma);
98 	else
99 		i915_gem_object_unpin_map(vma->obj);
100 
101 	i915_vma_make_purgeable(vma);
102 	i915_vma_unpin(vma);
103 }
104 
105 static struct i915_vma *create_ring_vma(struct i915_ggtt *ggtt, int size)
106 {
107 	struct i915_address_space *vm = &ggtt->vm;
108 	struct drm_i915_private *i915 = vm->i915;
109 	struct drm_i915_gem_object *obj;
110 	struct i915_vma *vma;
111 
112 	obj = ERR_PTR(-ENODEV);
113 	if (i915_ggtt_has_aperture(ggtt))
114 		obj = i915_gem_object_create_stolen(i915, size);
115 	if (IS_ERR(obj))
116 		obj = i915_gem_object_create_internal(i915, size);
117 	if (IS_ERR(obj))
118 		return ERR_CAST(obj);
119 
120 	/*
121 	 * Mark ring buffers as read-only from GPU side (so no stray overwrites)
122 	 * if supported by the platform's GGTT.
123 	 */
124 	if (vm->has_read_only)
125 		i915_gem_object_set_readonly(obj);
126 
127 	vma = i915_vma_instance(obj, vm, NULL);
128 	if (IS_ERR(vma))
129 		goto err;
130 
131 	return vma;
132 
133 err:
134 	i915_gem_object_put(obj);
135 	return vma;
136 }
137 
138 struct intel_ring *
139 intel_engine_create_ring(struct intel_engine_cs *engine, int size)
140 {
141 	struct drm_i915_private *i915 = engine->i915;
142 	struct intel_ring *ring;
143 	struct i915_vma *vma;
144 
145 	GEM_BUG_ON(!is_power_of_2(size));
146 	GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
147 
148 	ring = kzalloc(sizeof(*ring), GFP_KERNEL);
149 	if (!ring)
150 		return ERR_PTR(-ENOMEM);
151 
152 	kref_init(&ring->ref);
153 	ring->size = size;
154 	ring->wrap = BITS_PER_TYPE(ring->size) - ilog2(size);
155 
156 	/*
157 	 * Workaround an erratum on the i830 which causes a hang if
158 	 * the TAIL pointer points to within the last 2 cachelines
159 	 * of the buffer.
160 	 */
161 	ring->effective_size = size;
162 	if (IS_I830(i915) || IS_I845G(i915))
163 		ring->effective_size -= 2 * CACHELINE_BYTES;
164 
165 	intel_ring_update_space(ring);
166 
167 	vma = create_ring_vma(engine->gt->ggtt, size);
168 	if (IS_ERR(vma)) {
169 		kfree(ring);
170 		return ERR_CAST(vma);
171 	}
172 	ring->vma = vma;
173 
174 	return ring;
175 }
176 
177 void intel_ring_free(struct kref *ref)
178 {
179 	struct intel_ring *ring = container_of(ref, typeof(*ring), ref);
180 
181 	i915_vma_put(ring->vma);
182 	kfree(ring);
183 }
184 
185 static noinline int
186 wait_for_space(struct intel_ring *ring,
187 	       struct intel_timeline *tl,
188 	       unsigned int bytes)
189 {
190 	struct i915_request *target;
191 	long timeout;
192 
193 	if (intel_ring_update_space(ring) >= bytes)
194 		return 0;
195 
196 	GEM_BUG_ON(list_empty(&tl->requests));
197 	list_for_each_entry(target, &tl->requests, link) {
198 		if (target->ring != ring)
199 			continue;
200 
201 		/* Would completion of this request free enough space? */
202 		if (bytes <= __intel_ring_space(target->postfix,
203 						ring->emit, ring->size))
204 			break;
205 	}
206 
207 	if (GEM_WARN_ON(&target->link == &tl->requests))
208 		return -ENOSPC;
209 
210 	timeout = i915_request_wait(target,
211 				    I915_WAIT_INTERRUPTIBLE,
212 				    MAX_SCHEDULE_TIMEOUT);
213 	if (timeout < 0)
214 		return timeout;
215 
216 	i915_request_retire_upto(target);
217 
218 	intel_ring_update_space(ring);
219 	GEM_BUG_ON(ring->space < bytes);
220 	return 0;
221 }
222 
223 u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords)
224 {
225 	struct intel_ring *ring = rq->ring;
226 	const unsigned int remain_usable = ring->effective_size - ring->emit;
227 	const unsigned int bytes = num_dwords * sizeof(u32);
228 	unsigned int need_wrap = 0;
229 	unsigned int total_bytes;
230 	u32 *cs;
231 
232 	/* Packets must be qword aligned. */
233 	GEM_BUG_ON(num_dwords & 1);
234 
235 	total_bytes = bytes + rq->reserved_space;
236 	GEM_BUG_ON(total_bytes > ring->effective_size);
237 
238 	if (unlikely(total_bytes > remain_usable)) {
239 		const int remain_actual = ring->size - ring->emit;
240 
241 		if (bytes > remain_usable) {
242 			/*
243 			 * Not enough space for the basic request. So need to
244 			 * flush out the remainder and then wait for
245 			 * base + reserved.
246 			 */
247 			total_bytes += remain_actual;
248 			need_wrap = remain_actual | 1;
249 		} else  {
250 			/*
251 			 * The base request will fit but the reserved space
252 			 * falls off the end. So we don't need an immediate
253 			 * wrap and only need to effectively wait for the
254 			 * reserved size from the start of ringbuffer.
255 			 */
256 			total_bytes = rq->reserved_space + remain_actual;
257 		}
258 	}
259 
260 	if (unlikely(total_bytes > ring->space)) {
261 		int ret;
262 
263 		/*
264 		 * Space is reserved in the ringbuffer for finalising the
265 		 * request, as that cannot be allowed to fail. During request
266 		 * finalisation, reserved_space is set to 0 to stop the
267 		 * overallocation and the assumption is that then we never need
268 		 * to wait (which has the risk of failing with EINTR).
269 		 *
270 		 * See also i915_request_alloc() and i915_request_add().
271 		 */
272 		GEM_BUG_ON(!rq->reserved_space);
273 
274 		ret = wait_for_space(ring,
275 				     i915_request_timeline(rq),
276 				     total_bytes);
277 		if (unlikely(ret))
278 			return ERR_PTR(ret);
279 	}
280 
281 	if (unlikely(need_wrap)) {
282 		need_wrap &= ~1;
283 		GEM_BUG_ON(need_wrap > ring->space);
284 		GEM_BUG_ON(ring->emit + need_wrap > ring->size);
285 		GEM_BUG_ON(!IS_ALIGNED(need_wrap, sizeof(u64)));
286 
287 		/* Fill the tail with MI_NOOP */
288 		memset64(ring->vaddr + ring->emit, 0, need_wrap / sizeof(u64));
289 		ring->space -= need_wrap;
290 		ring->emit = 0;
291 	}
292 
293 	GEM_BUG_ON(ring->emit > ring->size - bytes);
294 	GEM_BUG_ON(ring->space < bytes);
295 	cs = ring->vaddr + ring->emit;
296 	GEM_DEBUG_EXEC(memset32(cs, POISON_INUSE, bytes / sizeof(*cs)));
297 	ring->emit += bytes;
298 	ring->space -= bytes;
299 
300 	return cs;
301 }
302 
303 /* Align the ring tail to a cacheline boundary */
304 int intel_ring_cacheline_align(struct i915_request *rq)
305 {
306 	int num_dwords;
307 	void *cs;
308 
309 	num_dwords = (rq->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(u32);
310 	if (num_dwords == 0)
311 		return 0;
312 
313 	num_dwords = CACHELINE_DWORDS - num_dwords;
314 	GEM_BUG_ON(num_dwords & 1);
315 
316 	cs = intel_ring_begin(rq, num_dwords);
317 	if (IS_ERR(cs))
318 		return PTR_ERR(cs);
319 
320 	memset64(cs, (u64)MI_NOOP << 32 | MI_NOOP, num_dwords / 2);
321 	intel_ring_advance(rq, cs + num_dwords);
322 
323 	GEM_BUG_ON(rq->ring->emit & (CACHELINE_BYTES - 1));
324 	return 0;
325 }
326 
327 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
328 #include "selftest_ring.c"
329 #endif
330