1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
3  *
4  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
5  * All Rights Reserved.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a
8  * copy of this software and associated documentation files (the
9  * "Software"), to deal in the Software without restriction, including
10  * without limitation the rights to use, copy, modify, merge, publish,
11  * distribute, sub license, and/or sell copies of the Software, and to
12  * permit persons to whom the Software is furnished to do so, subject to
13  * the following conditions:
14  *
15  * The above copyright notice and this permission notice (including the
16  * next paragraph) shall be included in all copies or substantial portions
17  * of the Software.
18  *
19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25  * USE OR OTHER DEALINGS IN THE SOFTWARE.
26  *
27  **************************************************************************/
28 /*
29  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30  */
31 
32 #define pr_fmt(fmt) "[TTM] " fmt
33 
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <drm/drm_vma_manager.h>
37 #include <drm/drm_drv.h>
38 #include <drm/drm_managed.h>
39 #include <linux/mm.h>
40 #include <linux/pfn_t.h>
41 #include <linux/rbtree.h>
42 #include <linux/module.h>
43 #include <linux/uaccess.h>
44 #include <linux/mem_encrypt.h>
45 
ttm_bo_vm_fault_idle(struct ttm_buffer_object * bo,struct vm_fault * vmf)46 static vm_fault_t ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
47 				struct vm_fault *vmf)
48 {
49 	long err = 0;
50 
51 	/*
52 	 * Quick non-stalling check for idle.
53 	 */
54 	if (dma_resv_test_signaled(bo->base.resv, DMA_RESV_USAGE_KERNEL))
55 		return 0;
56 
57 	/*
58 	 * If possible, avoid waiting for GPU with mmap_lock
59 	 * held.  We only do this if the fault allows retry and this
60 	 * is the first attempt.
61 	 */
62 	if (fault_flag_allow_retry_first(vmf->flags)) {
63 		if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
64 			return VM_FAULT_RETRY;
65 
66 		ttm_bo_get(bo);
67 		mmap_read_unlock(vmf->vma->vm_mm);
68 		(void)dma_resv_wait_timeout(bo->base.resv,
69 					    DMA_RESV_USAGE_KERNEL, true,
70 					    MAX_SCHEDULE_TIMEOUT);
71 		dma_resv_unlock(bo->base.resv);
72 		ttm_bo_put(bo);
73 		return VM_FAULT_RETRY;
74 	}
75 
76 	/*
77 	 * Ordinary wait.
78 	 */
79 	err = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_KERNEL, true,
80 				    MAX_SCHEDULE_TIMEOUT);
81 	if (unlikely(err < 0)) {
82 		return (err != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
83 			VM_FAULT_NOPAGE;
84 	}
85 
86 	return 0;
87 }
88 
ttm_bo_io_mem_pfn(struct ttm_buffer_object * bo,unsigned long page_offset)89 static unsigned long ttm_bo_io_mem_pfn(struct ttm_buffer_object *bo,
90 				       unsigned long page_offset)
91 {
92 	struct ttm_device *bdev = bo->bdev;
93 
94 	if (bdev->funcs->io_mem_pfn)
95 		return bdev->funcs->io_mem_pfn(bo, page_offset);
96 
97 	return (bo->resource->bus.offset >> PAGE_SHIFT) + page_offset;
98 }
99 
100 /**
101  * ttm_bo_vm_reserve - Reserve a buffer object in a retryable vm callback
102  * @bo: The buffer object
103  * @vmf: The fault structure handed to the callback
104  *
105  * vm callbacks like fault() and *_mkwrite() allow for the mmap_lock to be dropped
106  * during long waits, and after the wait the callback will be restarted. This
107  * is to allow other threads using the same virtual memory space concurrent
108  * access to map(), unmap() completely unrelated buffer objects. TTM buffer
109  * object reservations sometimes wait for GPU and should therefore be
110  * considered long waits. This function reserves the buffer object interruptibly
111  * taking this into account. Starvation is avoided by the vm system not
112  * allowing too many repeated restarts.
113  * This function is intended to be used in customized fault() and _mkwrite()
114  * handlers.
115  *
116  * Return:
117  *    0 on success and the bo was reserved.
118  *    VM_FAULT_RETRY if blocking wait.
119  *    VM_FAULT_NOPAGE if blocking wait and retrying was not allowed.
120  */
ttm_bo_vm_reserve(struct ttm_buffer_object * bo,struct vm_fault * vmf)121 vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
122 			     struct vm_fault *vmf)
123 {
124 	/*
125 	 * Work around locking order reversal in fault / nopfn
126 	 * between mmap_lock and bo_reserve: Perform a trylock operation
127 	 * for reserve, and if it fails, retry the fault after waiting
128 	 * for the buffer to become unreserved.
129 	 */
130 	if (unlikely(!dma_resv_trylock(bo->base.resv))) {
131 		/*
132 		 * If the fault allows retry and this is the first
133 		 * fault attempt, we try to release the mmap_lock
134 		 * before waiting
135 		 */
136 		if (fault_flag_allow_retry_first(vmf->flags)) {
137 			if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
138 				ttm_bo_get(bo);
139 				mmap_read_unlock(vmf->vma->vm_mm);
140 				if (!dma_resv_lock_interruptible(bo->base.resv,
141 								 NULL))
142 					dma_resv_unlock(bo->base.resv);
143 				ttm_bo_put(bo);
144 			}
145 
146 			return VM_FAULT_RETRY;
147 		}
148 
149 		if (dma_resv_lock_interruptible(bo->base.resv, NULL))
150 			return VM_FAULT_NOPAGE;
151 	}
152 
153 	/*
154 	 * Refuse to fault imported pages. This should be handled
155 	 * (if at all) by redirecting mmap to the exporter.
156 	 */
157 	if (bo->ttm && (bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL)) {
158 		if (!(bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL_MAPPABLE)) {
159 			dma_resv_unlock(bo->base.resv);
160 			return VM_FAULT_SIGBUS;
161 		}
162 	}
163 
164 	return 0;
165 }
166 EXPORT_SYMBOL(ttm_bo_vm_reserve);
167 
168 /**
169  * ttm_bo_vm_fault_reserved - TTM fault helper
170  * @vmf: The struct vm_fault given as argument to the fault callback
171  * @prot: The page protection to be used for this memory area.
172  * @num_prefault: Maximum number of prefault pages. The caller may want to
173  * specify this based on madvice settings and the size of the GPU object
174  * backed by the memory.
175  *
176  * This function inserts one or more page table entries pointing to the
177  * memory backing the buffer object, and then returns a return code
178  * instructing the caller to retry the page access.
179  *
180  * Return:
181  *   VM_FAULT_NOPAGE on success or pending signal
182  *   VM_FAULT_SIGBUS on unspecified error
183  *   VM_FAULT_OOM on out-of-memory
184  *   VM_FAULT_RETRY if retryable wait
185  */
ttm_bo_vm_fault_reserved(struct vm_fault * vmf,pgprot_t prot,pgoff_t num_prefault)186 vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
187 				    pgprot_t prot,
188 				    pgoff_t num_prefault)
189 {
190 	struct vm_area_struct *vma = vmf->vma;
191 	struct ttm_buffer_object *bo = vma->vm_private_data;
192 	struct ttm_device *bdev = bo->bdev;
193 	unsigned long page_offset;
194 	unsigned long page_last;
195 	unsigned long pfn;
196 	struct ttm_tt *ttm = NULL;
197 	struct page *page;
198 	int err;
199 	pgoff_t i;
200 	vm_fault_t ret = VM_FAULT_NOPAGE;
201 	unsigned long address = vmf->address;
202 
203 	/*
204 	 * Wait for buffer data in transit, due to a pipelined
205 	 * move.
206 	 */
207 	ret = ttm_bo_vm_fault_idle(bo, vmf);
208 	if (unlikely(ret != 0))
209 		return ret;
210 
211 	err = ttm_mem_io_reserve(bdev, bo->resource);
212 	if (unlikely(err != 0))
213 		return VM_FAULT_SIGBUS;
214 
215 	page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
216 		vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node);
217 	page_last = vma_pages(vma) + vma->vm_pgoff -
218 		drm_vma_node_start(&bo->base.vma_node);
219 
220 	if (unlikely(page_offset >= bo->resource->num_pages))
221 		return VM_FAULT_SIGBUS;
222 
223 	prot = ttm_io_prot(bo, bo->resource, prot);
224 	if (!bo->resource->bus.is_iomem) {
225 		struct ttm_operation_ctx ctx = {
226 			.interruptible = false,
227 			.no_wait_gpu = false,
228 			.force_alloc = true
229 		};
230 
231 		ttm = bo->ttm;
232 		if (ttm_tt_populate(bdev, bo->ttm, &ctx))
233 			return VM_FAULT_OOM;
234 	} else {
235 		/* Iomem should not be marked encrypted */
236 		prot = pgprot_decrypted(prot);
237 	}
238 
239 	/*
240 	 * Speculatively prefault a number of pages. Only error on
241 	 * first page.
242 	 */
243 	for (i = 0; i < num_prefault; ++i) {
244 		if (bo->resource->bus.is_iomem) {
245 			pfn = ttm_bo_io_mem_pfn(bo, page_offset);
246 		} else {
247 			page = ttm->pages[page_offset];
248 			if (unlikely(!page && i == 0)) {
249 				return VM_FAULT_OOM;
250 			} else if (unlikely(!page)) {
251 				break;
252 			}
253 			pfn = page_to_pfn(page);
254 		}
255 
256 		/*
257 		 * Note that the value of @prot at this point may differ from
258 		 * the value of @vma->vm_page_prot in the caching- and
259 		 * encryption bits. This is because the exact location of the
260 		 * data may not be known at mmap() time and may also change
261 		 * at arbitrary times while the data is mmap'ed.
262 		 * See vmf_insert_mixed_prot() for a discussion.
263 		 */
264 		ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
265 
266 		/* Never error on prefaulted PTEs */
267 		if (unlikely((ret & VM_FAULT_ERROR))) {
268 			if (i == 0)
269 				return VM_FAULT_NOPAGE;
270 			else
271 				break;
272 		}
273 
274 		address += PAGE_SIZE;
275 		if (unlikely(++page_offset >= page_last))
276 			break;
277 	}
278 	return ret;
279 }
280 EXPORT_SYMBOL(ttm_bo_vm_fault_reserved);
281 
ttm_bo_release_dummy_page(struct drm_device * dev,void * res)282 static void ttm_bo_release_dummy_page(struct drm_device *dev, void *res)
283 {
284 	struct page *dummy_page = (struct page *)res;
285 
286 	__free_page(dummy_page);
287 }
288 
ttm_bo_vm_dummy_page(struct vm_fault * vmf,pgprot_t prot)289 vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot)
290 {
291 	struct vm_area_struct *vma = vmf->vma;
292 	struct ttm_buffer_object *bo = vma->vm_private_data;
293 	struct drm_device *ddev = bo->base.dev;
294 	vm_fault_t ret = VM_FAULT_NOPAGE;
295 	unsigned long address;
296 	unsigned long pfn;
297 	struct page *page;
298 
299 	/* Allocate new dummy page to map all the VA range in this VMA to it*/
300 	page = alloc_page(GFP_KERNEL | __GFP_ZERO);
301 	if (!page)
302 		return VM_FAULT_OOM;
303 
304 	/* Set the page to be freed using drmm release action */
305 	if (drmm_add_action_or_reset(ddev, ttm_bo_release_dummy_page, page))
306 		return VM_FAULT_OOM;
307 
308 	pfn = page_to_pfn(page);
309 
310 	/* Prefault the entire VMA range right away to avoid further faults */
311 	for (address = vma->vm_start; address < vma->vm_end;
312 	     address += PAGE_SIZE)
313 		ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
314 
315 	return ret;
316 }
317 EXPORT_SYMBOL(ttm_bo_vm_dummy_page);
318 
ttm_bo_vm_fault(struct vm_fault * vmf)319 vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
320 {
321 	struct vm_area_struct *vma = vmf->vma;
322 	pgprot_t prot;
323 	struct ttm_buffer_object *bo = vma->vm_private_data;
324 	struct drm_device *ddev = bo->base.dev;
325 	vm_fault_t ret;
326 	int idx;
327 
328 	ret = ttm_bo_vm_reserve(bo, vmf);
329 	if (ret)
330 		return ret;
331 
332 	prot = vma->vm_page_prot;
333 	if (drm_dev_enter(ddev, &idx)) {
334 		ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT);
335 		drm_dev_exit(idx);
336 	} else {
337 		ret = ttm_bo_vm_dummy_page(vmf, prot);
338 	}
339 	if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
340 		return ret;
341 
342 	dma_resv_unlock(bo->base.resv);
343 
344 	return ret;
345 }
346 EXPORT_SYMBOL(ttm_bo_vm_fault);
347 
ttm_bo_vm_open(struct vm_area_struct * vma)348 void ttm_bo_vm_open(struct vm_area_struct *vma)
349 {
350 	struct ttm_buffer_object *bo = vma->vm_private_data;
351 
352 	WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
353 
354 	ttm_bo_get(bo);
355 }
356 EXPORT_SYMBOL(ttm_bo_vm_open);
357 
ttm_bo_vm_close(struct vm_area_struct * vma)358 void ttm_bo_vm_close(struct vm_area_struct *vma)
359 {
360 	struct ttm_buffer_object *bo = vma->vm_private_data;
361 
362 	ttm_bo_put(bo);
363 	vma->vm_private_data = NULL;
364 }
365 EXPORT_SYMBOL(ttm_bo_vm_close);
366 
ttm_bo_vm_access_kmap(struct ttm_buffer_object * bo,unsigned long offset,uint8_t * buf,int len,int write)367 static int ttm_bo_vm_access_kmap(struct ttm_buffer_object *bo,
368 				 unsigned long offset,
369 				 uint8_t *buf, int len, int write)
370 {
371 	unsigned long page = offset >> PAGE_SHIFT;
372 	unsigned long bytes_left = len;
373 	int ret;
374 
375 	/* Copy a page at a time, that way no extra virtual address
376 	 * mapping is needed
377 	 */
378 	offset -= page << PAGE_SHIFT;
379 	do {
380 		unsigned long bytes = min(bytes_left, PAGE_SIZE - offset);
381 		struct ttm_bo_kmap_obj map;
382 		void *ptr;
383 		bool is_iomem;
384 
385 		ret = ttm_bo_kmap(bo, page, 1, &map);
386 		if (ret)
387 			return ret;
388 
389 		ptr = (uint8_t *)ttm_kmap_obj_virtual(&map, &is_iomem) + offset;
390 		WARN_ON_ONCE(is_iomem);
391 		if (write)
392 			memcpy(ptr, buf, bytes);
393 		else
394 			memcpy(buf, ptr, bytes);
395 		ttm_bo_kunmap(&map);
396 
397 		page++;
398 		buf += bytes;
399 		bytes_left -= bytes;
400 		offset = 0;
401 	} while (bytes_left);
402 
403 	return len;
404 }
405 
ttm_bo_vm_access(struct vm_area_struct * vma,unsigned long addr,void * buf,int len,int write)406 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
407 		     void *buf, int len, int write)
408 {
409 	struct ttm_buffer_object *bo = vma->vm_private_data;
410 	unsigned long offset = (addr) - vma->vm_start +
411 		((vma->vm_pgoff - drm_vma_node_start(&bo->base.vma_node))
412 		 << PAGE_SHIFT);
413 	int ret;
414 
415 	if (len < 1 || (offset + len) >> PAGE_SHIFT > bo->resource->num_pages)
416 		return -EIO;
417 
418 	ret = ttm_bo_reserve(bo, true, false, NULL);
419 	if (ret)
420 		return ret;
421 
422 	switch (bo->resource->mem_type) {
423 	case TTM_PL_SYSTEM:
424 		fallthrough;
425 	case TTM_PL_TT:
426 		ret = ttm_bo_vm_access_kmap(bo, offset, buf, len, write);
427 		break;
428 	default:
429 		if (bo->bdev->funcs->access_memory)
430 			ret = bo->bdev->funcs->access_memory(
431 				bo, offset, buf, len, write);
432 		else
433 			ret = -EIO;
434 	}
435 
436 	ttm_bo_unreserve(bo);
437 
438 	return ret;
439 }
440 EXPORT_SYMBOL(ttm_bo_vm_access);
441 
442 static const struct vm_operations_struct ttm_bo_vm_ops = {
443 	.fault = ttm_bo_vm_fault,
444 	.open = ttm_bo_vm_open,
445 	.close = ttm_bo_vm_close,
446 	.access = ttm_bo_vm_access,
447 };
448 
ttm_bo_mmap_obj(struct vm_area_struct * vma,struct ttm_buffer_object * bo)449 int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
450 {
451 	/* Enforce no COW since would have really strange behavior with it. */
452 	if (is_cow_mapping(vma->vm_flags))
453 		return -EINVAL;
454 
455 	ttm_bo_get(bo);
456 
457 	/*
458 	 * Drivers may want to override the vm_ops field. Otherwise we
459 	 * use TTM's default callbacks.
460 	 */
461 	if (!vma->vm_ops)
462 		vma->vm_ops = &ttm_bo_vm_ops;
463 
464 	/*
465 	 * Note: We're transferring the bo reference to
466 	 * vma->vm_private_data here.
467 	 */
468 
469 	vma->vm_private_data = bo;
470 
471 	vma->vm_flags |= VM_PFNMAP;
472 	vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
473 	return 0;
474 }
475 EXPORT_SYMBOL(ttm_bo_mmap_obj);
476