xref: /freebsd/sys/dev/drm2/ttm/ttm_memory.c (revision d4eeb02986980bf33dd56c41ceb9fc5f180c0d47)
1 /**************************************************************************
2  *
3  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
4  * All Rights Reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
25  *
26  **************************************************************************/
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include <dev/drm2/drmP.h>
32 #include <dev/drm2/ttm/ttm_memory.h>
33 #include <dev/drm2/ttm/ttm_module.h>
34 #include <dev/drm2/ttm/ttm_page_alloc.h>
35 
36 #define TTM_MEMORY_ALLOC_RETRIES 4
37 
38 struct ttm_mem_zone {
39 	u_int kobj_ref;
40 	struct ttm_mem_global *glob;
41 	const char *name;
42 	uint64_t zone_mem;
43 	uint64_t emer_mem;
44 	uint64_t max_mem;
45 	uint64_t swap_limit;
46 	uint64_t used_mem;
47 };
48 
49 MALLOC_DEFINE(M_TTM_ZONE, "ttm_zone", "TTM Zone");
50 
51 static void ttm_mem_zone_kobj_release(struct ttm_mem_zone *zone)
52 {
53 
54 	printf("[TTM] Zone %7s: Used memory at exit: %llu kiB\n",
55 		zone->name, (unsigned long long)zone->used_mem >> 10);
56 	free(zone, M_TTM_ZONE);
57 }
58 
59 #if 0
60 /* XXXKIB sysctl */
61 static ssize_t ttm_mem_zone_show(struct ttm_mem_zone *zone;
62 				 struct attribute *attr,
63 				 char *buffer)
64 {
65 	uint64_t val = 0;
66 
67 	mtx_lock(&zone->glob->lock);
68 	if (attr == &ttm_mem_sys)
69 		val = zone->zone_mem;
70 	else if (attr == &ttm_mem_emer)
71 		val = zone->emer_mem;
72 	else if (attr == &ttm_mem_max)
73 		val = zone->max_mem;
74 	else if (attr == &ttm_mem_swap)
75 		val = zone->swap_limit;
76 	else if (attr == &ttm_mem_used)
77 		val = zone->used_mem;
78 	mtx_unlock(&zone->glob->lock);
79 
80 	return snprintf(buffer, PAGE_SIZE, "%llu\n",
81 			(unsigned long long) val >> 10);
82 }
83 #endif
84 
85 static void ttm_check_swapping(struct ttm_mem_global *glob);
86 
87 #if 0
88 /* XXXKIB sysctl */
89 static ssize_t ttm_mem_zone_store(struct ttm_mem_zone *zone,
90 				  struct attribute *attr,
91 				  const char *buffer,
92 				  size_t size)
93 {
94 	int chars;
95 	unsigned long val;
96 	uint64_t val64;
97 
98 	chars = sscanf(buffer, "%lu", &val);
99 	if (chars == 0)
100 		return size;
101 
102 	val64 = val;
103 	val64 <<= 10;
104 
105 	mtx_lock(&zone->glob->lock);
106 	if (val64 > zone->zone_mem)
107 		val64 = zone->zone_mem;
108 	if (attr == &ttm_mem_emer) {
109 		zone->emer_mem = val64;
110 		if (zone->max_mem > val64)
111 			zone->max_mem = val64;
112 	} else if (attr == &ttm_mem_max) {
113 		zone->max_mem = val64;
114 		if (zone->emer_mem < val64)
115 			zone->emer_mem = val64;
116 	} else if (attr == &ttm_mem_swap)
117 		zone->swap_limit = val64;
118 	mtx_unlock(&zone->glob->lock);
119 
120 	ttm_check_swapping(zone->glob);
121 
122 	return size;
123 }
124 #endif
125 
126 static void ttm_mem_global_kobj_release(struct ttm_mem_global *glob)
127 {
128 }
129 
130 static bool ttm_zones_above_swap_target(struct ttm_mem_global *glob,
131 					bool from_wq, uint64_t extra)
132 {
133 	unsigned int i;
134 	struct ttm_mem_zone *zone;
135 	uint64_t target;
136 
137 	for (i = 0; i < glob->num_zones; ++i) {
138 		zone = glob->zones[i];
139 
140 		if (from_wq)
141 			target = zone->swap_limit;
142 		else if (priv_check(curthread, PRIV_VM_MLOCK) == 0)
143 			target = zone->emer_mem;
144 		else
145 			target = zone->max_mem;
146 
147 		target = (extra > target) ? 0ULL : target;
148 
149 		if (zone->used_mem > target)
150 			return true;
151 	}
152 	return false;
153 }
154 
155 /**
156  * At this point we only support a single shrink callback.
157  * Extend this if needed, perhaps using a linked list of callbacks.
158  * Note that this function is reentrant:
159  * many threads may try to swap out at any given time.
160  */
161 
162 static void ttm_shrink(struct ttm_mem_global *glob, bool from_wq,
163 		       uint64_t extra)
164 {
165 	int ret;
166 	struct ttm_mem_shrink *shrink;
167 
168 	mtx_lock(&glob->lock);
169 	if (glob->shrink == NULL)
170 		goto out;
171 
172 	while (ttm_zones_above_swap_target(glob, from_wq, extra)) {
173 		shrink = glob->shrink;
174 		mtx_unlock(&glob->lock);
175 		ret = shrink->do_shrink(shrink);
176 		mtx_lock(&glob->lock);
177 		if (unlikely(ret != 0))
178 			goto out;
179 	}
180 out:
181 	mtx_unlock(&glob->lock);
182 }
183 
184 
185 
186 static void ttm_shrink_work(void *arg, int pending __unused)
187 {
188 	struct ttm_mem_global *glob = arg;
189 
190 	ttm_shrink(glob, true, 0ULL);
191 }
192 
193 static int ttm_mem_init_kernel_zone(struct ttm_mem_global *glob,
194     uint64_t mem)
195 {
196 	struct ttm_mem_zone *zone;
197 
198 	zone = malloc(sizeof(*zone), M_TTM_ZONE, M_WAITOK | M_ZERO);
199 
200 	zone->name = "kernel";
201 	zone->zone_mem = mem;
202 	zone->max_mem = mem >> 1;
203 	zone->emer_mem = (mem >> 1) + (mem >> 2);
204 	zone->swap_limit = zone->max_mem - (mem >> 3);
205 	zone->used_mem = 0;
206 	zone->glob = glob;
207 	glob->zone_kernel = zone;
208 	refcount_init(&zone->kobj_ref, 1);
209 	glob->zones[glob->num_zones++] = zone;
210 	return 0;
211 }
212 
213 static int ttm_mem_init_dma32_zone(struct ttm_mem_global *glob,
214     uint64_t mem)
215 {
216 	struct ttm_mem_zone *zone;
217 
218 	zone = malloc(sizeof(*zone), M_TTM_ZONE, M_WAITOK | M_ZERO);
219 
220 	/**
221 	 * No special dma32 zone needed.
222 	 */
223 
224 	if (mem <= ((uint64_t) 1ULL << 32)) {
225 		free(zone, M_TTM_ZONE);
226 		return 0;
227 	}
228 
229 	/*
230 	 * Limit max dma32 memory to 4GB for now
231 	 * until we can figure out how big this
232 	 * zone really is.
233 	 */
234 
235 	mem = ((uint64_t) 1ULL << 32);
236 	zone->name = "dma32";
237 	zone->zone_mem = mem;
238 	zone->max_mem = mem >> 1;
239 	zone->emer_mem = (mem >> 1) + (mem >> 2);
240 	zone->swap_limit = zone->max_mem - (mem >> 3);
241 	zone->used_mem = 0;
242 	zone->glob = glob;
243 	glob->zone_dma32 = zone;
244 	refcount_init(&zone->kobj_ref, 1);
245 	glob->zones[glob->num_zones++] = zone;
246 	return 0;
247 }
248 
249 int ttm_mem_global_init(struct ttm_mem_global *glob)
250 {
251 	u_int64_t mem;
252 	int ret;
253 	int i;
254 	struct ttm_mem_zone *zone;
255 
256 	mtx_init(&glob->lock, "ttmgz", NULL, MTX_DEF);
257 	glob->swap_queue = taskqueue_create("ttm_swap", M_WAITOK,
258 	    taskqueue_thread_enqueue, &glob->swap_queue);
259 	taskqueue_start_threads(&glob->swap_queue, 1, PVM, "ttm swap");
260 	TASK_INIT(&glob->work, 0, ttm_shrink_work, glob);
261 
262 	refcount_init(&glob->kobj_ref, 1);
263 
264 	mem = physmem * PAGE_SIZE;
265 
266 	ret = ttm_mem_init_kernel_zone(glob, mem);
267 	if (unlikely(ret != 0))
268 		goto out_no_zone;
269 	ret = ttm_mem_init_dma32_zone(glob, mem);
270 	if (unlikely(ret != 0))
271 		goto out_no_zone;
272 	for (i = 0; i < glob->num_zones; ++i) {
273 		zone = glob->zones[i];
274 		printf("[TTM] Zone %7s: Available graphics memory: %llu kiB\n",
275 			zone->name, (unsigned long long)zone->max_mem >> 10);
276 	}
277 	ttm_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
278 	ttm_dma_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
279 	return 0;
280 out_no_zone:
281 	ttm_mem_global_release(glob);
282 	return ret;
283 }
284 
285 void ttm_mem_global_release(struct ttm_mem_global *glob)
286 {
287 	unsigned int i;
288 	struct ttm_mem_zone *zone;
289 
290 	/* let the page allocator first stop the shrink work. */
291 	ttm_page_alloc_fini();
292 	ttm_dma_page_alloc_fini();
293 
294 	taskqueue_drain(glob->swap_queue, &glob->work);
295 	taskqueue_free(glob->swap_queue);
296 	glob->swap_queue = NULL;
297 	for (i = 0; i < glob->num_zones; ++i) {
298 		zone = glob->zones[i];
299 		if (refcount_release(&zone->kobj_ref))
300 			ttm_mem_zone_kobj_release(zone);
301 	}
302 	if (refcount_release(&glob->kobj_ref))
303 		ttm_mem_global_kobj_release(glob);
304 }
305 
306 static void ttm_check_swapping(struct ttm_mem_global *glob)
307 {
308 	bool needs_swapping = false;
309 	unsigned int i;
310 	struct ttm_mem_zone *zone;
311 
312 	mtx_lock(&glob->lock);
313 	for (i = 0; i < glob->num_zones; ++i) {
314 		zone = glob->zones[i];
315 		if (zone->used_mem > zone->swap_limit) {
316 			needs_swapping = true;
317 			break;
318 		}
319 	}
320 
321 	mtx_unlock(&glob->lock);
322 
323 	if (unlikely(needs_swapping))
324 		taskqueue_enqueue(glob->swap_queue, &glob->work);
325 
326 }
327 
328 static void ttm_mem_global_free_zone(struct ttm_mem_global *glob,
329 				     struct ttm_mem_zone *single_zone,
330 				     uint64_t amount)
331 {
332 	unsigned int i;
333 	struct ttm_mem_zone *zone;
334 
335 	mtx_lock(&glob->lock);
336 	for (i = 0; i < glob->num_zones; ++i) {
337 		zone = glob->zones[i];
338 		if (single_zone && zone != single_zone)
339 			continue;
340 		zone->used_mem -= amount;
341 	}
342 	mtx_unlock(&glob->lock);
343 }
344 
345 void ttm_mem_global_free(struct ttm_mem_global *glob,
346 			 uint64_t amount)
347 {
348 	return ttm_mem_global_free_zone(glob, NULL, amount);
349 }
350 
351 static int ttm_mem_global_reserve(struct ttm_mem_global *glob,
352 				  struct ttm_mem_zone *single_zone,
353 				  uint64_t amount, bool reserve)
354 {
355 	uint64_t limit;
356 	int ret = -ENOMEM;
357 	unsigned int i;
358 	struct ttm_mem_zone *zone;
359 
360 	mtx_lock(&glob->lock);
361 	for (i = 0; i < glob->num_zones; ++i) {
362 		zone = glob->zones[i];
363 		if (single_zone && zone != single_zone)
364 			continue;
365 
366 		limit = (priv_check(curthread, PRIV_VM_MLOCK) == 0) ?
367 			zone->emer_mem : zone->max_mem;
368 
369 		if (zone->used_mem > limit)
370 			goto out_unlock;
371 	}
372 
373 	if (reserve) {
374 		for (i = 0; i < glob->num_zones; ++i) {
375 			zone = glob->zones[i];
376 			if (single_zone && zone != single_zone)
377 				continue;
378 			zone->used_mem += amount;
379 		}
380 	}
381 
382 	ret = 0;
383 out_unlock:
384 	mtx_unlock(&glob->lock);
385 	ttm_check_swapping(glob);
386 
387 	return ret;
388 }
389 
390 
391 static int ttm_mem_global_alloc_zone(struct ttm_mem_global *glob,
392 				     struct ttm_mem_zone *single_zone,
393 				     uint64_t memory,
394 				     bool no_wait, bool interruptible)
395 {
396 	int count = TTM_MEMORY_ALLOC_RETRIES;
397 
398 	while (unlikely(ttm_mem_global_reserve(glob,
399 					       single_zone,
400 					       memory, true)
401 			!= 0)) {
402 		if (no_wait)
403 			return -ENOMEM;
404 		if (unlikely(count-- == 0))
405 			return -ENOMEM;
406 		ttm_shrink(glob, false, memory + (memory >> 2) + 16);
407 	}
408 
409 	return 0;
410 }
411 
412 int ttm_mem_global_alloc(struct ttm_mem_global *glob, uint64_t memory,
413 			 bool no_wait, bool interruptible)
414 {
415 	/**
416 	 * Normal allocations of kernel memory are registered in
417 	 * all zones.
418 	 */
419 
420 	return ttm_mem_global_alloc_zone(glob, NULL, memory, no_wait,
421 					 interruptible);
422 }
423 
424 #define page_to_pfn(pp) OFF_TO_IDX(VM_PAGE_TO_PHYS(pp))
425 
426 int ttm_mem_global_alloc_page(struct ttm_mem_global *glob,
427 			      struct vm_page *page,
428 			      bool no_wait, bool interruptible)
429 {
430 
431 	struct ttm_mem_zone *zone = NULL;
432 
433 	/**
434 	 * Page allocations may be registed in a single zone
435 	 * only if highmem or !dma32.
436 	 */
437 
438 	if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
439 		zone = glob->zone_kernel;
440 	return ttm_mem_global_alloc_zone(glob, zone, PAGE_SIZE, no_wait,
441 					 interruptible);
442 }
443 
444 void ttm_mem_global_free_page(struct ttm_mem_global *glob, struct vm_page *page)
445 {
446 	struct ttm_mem_zone *zone = NULL;
447 
448 	if (glob->zone_dma32 && page_to_pfn(page) > 0x00100000UL)
449 		zone = glob->zone_kernel;
450 	ttm_mem_global_free_zone(glob, zone, PAGE_SIZE);
451 }
452 
453 
454 size_t ttm_round_pot(size_t size)
455 {
456 	if ((size & (size - 1)) == 0)
457 		return size;
458 	else if (size > PAGE_SIZE)
459 		return PAGE_ALIGN(size);
460 	else {
461 		size_t tmp_size = 4;
462 
463 		while (tmp_size < size)
464 			tmp_size <<= 1;
465 
466 		return tmp_size;
467 	}
468 	return 0;
469 }
470