xref: /openbmc/linux/drivers/gpu/drm/i915/gt/intel_ring.c (revision 890f0b0d)
1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2019 Intel Corporation
5  */
6 
7 #include "gem/i915_gem_object.h"
8 #include "i915_drv.h"
9 #include "i915_vma.h"
10 #include "intel_engine.h"
11 #include "intel_ring.h"
12 #include "intel_timeline.h"
13 
14 unsigned int intel_ring_update_space(struct intel_ring *ring)
15 {
16 	unsigned int space;
17 
18 	space = __intel_ring_space(ring->head, ring->emit, ring->size);
19 
20 	ring->space = space;
21 	return space;
22 }
23 
24 int intel_ring_pin(struct intel_ring *ring)
25 {
26 	struct i915_vma *vma = ring->vma;
27 	unsigned int flags;
28 	void *addr;
29 	int ret;
30 
31 	if (atomic_fetch_inc(&ring->pin_count))
32 		return 0;
33 
34 	flags = PIN_GLOBAL;
35 
36 	/* Ring wraparound at offset 0 sometimes hangs. No idea why. */
37 	flags |= PIN_OFFSET_BIAS | i915_ggtt_pin_bias(vma);
38 
39 	if (vma->obj->stolen)
40 		flags |= PIN_MAPPABLE;
41 	else
42 		flags |= PIN_HIGH;
43 
44 	ret = i915_vma_pin(vma, 0, 0, flags);
45 	if (unlikely(ret))
46 		goto err_unpin;
47 
48 	if (i915_vma_is_map_and_fenceable(vma))
49 		addr = (void __force *)i915_vma_pin_iomap(vma);
50 	else
51 		addr = i915_gem_object_pin_map(vma->obj,
52 					       i915_coherent_map_type(vma->vm->i915));
53 	if (IS_ERR(addr)) {
54 		ret = PTR_ERR(addr);
55 		goto err_ring;
56 	}
57 
58 	i915_vma_make_unshrinkable(vma);
59 
60 	/* Discard any unused bytes beyond that submitted to hw. */
61 	intel_ring_reset(ring, ring->emit);
62 
63 	ring->vaddr = addr;
64 	return 0;
65 
66 err_ring:
67 	i915_vma_unpin(vma);
68 err_unpin:
69 	atomic_dec(&ring->pin_count);
70 	return ret;
71 }
72 
73 void intel_ring_reset(struct intel_ring *ring, u32 tail)
74 {
75 	tail = intel_ring_wrap(ring, tail);
76 	ring->tail = tail;
77 	ring->head = tail;
78 	ring->emit = tail;
79 	intel_ring_update_space(ring);
80 }
81 
82 void intel_ring_unpin(struct intel_ring *ring)
83 {
84 	struct i915_vma *vma = ring->vma;
85 
86 	if (!atomic_dec_and_test(&ring->pin_count))
87 		return;
88 
89 	i915_vma_unset_ggtt_write(vma);
90 	if (i915_vma_is_map_and_fenceable(vma))
91 		i915_vma_unpin_iomap(vma);
92 	else
93 		i915_gem_object_unpin_map(vma->obj);
94 
95 	i915_vma_make_purgeable(vma);
96 	i915_vma_unpin(vma);
97 }
98 
99 static struct i915_vma *create_ring_vma(struct i915_ggtt *ggtt, int size)
100 {
101 	struct i915_address_space *vm = &ggtt->vm;
102 	struct drm_i915_private *i915 = vm->i915;
103 	struct drm_i915_gem_object *obj;
104 	struct i915_vma *vma;
105 
106 	obj = ERR_PTR(-ENODEV);
107 	if (i915_ggtt_has_aperture(ggtt))
108 		obj = i915_gem_object_create_stolen(i915, size);
109 	if (IS_ERR(obj))
110 		obj = i915_gem_object_create_internal(i915, size);
111 	if (IS_ERR(obj))
112 		return ERR_CAST(obj);
113 
114 	/*
115 	 * Mark ring buffers as read-only from GPU side (so no stray overwrites)
116 	 * if supported by the platform's GGTT.
117 	 */
118 	if (vm->has_read_only)
119 		i915_gem_object_set_readonly(obj);
120 
121 	vma = i915_vma_instance(obj, vm, NULL);
122 	if (IS_ERR(vma))
123 		goto err;
124 
125 	return vma;
126 
127 err:
128 	i915_gem_object_put(obj);
129 	return vma;
130 }
131 
132 struct intel_ring *
133 intel_engine_create_ring(struct intel_engine_cs *engine, int size)
134 {
135 	struct drm_i915_private *i915 = engine->i915;
136 	struct intel_ring *ring;
137 	struct i915_vma *vma;
138 
139 	GEM_BUG_ON(!is_power_of_2(size));
140 	GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
141 
142 	ring = kzalloc(sizeof(*ring), GFP_KERNEL);
143 	if (!ring)
144 		return ERR_PTR(-ENOMEM);
145 
146 	kref_init(&ring->ref);
147 	ring->size = size;
148 	ring->wrap = BITS_PER_TYPE(ring->size) - ilog2(size);
149 
150 	/*
151 	 * Workaround an erratum on the i830 which causes a hang if
152 	 * the TAIL pointer points to within the last 2 cachelines
153 	 * of the buffer.
154 	 */
155 	ring->effective_size = size;
156 	if (IS_I830(i915) || IS_I845G(i915))
157 		ring->effective_size -= 2 * CACHELINE_BYTES;
158 
159 	intel_ring_update_space(ring);
160 
161 	vma = create_ring_vma(engine->gt->ggtt, size);
162 	if (IS_ERR(vma)) {
163 		kfree(ring);
164 		return ERR_CAST(vma);
165 	}
166 	ring->vma = vma;
167 
168 	return ring;
169 }
170 
171 void intel_ring_free(struct kref *ref)
172 {
173 	struct intel_ring *ring = container_of(ref, typeof(*ring), ref);
174 
175 	i915_vma_put(ring->vma);
176 	kfree(ring);
177 }
178 
179 static noinline int
180 wait_for_space(struct intel_ring *ring,
181 	       struct intel_timeline *tl,
182 	       unsigned int bytes)
183 {
184 	struct i915_request *target;
185 	long timeout;
186 
187 	if (intel_ring_update_space(ring) >= bytes)
188 		return 0;
189 
190 	GEM_BUG_ON(list_empty(&tl->requests));
191 	list_for_each_entry(target, &tl->requests, link) {
192 		if (target->ring != ring)
193 			continue;
194 
195 		/* Would completion of this request free enough space? */
196 		if (bytes <= __intel_ring_space(target->postfix,
197 						ring->emit, ring->size))
198 			break;
199 	}
200 
201 	if (GEM_WARN_ON(&target->link == &tl->requests))
202 		return -ENOSPC;
203 
204 	timeout = i915_request_wait(target,
205 				    I915_WAIT_INTERRUPTIBLE,
206 				    MAX_SCHEDULE_TIMEOUT);
207 	if (timeout < 0)
208 		return timeout;
209 
210 	i915_request_retire_upto(target);
211 
212 	intel_ring_update_space(ring);
213 	GEM_BUG_ON(ring->space < bytes);
214 	return 0;
215 }
216 
217 u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords)
218 {
219 	struct intel_ring *ring = rq->ring;
220 	const unsigned int remain_usable = ring->effective_size - ring->emit;
221 	const unsigned int bytes = num_dwords * sizeof(u32);
222 	unsigned int need_wrap = 0;
223 	unsigned int total_bytes;
224 	u32 *cs;
225 
226 	/* Packets must be qword aligned. */
227 	GEM_BUG_ON(num_dwords & 1);
228 
229 	total_bytes = bytes + rq->reserved_space;
230 	GEM_BUG_ON(total_bytes > ring->effective_size);
231 
232 	if (unlikely(total_bytes > remain_usable)) {
233 		const int remain_actual = ring->size - ring->emit;
234 
235 		if (bytes > remain_usable) {
236 			/*
237 			 * Not enough space for the basic request. So need to
238 			 * flush out the remainder and then wait for
239 			 * base + reserved.
240 			 */
241 			total_bytes += remain_actual;
242 			need_wrap = remain_actual | 1;
243 		} else  {
244 			/*
245 			 * The base request will fit but the reserved space
246 			 * falls off the end. So we don't need an immediate
247 			 * wrap and only need to effectively wait for the
248 			 * reserved size from the start of ringbuffer.
249 			 */
250 			total_bytes = rq->reserved_space + remain_actual;
251 		}
252 	}
253 
254 	if (unlikely(total_bytes > ring->space)) {
255 		int ret;
256 
257 		/*
258 		 * Space is reserved in the ringbuffer for finalising the
259 		 * request, as that cannot be allowed to fail. During request
260 		 * finalisation, reserved_space is set to 0 to stop the
261 		 * overallocation and the assumption is that then we never need
262 		 * to wait (which has the risk of failing with EINTR).
263 		 *
264 		 * See also i915_request_alloc() and i915_request_add().
265 		 */
266 		GEM_BUG_ON(!rq->reserved_space);
267 
268 		ret = wait_for_space(ring,
269 				     i915_request_timeline(rq),
270 				     total_bytes);
271 		if (unlikely(ret))
272 			return ERR_PTR(ret);
273 	}
274 
275 	if (unlikely(need_wrap)) {
276 		need_wrap &= ~1;
277 		GEM_BUG_ON(need_wrap > ring->space);
278 		GEM_BUG_ON(ring->emit + need_wrap > ring->size);
279 		GEM_BUG_ON(!IS_ALIGNED(need_wrap, sizeof(u64)));
280 
281 		/* Fill the tail with MI_NOOP */
282 		memset64(ring->vaddr + ring->emit, 0, need_wrap / sizeof(u64));
283 		ring->space -= need_wrap;
284 		ring->emit = 0;
285 	}
286 
287 	GEM_BUG_ON(ring->emit > ring->size - bytes);
288 	GEM_BUG_ON(ring->space < bytes);
289 	cs = ring->vaddr + ring->emit;
290 	GEM_DEBUG_EXEC(memset32(cs, POISON_INUSE, bytes / sizeof(*cs)));
291 	ring->emit += bytes;
292 	ring->space -= bytes;
293 
294 	return cs;
295 }
296 
297 /* Align the ring tail to a cacheline boundary */
298 int intel_ring_cacheline_align(struct i915_request *rq)
299 {
300 	int num_dwords;
301 	void *cs;
302 
303 	num_dwords = (rq->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(u32);
304 	if (num_dwords == 0)
305 		return 0;
306 
307 	num_dwords = CACHELINE_DWORDS - num_dwords;
308 	GEM_BUG_ON(num_dwords & 1);
309 
310 	cs = intel_ring_begin(rq, num_dwords);
311 	if (IS_ERR(cs))
312 		return PTR_ERR(cs);
313 
314 	memset64(cs, (u64)MI_NOOP << 32 | MI_NOOP, num_dwords / 2);
315 	intel_ring_advance(rq, cs + num_dwords);
316 
317 	GEM_BUG_ON(rq->ring->emit & (CACHELINE_BYTES - 1));
318 	return 0;
319 }
320