summaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/i915/gem/i915_gem_mman.c
blob: 39a661927d8e0f02e2d4aff967eb4b3d9bd3d989 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
/*
 * SPDX-License-Identifier: MIT
 *
 * Copyright © 2014-2016 Intel Corporation
 */

#include <linux/mman.h>
#include <linux/sizes.h>

#include "i915_drv.h"
#include "i915_gem_gtt.h"
#include "i915_gem_ioctls.h"
#include "i915_gem_object.h"
#include "i915_vma.h"
#include "intel_drv.h"

static inline bool
__vma_matches(struct vm_area_struct *vma, struct file *filp,
	      unsigned long addr, unsigned long size)
{
	if (vma->vm_file != filp)
		return false;

	return vma->vm_start == addr &&
	       (vma->vm_end - vma->vm_start) == PAGE_ALIGN(size);
}

/**
 * i915_gem_mmap_ioctl - Maps the contents of an object, returning the address
 *			 it is mapped to.
 * @dev: drm device
 * @data: ioctl data blob
 * @file: drm file
 *
 * While the mapping holds a reference on the contents of the object, it doesn't
 * imply a ref on the object itself.
 *
 * IMPORTANT:
 *
 * DRM driver writers who look a this function as an example for how to do GEM
 * mmap support, please don't implement mmap support like here. The modern way
 * to implement DRM mmap support is with an mmap offset ioctl (like
 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
 * That way debug tooling like valgrind will understand what's going on, hiding
 * the mmap call in a driver private ioctl will break that. The i915 driver only
 * does cpu mmaps this way because we didn't know better.
 */
int
i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
		    struct drm_file *file)
{
	struct drm_i915_gem_mmap *args = data;
	struct drm_i915_gem_object *obj;
	unsigned long addr;

	if (args->flags & ~(I915_MMAP_WC))
		return -EINVAL;

	if (args->flags & I915_MMAP_WC && !boot_cpu_has(X86_FEATURE_PAT))
		return -ENODEV;

	obj = i915_gem_object_lookup(file, args->handle);
	if (!obj)
		return -ENOENT;

	/* prime objects have no backing filp to GEM mmap
	 * pages from.
	 */
	if (!obj->base.filp) {
		addr = -ENXIO;
		goto err;
	}

	if (range_overflows(args->offset, args->size, (u64)obj->base.size)) {
		addr = -EINVAL;
		goto err;
	}

	addr = vm_mmap(obj->base.filp, 0, args->size,
		       PROT_READ | PROT_WRITE, MAP_SHARED,
		       args->offset);
	if (IS_ERR_VALUE(addr))
		goto err;

	if (args->flags & I915_MMAP_WC) {
		struct mm_struct *mm = current->mm;
		struct vm_area_struct *vma;

		if (down_write_killable(&mm->mmap_sem)) {
			addr = -EINTR;
			goto err;
		}
		vma = find_vma(mm, addr);
		if (vma && __vma_matches(vma, obj->base.filp, addr, args->size))
			vma->vm_page_prot =
				pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
		else
			addr = -ENOMEM;
		up_write(&mm->mmap_sem);
		if (IS_ERR_VALUE(addr))
			goto err;

		/* This may race, but that's ok, it only gets set */
		WRITE_ONCE(obj->frontbuffer_ggtt_origin, ORIGIN_CPU);
	}
	i915_gem_object_put(obj);

	args->addr_ptr = (u64)addr;
	return 0;

err:
	i915_gem_object_put(obj);
	return addr;
}

static unsigned int tile_row_pages(const struct drm_i915_gem_object *obj)
{
	return i915_gem_object_get_tile_row_size(obj) >> PAGE_SHIFT;
}

/**
 * i915_gem_mmap_gtt_version - report the current feature set for GTT mmaps
 *
 * A history of the GTT mmap interface:
 *
 * 0 - Everything had to fit into the GTT. Both parties of a memcpy had to
 *     aligned and suitable for fencing, and still fit into the available
 *     mappable space left by the pinned display objects. A classic problem
 *     we called the page-fault-of-doom where we would ping-pong between
 *     two objects that could not fit inside the GTT and so the memcpy
 *     would page one object in at the expense of the other between every
 *     single byte.
 *
 * 1 - Objects can be any size, and have any compatible fencing (X Y, or none
 *     as set via i915_gem_set_tiling() [DRM_I915_GEM_SET_TILING]). If the
 *     object is too large for the available space (or simply too large
 *     for the mappable aperture!), a view is created instead and faulted
 *     into userspace. (This view is aligned and sized appropriately for
 *     fenced access.)
 *
 * 2 - Recognise WC as a separate cache domain so that we can flush the
 *     delayed writes via GTT before performing direct access via WC.
 *
 * 3 - Remove implicit set-domain(GTT) and synchronisation on initial
 *     pagefault; swapin remains transparent.
 *
 * Restrictions:
 *
 *  * snoopable objects cannot be accessed via the GTT. It can cause machine
 *    hangs on some architectures, corruption on others. An attempt to service
 *    a GTT page fault from a snoopable object will generate a SIGBUS.
 *
 *  * the object must be able to fit into RAM (physical memory, though no
 *    limited to the mappable aperture).
 *
 *
 * Caveats:
 *
 *  * a new GTT page fault will synchronize rendering from the GPU and flush
 *    all data to system memory. Subsequent access will not be synchronized.
 *
 *  * all mappings are revoked on runtime device suspend.
 *
 *  * there are only 8, 16 or 32 fence registers to share between all users
 *    (older machines require fence register for display and blitter access
 *    as well). Contention of the fence registers will cause the previous users
 *    to be unmapped and any new access will generate new page faults.
 *
 *  * running out of memory while servicing a fault may generate a SIGBUS,
 *    rather than the expected SIGSEGV.
 */
int i915_gem_mmap_gtt_version(void)
{
	return 3;
}

static inline struct i915_ggtt_view
compute_partial_view(const struct drm_i915_gem_object *obj,
		     pgoff_t page_offset,
		     unsigned int chunk)
{
	struct i915_ggtt_view view;

	if (i915_gem_object_is_tiled(obj))
		chunk = roundup(chunk, tile_row_pages(obj));

	view.type = I915_GGTT_VIEW_PARTIAL;
	view.partial.offset = rounddown(page_offset, chunk);
	view.partial.size =
		min_t(unsigned int, chunk,
		      (obj->base.size >> PAGE_SHIFT) - view.partial.offset);

	/* If the partial covers the entire object, just create a normal VMA. */
	if (chunk >= obj->base.size >> PAGE_SHIFT)
		view.type = I915_GGTT_VIEW_NORMAL;

	return view;
}

/**
 * i915_gem_fault - fault a page into the GTT
 * @vmf: fault info
 *
 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
 * from userspace.  The fault handler takes care of binding the object to
 * the GTT (if needed), allocating and programming a fence register (again,
 * only if needed based on whether the old reg is still valid or the object
 * is tiled) and inserting a new PTE into the faulting process.
 *
 * Note that the faulting process may involve evicting existing objects
 * from the GTT and/or fence registers to make room.  So performance may
 * suffer if the GTT working set is large or there are few fence registers
 * left.
 *
 * The current feature set supported by i915_gem_fault() and thus GTT mmaps
 * is exposed via I915_PARAM_MMAP_GTT_VERSION (see i915_gem_mmap_gtt_version).
 */
vm_fault_t i915_gem_fault(struct vm_fault *vmf)
{
#define MIN_CHUNK_PAGES (SZ_1M >> PAGE_SHIFT)
	struct vm_area_struct *area = vmf->vma;
	struct drm_i915_gem_object *obj = to_intel_bo(area->vm_private_data);
	struct drm_device *dev = obj->base.dev;
	struct drm_i915_private *i915 = to_i915(dev);
	struct intel_runtime_pm *rpm = &i915->runtime_pm;
	struct i915_ggtt *ggtt = &i915->ggtt;
	bool write = area->vm_flags & VM_WRITE;
	intel_wakeref_t wakeref;
	struct i915_vma *vma;
	pgoff_t page_offset;
	int srcu;
	int ret;

	/* Sanity check that we allow writing into this object */
	if (i915_gem_object_is_readonly(obj) && write)
		return VM_FAULT_SIGBUS;

	/* We don't use vmf->pgoff since that has the fake offset */
	page_offset = (vmf->address - area->vm_start) >> PAGE_SHIFT;

	trace_i915_gem_object_fault(obj, page_offset, true, write);

	ret = i915_gem_object_pin_pages(obj);
	if (ret)
		goto err;

	wakeref = intel_runtime_pm_get(rpm);

	srcu = i915_reset_trylock(i915);
	if (srcu < 0) {
		ret = srcu;
		goto err_rpm;
	}

	ret = i915_mutex_lock_interruptible(dev);
	if (ret)
		goto err_reset;

	/* Access to snoopable pages through the GTT is incoherent. */
	if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(i915)) {
		ret = -EFAULT;
		goto err_unlock;
	}

	/* Now pin it into the GTT as needed */
	vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
				       PIN_MAPPABLE |
				       PIN_NONBLOCK |
				       PIN_NONFAULT);
	if (IS_ERR(vma)) {
		/* Use a partial view if it is bigger than available space */
		struct i915_ggtt_view view =
			compute_partial_view(obj, page_offset, MIN_CHUNK_PAGES);
		unsigned int flags;

		flags = PIN_MAPPABLE;
		if (view.type == I915_GGTT_VIEW_NORMAL)
			flags |= PIN_NONBLOCK; /* avoid warnings for pinned */

		/*
		 * Userspace is now writing through an untracked VMA, abandon
		 * all hope that the hardware is able to track future writes.
		 */
		obj->frontbuffer_ggtt_origin = ORIGIN_CPU;

		vma = i915_gem_object_ggtt_pin(obj, &view, 0, 0, flags);
		if (IS_ERR(vma) && !view.type) {
			flags = PIN_MAPPABLE;
			view.type = I915_GGTT_VIEW_PARTIAL;
			vma = i915_gem_object_ggtt_pin(obj, &view, 0, 0, flags);
		}
	}
	if (IS_ERR(vma)) {
		ret = PTR_ERR(vma);
		goto err_unlock;
	}

	ret = i915_vma_pin_fence(vma);
	if (ret)
		goto err_unpin;

	/* Finally, remap it using the new GTT offset */
	ret = remap_io_mapping(area,
			       area->vm_start + (vma->ggtt_view.partial.offset << PAGE_SHIFT),
			       (ggtt->gmadr.start + vma->node.start) >> PAGE_SHIFT,
			       min_t(u64, vma->size, area->vm_end - area->vm_start),
			       &ggtt->iomap);
	if (ret)
		goto err_fence;

	/* Mark as being mmapped into userspace for later revocation */
	assert_rpm_wakelock_held(rpm);
	if (!i915_vma_set_userfault(vma) && !obj->userfault_count++)
		list_add(&obj->userfault_link, &i915->ggtt.userfault_list);
	if (CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND)
		intel_wakeref_auto(&i915->ggtt.userfault_wakeref,
				   msecs_to_jiffies_timeout(CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND));
	GEM_BUG_ON(!obj->userfault_count);

	i915_vma_set_ggtt_write(vma);

err_fence:
	i915_vma_unpin_fence(vma);
err_unpin:
	__i915_vma_unpin(vma);
err_unlock:
	mutex_unlock(&dev->struct_mutex);
err_reset:
	i915_reset_unlock(i915, srcu);
err_rpm:
	intel_runtime_pm_put(rpm, wakeref);
	i915_gem_object_unpin_pages(obj);
err:
	switch (ret) {
	case -EIO:
		/*
		 * We eat errors when the gpu is terminally wedged to avoid
		 * userspace unduly crashing (gl has no provisions for mmaps to
		 * fail). But any other -EIO isn't ours (e.g. swap in failure)
		 * and so needs to be reported.
		 */
		if (!i915_terminally_wedged(i915))
			return VM_FAULT_SIGBUS;
		/* else, fall through */
	case -EAGAIN:
		/*
		 * EAGAIN means the gpu is hung and we'll wait for the error
		 * handler to reset everything when re-faulting in
		 * i915_mutex_lock_interruptible.
		 */
	case 0:
	case -ERESTARTSYS:
	case -EINTR:
	case -EBUSY:
		/*
		 * EBUSY is ok: this just means that another thread
		 * already did the job.
		 */
		return VM_FAULT_NOPAGE;
	case -ENOMEM:
		return VM_FAULT_OOM;
	case -ENOSPC:
	case -EFAULT:
		return VM_FAULT_SIGBUS;
	default:
		WARN_ONCE(ret, "unhandled error in %s: %i\n", __func__, ret);
		return VM_FAULT_SIGBUS;
	}
}

void __i915_gem_object_release_mmap(struct drm_i915_gem_object *obj)
{
	struct i915_vma *vma;

	GEM_BUG_ON(!obj->userfault_count);

	obj->userfault_count = 0;
	list_del(&obj->userfault_link);
	drm_vma_node_unmap(&obj->base.vma_node,
			   obj->base.dev->anon_inode->i_mapping);

	for_each_ggtt_vma(vma, obj)
		i915_vma_unset_userfault(vma);
}

/**
 * i915_gem_object_release_mmap - remove physical page mappings
 * @obj: obj in question
 *
 * Preserve the reservation of the mmapping with the DRM core code, but
 * relinquish ownership of the pages back to the system.
 *
 * It is vital that we remove the page mapping if we have mapped a tiled
 * object through the GTT and then lose the fence register due to
 * resource pressure. Similarly if the object has been moved out of the
 * aperture, than pages mapped into userspace must be revoked. Removing the
 * mapping will then trigger a page fault on the next user access, allowing
 * fixup by i915_gem_fault().
 */
void i915_gem_object_release_mmap(struct drm_i915_gem_object *obj)
{
	struct drm_i915_private *i915 = to_i915(obj->base.dev);
	intel_wakeref_t wakeref;

	/* Serialisation between user GTT access and our code depends upon
	 * revoking the CPU's PTE whilst the mutex is held. The next user
	 * pagefault then has to wait until we release the mutex.
	 *
	 * Note that RPM complicates somewhat by adding an additional
	 * requirement that operations to the GGTT be made holding the RPM
	 * wakeref.
	 */
	lockdep_assert_held(&i915->drm.struct_mutex);
	wakeref = intel_runtime_pm_get(&i915->runtime_pm);

	if (!obj->userfault_count)
		goto out;

	__i915_gem_object_release_mmap(obj);

	/* Ensure that the CPU's PTE are revoked and there are not outstanding
	 * memory transactions from userspace before we return. The TLB
	 * flushing implied above by changing the PTE above *should* be
	 * sufficient, an extra barrier here just provides us with a bit
	 * of paranoid documentation about our requirement to serialise
	 * memory writes before touching registers / GSM.
	 */
	wmb();

out:
	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
}

static int create_mmap_offset(struct drm_i915_gem_object *obj)
{
	struct drm_i915_private *i915 = to_i915(obj->base.dev);
	int err;

	err = drm_gem_create_mmap_offset(&obj->base);
	if (likely(!err))
		return 0;

	/* Attempt to reap some mmap space from dead objects */
	do {
		err = i915_gem_wait_for_idle(i915,
					     I915_WAIT_INTERRUPTIBLE,
					     MAX_SCHEDULE_TIMEOUT);
		if (err)
			break;

		i915_gem_drain_freed_objects(i915);
		err = drm_gem_create_mmap_offset(&obj->base);
		if (!err)
			break;

	} while (flush_delayed_work(&i915->gem.retire_work));

	return err;
}

int
i915_gem_mmap_gtt(struct drm_file *file,
		  struct drm_device *dev,
		  u32 handle,
		  u64 *offset)
{
	struct drm_i915_gem_object *obj;
	int ret;

	obj = i915_gem_object_lookup(file, handle);
	if (!obj)
		return -ENOENT;

	ret = create_mmap_offset(obj);
	if (ret == 0)
		*offset = drm_vma_node_offset_addr(&obj->base.vma_node);

	i915_gem_object_put(obj);
	return ret;
}

/**
 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
 * @dev: DRM device
 * @data: GTT mapping ioctl data
 * @file: GEM object info
 *
 * Simply returns the fake offset to userspace so it can mmap it.
 * The mmap call will end up in drm_gem_mmap(), which will set things
 * up so we can get faults in the handler above.
 *
 * The fault handler will take care of binding the object into the GTT
 * (since it may have been evicted to make room for something), allocating
 * a fence register, and mapping the appropriate aperture address into
 * userspace.
 */
int
i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
			struct drm_file *file)
{
	struct drm_i915_gem_mmap_gtt *args = data;

	return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
}

#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
#include "selftests/i915_gem_mman.c"
#endif