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 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
29 */
30
31#define pr_fmt(fmt) "[TTM] " fmt
32
33#include <ttm/ttm_module.h>
34#include <ttm/ttm_bo_driver.h>
35#include <ttm/ttm_placement.h>
36#include <drm/drm_vma_manager.h>
37#include <linux/mm.h>
38#include <linux/rbtree.h>
39#include <linux/module.h>
40#include <linux/uaccess.h>
41
42#define TTM_BO_VM_NUM_PREFAULT 16
43
44static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
45				struct vm_area_struct *vma,
46				struct vm_fault *vmf)
47{
48	int ret = 0;
49
50	if (likely(!test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)))
51		goto out_unlock;
52
53	/*
54	 * Quick non-stalling check for idle.
55	 */
56	ret = ttm_bo_wait(bo, false, false, true);
57	if (likely(ret == 0))
58		goto out_unlock;
59
60	/*
61	 * If possible, avoid waiting for GPU with mmap_sem
62	 * held.
63	 */
64	if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
65		ret = VM_FAULT_RETRY;
66		if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
67			goto out_unlock;
68
69		up_read(&vma->vm_mm->mmap_sem);
70		(void) ttm_bo_wait(bo, false, true, false);
71		goto out_unlock;
72	}
73
74	/*
75	 * Ordinary wait.
76	 */
77	ret = ttm_bo_wait(bo, false, true, false);
78	if (unlikely(ret != 0))
79		ret = (ret != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
80			VM_FAULT_NOPAGE;
81
82out_unlock:
83	return ret;
84}
85
86static int ttm_bo_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
87{
88	struct ttm_buffer_object *bo = (struct ttm_buffer_object *)
89	    vma->vm_private_data;
90	struct ttm_bo_device *bdev = bo->bdev;
91	unsigned long page_offset;
92	unsigned long page_last;
93	unsigned long pfn;
94	struct ttm_tt *ttm = NULL;
95	struct page *page;
96	int ret;
97	int i;
98	unsigned long address = (unsigned long)vmf->virtual_address;
99	int retval = VM_FAULT_NOPAGE;
100	struct ttm_mem_type_manager *man =
101		&bdev->man[bo->mem.mem_type];
102	struct vm_area_struct cvma;
103
104	/*
105	 * Work around locking order reversal in fault / nopfn
106	 * between mmap_sem and bo_reserve: Perform a trylock operation
107	 * for reserve, and if it fails, retry the fault after waiting
108	 * for the buffer to become unreserved.
109	 */
110	ret = ttm_bo_reserve(bo, true, true, false, NULL);
111	if (unlikely(ret != 0)) {
112		if (ret != -EBUSY)
113			return VM_FAULT_NOPAGE;
114
115		if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
116			if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
117				up_read(&vma->vm_mm->mmap_sem);
118				(void) ttm_bo_wait_unreserved(bo);
119			}
120
121			return VM_FAULT_RETRY;
122		}
123
124		/*
125		 * If we'd want to change locking order to
126		 * mmap_sem -> bo::reserve, we'd use a blocking reserve here
127		 * instead of retrying the fault...
128		 */
129		return VM_FAULT_NOPAGE;
130	}
131
132	/*
133	 * Refuse to fault imported pages. This should be handled
134	 * (if at all) by redirecting mmap to the exporter.
135	 */
136	if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
137		retval = VM_FAULT_SIGBUS;
138		goto out_unlock;
139	}
140
141	if (bdev->driver->fault_reserve_notify) {
142		ret = bdev->driver->fault_reserve_notify(bo);
143		switch (ret) {
144		case 0:
145			break;
146		case -EBUSY:
147		case -ERESTARTSYS:
148			retval = VM_FAULT_NOPAGE;
149			goto out_unlock;
150		default:
151			retval = VM_FAULT_SIGBUS;
152			goto out_unlock;
153		}
154	}
155
156	/*
157	 * Wait for buffer data in transit, due to a pipelined
158	 * move.
159	 */
160	ret = ttm_bo_vm_fault_idle(bo, vma, vmf);
161	if (unlikely(ret != 0)) {
162		retval = ret;
163		goto out_unlock;
164	}
165
166	ret = ttm_mem_io_lock(man, true);
167	if (unlikely(ret != 0)) {
168		retval = VM_FAULT_NOPAGE;
169		goto out_unlock;
170	}
171	ret = ttm_mem_io_reserve_vm(bo);
172	if (unlikely(ret != 0)) {
173		retval = VM_FAULT_SIGBUS;
174		goto out_io_unlock;
175	}
176
177	page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
178		vma->vm_pgoff - drm_vma_node_start(&bo->vma_node);
179	page_last = vma_pages(vma) + vma->vm_pgoff -
180		drm_vma_node_start(&bo->vma_node);
181
182	if (unlikely(page_offset >= bo->num_pages)) {
183		retval = VM_FAULT_SIGBUS;
184		goto out_io_unlock;
185	}
186
187	/*
188	 * Make a local vma copy to modify the page_prot member
189	 * and vm_flags if necessary. The vma parameter is protected
190	 * by mmap_sem in write mode.
191	 */
192	cvma = *vma;
193	cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags);
194
195	if (bo->mem.bus.is_iomem) {
196		cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
197						cvma.vm_page_prot);
198	} else {
199		ttm = bo->ttm;
200		cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
201						cvma.vm_page_prot);
202
203		/* Allocate all page at once, most common usage */
204		if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
205			retval = VM_FAULT_OOM;
206			goto out_io_unlock;
207		}
208	}
209
210	/*
211	 * Speculatively prefault a number of pages. Only error on
212	 * first page.
213	 */
214	for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
215		if (bo->mem.bus.is_iomem)
216			pfn = ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT) + page_offset;
217		else {
218			page = ttm->pages[page_offset];
219			if (unlikely(!page && i == 0)) {
220				retval = VM_FAULT_OOM;
221				goto out_io_unlock;
222			} else if (unlikely(!page)) {
223				break;
224			}
225			page->mapping = vma->vm_file->f_mapping;
226			page->index = drm_vma_node_start(&bo->vma_node) +
227				page_offset;
228			pfn = page_to_pfn(page);
229		}
230
231		if (vma->vm_flags & VM_MIXEDMAP)
232			ret = vm_insert_mixed(&cvma, address, pfn);
233		else
234			ret = vm_insert_pfn(&cvma, address, pfn);
235
236		/*
237		 * Somebody beat us to this PTE or prefaulting to
238		 * an already populated PTE, or prefaulting error.
239		 */
240
241		if (unlikely((ret == -EBUSY) || (ret != 0 && i > 0)))
242			break;
243		else if (unlikely(ret != 0)) {
244			retval =
245			    (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
246			goto out_io_unlock;
247		}
248
249		address += PAGE_SIZE;
250		if (unlikely(++page_offset >= page_last))
251			break;
252	}
253out_io_unlock:
254	ttm_mem_io_unlock(man);
255out_unlock:
256	ttm_bo_unreserve(bo);
257	return retval;
258}
259
260static void ttm_bo_vm_open(struct vm_area_struct *vma)
261{
262	struct ttm_buffer_object *bo =
263	    (struct ttm_buffer_object *)vma->vm_private_data;
264
265	WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
266
267	(void)ttm_bo_reference(bo);
268}
269
270static void ttm_bo_vm_close(struct vm_area_struct *vma)
271{
272	struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data;
273
274	ttm_bo_unref(&bo);
275	vma->vm_private_data = NULL;
276}
277
278static const struct vm_operations_struct ttm_bo_vm_ops = {
279	.fault = ttm_bo_vm_fault,
280	.open = ttm_bo_vm_open,
281	.close = ttm_bo_vm_close
282};
283
284static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
285						  unsigned long offset,
286						  unsigned long pages)
287{
288	struct drm_vma_offset_node *node;
289	struct ttm_buffer_object *bo = NULL;
290
291	drm_vma_offset_lock_lookup(&bdev->vma_manager);
292
293	node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages);
294	if (likely(node)) {
295		bo = container_of(node, struct ttm_buffer_object, vma_node);
296		if (!kref_get_unless_zero(&bo->kref))
297			bo = NULL;
298	}
299
300	drm_vma_offset_unlock_lookup(&bdev->vma_manager);
301
302	if (!bo)
303		pr_err("Could not find buffer object to map\n");
304
305	return bo;
306}
307
308int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
309		struct ttm_bo_device *bdev)
310{
311	struct ttm_bo_driver *driver;
312	struct ttm_buffer_object *bo;
313	int ret;
314
315	bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
316	if (unlikely(!bo))
317		return -EINVAL;
318
319	driver = bo->bdev->driver;
320	if (unlikely(!driver->verify_access)) {
321		ret = -EPERM;
322		goto out_unref;
323	}
324	ret = driver->verify_access(bo, filp);
325	if (unlikely(ret != 0))
326		goto out_unref;
327
328	vma->vm_ops = &ttm_bo_vm_ops;
329
330	/*
331	 * Note: We're transferring the bo reference to
332	 * vma->vm_private_data here.
333	 */
334
335	vma->vm_private_data = bo;
336
337	/*
338	 * We'd like to use VM_PFNMAP on shared mappings, where
339	 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
340	 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
341	 * bad for performance. Until that has been sorted out, use
342	 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
343	 */
344	vma->vm_flags |= VM_MIXEDMAP;
345	vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
346	return 0;
347out_unref:
348	ttm_bo_unref(&bo);
349	return ret;
350}
351EXPORT_SYMBOL(ttm_bo_mmap);
352
353int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
354{
355	if (vma->vm_pgoff != 0)
356		return -EACCES;
357
358	vma->vm_ops = &ttm_bo_vm_ops;
359	vma->vm_private_data = ttm_bo_reference(bo);
360	vma->vm_flags |= VM_MIXEDMAP;
361	vma->vm_flags |= VM_IO | VM_DONTEXPAND;
362	return 0;
363}
364EXPORT_SYMBOL(ttm_fbdev_mmap);
365