1/*
2 * Copyright 2013 Red Hat Inc.
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * Authors: Jérôme Glisse <jglisse@redhat.com>
15 */
16/*
17 * Heterogeneous Memory Management (HMM)
18 *
19 * See Documentation/vm/hmm.rst for reasons and overview of what HMM is and it
20 * is for. Here we focus on the HMM API description, with some explanation of
21 * the underlying implementation.
22 *
23 * Short description: HMM provides a set of helpers to share a virtual address
24 * space between CPU and a device, so that the device can access any valid
25 * address of the process (while still obeying memory protection). HMM also
26 * provides helpers to migrate process memory to device memory, and back. Each
27 * set of functionality (address space mirroring, and migration to and from
28 * device memory) can be used independently of the other.
29 *
30 *
31 * HMM address space mirroring API:
32 *
33 * Use HMM address space mirroring if you want to mirror range of the CPU page
34 * table of a process into a device page table. Here, "mirror" means "keep
35 * synchronized". Prerequisites: the device must provide the ability to write-
36 * protect its page tables (at PAGE_SIZE granularity), and must be able to
37 * recover from the resulting potential page faults.
38 *
39 * HMM guarantees that at any point in time, a given virtual address points to
40 * either the same memory in both CPU and device page tables (that is: CPU and
41 * device page tables each point to the same pages), or that one page table (CPU
42 * or device) points to no entry, while the other still points to the old page
43 * for the address. The latter case happens when the CPU page table update
44 * happens first, and then the update is mirrored over to the device page table.
45 * This does not cause any issue, because the CPU page table cannot start
46 * pointing to a new page until the device page table is invalidated.
47 *
48 * HMM uses mmu_notifiers to monitor the CPU page tables, and forwards any
49 * updates to each device driver that has registered a mirror. It also provides
50 * some API calls to help with taking a snapshot of the CPU page table, and to
51 * synchronize with any updates that might happen concurrently.
52 *
53 *
54 * HMM migration to and from device memory:
55 *
56 * HMM provides a set of helpers to hotplug device memory as ZONE_DEVICE, with
57 * a new MEMORY_DEVICE_PRIVATE type. This provides a struct page for each page
58 * of the device memory, and allows the device driver to manage its memory
59 * using those struct pages. Having struct pages for device memory makes
60 * migration easier. Because that memory is not addressable by the CPU it must
61 * never be pinned to the device; in other words, any CPU page fault can always
62 * cause the device memory to be migrated (copied/moved) back to regular memory.
63 *
64 * A new migrate helper (migrate_vma()) has been added (see mm/migrate.c) that
65 * allows use of a device DMA engine to perform the copy operation between
66 * regular system memory and device memory.
67 */
68#ifndef LINUX_HMM_H
69#define LINUX_HMM_H
70
71#include <linux/kconfig.h>
72
73#if IS_ENABLED(CONFIG_HMM)
74
75#include <linux/device.h>
76#include <linux/migrate.h>
77#include <linux/memremap.h>
78#include <linux/completion.h>
79
80struct hmm;
81
82/*
83 * hmm_pfn_flag_e - HMM flag enums
84 *
85 * Flags:
86 * HMM_PFN_VALID: pfn is valid. It has, at least, read permission.
87 * HMM_PFN_WRITE: CPU page table has write permission set
88 * HMM_PFN_DEVICE_PRIVATE: private device memory (ZONE_DEVICE)
89 *
90 * The driver provide a flags array, if driver valid bit for an entry is bit
91 * 3 ie (entry & (1 << 3)) is true if entry is valid then driver must provide
92 * an array in hmm_range.flags with hmm_range.flags[HMM_PFN_VALID] == 1 << 3.
93 * Same logic apply to all flags. This is same idea as vm_page_prot in vma
94 * except that this is per device driver rather than per architecture.
95 */
96enum hmm_pfn_flag_e {
97 HMM_PFN_VALID = 0,
98 HMM_PFN_WRITE,
99 HMM_PFN_DEVICE_PRIVATE,
100 HMM_PFN_FLAG_MAX
101};
102
103/*
104 * hmm_pfn_value_e - HMM pfn special value
105 *
106 * Flags:
107 * HMM_PFN_ERROR: corresponding CPU page table entry points to poisoned memory
108 * HMM_PFN_NONE: corresponding CPU page table entry is pte_none()
109 * HMM_PFN_SPECIAL: corresponding CPU page table entry is special; i.e., the
110 * result of vm_insert_pfn() or vm_insert_page(). Therefore, it should not
111 * be mirrored by a device, because the entry will never have HMM_PFN_VALID
112 * set and the pfn value is undefined.
113 *
114 * Driver provide entry value for none entry, error entry and special entry,
115 * driver can alias (ie use same value for error and special for instance). It
116 * should not alias none and error or special.
117 *
118 * HMM pfn value returned by hmm_vma_get_pfns() or hmm_vma_fault() will be:
119 * hmm_range.values[HMM_PFN_ERROR] if CPU page table entry is poisonous,
120 * hmm_range.values[HMM_PFN_NONE] if there is no CPU page table
121 * hmm_range.values[HMM_PFN_SPECIAL] if CPU page table entry is a special one
122 */
123enum hmm_pfn_value_e {
124 HMM_PFN_ERROR,
125 HMM_PFN_NONE,
126 HMM_PFN_SPECIAL,
127 HMM_PFN_VALUE_MAX
128};
129
130/*
131 * struct hmm_range - track invalidation lock on virtual address range
132 *
133 * @vma: the vm area struct for the range
134 * @list: all range lock are on a list
135 * @start: range virtual start address (inclusive)
136 * @end: range virtual end address (exclusive)
137 * @pfns: array of pfns (big enough for the range)
138 * @flags: pfn flags to match device driver page table
139 * @values: pfn value for some special case (none, special, error, ...)
140 * @pfn_shifts: pfn shift value (should be <= PAGE_SHIFT)
141 * @valid: pfns array did not change since it has been fill by an HMM function
142 */
143struct hmm_range {
144 struct vm_area_struct *vma;
145 struct list_head list;
146 unsigned long start;
147 unsigned long end;
148 uint64_t *pfns;
149 const uint64_t *flags;
150 const uint64_t *values;
151 uint8_t pfn_shift;
152 bool valid;
153};
154
155/*
156 * hmm_pfn_to_page() - return struct page pointed to by a valid HMM pfn
157 * @range: range use to decode HMM pfn value
158 * @pfn: HMM pfn value to get corresponding struct page from
159 * Returns: struct page pointer if pfn is a valid HMM pfn, NULL otherwise
160 *
161 * If the HMM pfn is valid (ie valid flag set) then return the struct page
162 * matching the pfn value stored in the HMM pfn. Otherwise return NULL.
163 */
164static inline struct page *hmm_pfn_to_page(const struct hmm_range *range,
165 uint64_t pfn)
166{
167 if (pfn == range->values[HMM_PFN_NONE])
168 return NULL;
169 if (pfn == range->values[HMM_PFN_ERROR])
170 return NULL;
171 if (pfn == range->values[HMM_PFN_SPECIAL])
172 return NULL;
173 if (!(pfn & range->flags[HMM_PFN_VALID]))
174 return NULL;
175 return pfn_to_page(pfn >> range->pfn_shift);
176}
177
178/*
179 * hmm_pfn_to_pfn() - return pfn value store in a HMM pfn
180 * @range: range use to decode HMM pfn value
181 * @pfn: HMM pfn value to extract pfn from
182 * Returns: pfn value if HMM pfn is valid, -1UL otherwise
183 */
184static inline unsigned long hmm_pfn_to_pfn(const struct hmm_range *range,
185 uint64_t pfn)
186{
187 if (pfn == range->values[HMM_PFN_NONE])
188 return -1UL;
189 if (pfn == range->values[HMM_PFN_ERROR])
190 return -1UL;
191 if (pfn == range->values[HMM_PFN_SPECIAL])
192 return -1UL;
193 if (!(pfn & range->flags[HMM_PFN_VALID]))
194 return -1UL;
195 return (pfn >> range->pfn_shift);
196}
197
198/*
199 * hmm_pfn_from_page() - create a valid HMM pfn value from struct page
200 * @range: range use to encode HMM pfn value
201 * @page: struct page pointer for which to create the HMM pfn
202 * Returns: valid HMM pfn for the page
203 */
204static inline uint64_t hmm_pfn_from_page(const struct hmm_range *range,
205 struct page *page)
206{
207 return (page_to_pfn(page) << range->pfn_shift) |
208 range->flags[HMM_PFN_VALID];
209}
210
211/*
212 * hmm_pfn_from_pfn() - create a valid HMM pfn value from pfn
213 * @range: range use to encode HMM pfn value
214 * @pfn: pfn value for which to create the HMM pfn
215 * Returns: valid HMM pfn for the pfn
216 */
217static inline uint64_t hmm_pfn_from_pfn(const struct hmm_range *range,
218 unsigned long pfn)
219{
220 return (pfn << range->pfn_shift) |
221 range->flags[HMM_PFN_VALID];
222}
223
224
225#if IS_ENABLED(CONFIG_HMM_MIRROR)
226/*
227 * Mirroring: how to synchronize device page table with CPU page table.
228 *
229 * A device driver that is participating in HMM mirroring must always
230 * synchronize with CPU page table updates. For this, device drivers can either
231 * directly use mmu_notifier APIs or they can use the hmm_mirror API. Device
232 * drivers can decide to register one mirror per device per process, or just
233 * one mirror per process for a group of devices. The pattern is:
234 *
235 * int device_bind_address_space(..., struct mm_struct *mm, ...)
236 * {
237 * struct device_address_space *das;
238 *
239 * // Device driver specific initialization, and allocation of das
240 * // which contains an hmm_mirror struct as one of its fields.
241 * ...
242 *
243 * ret = hmm_mirror_register(&das->mirror, mm, &device_mirror_ops);
244 * if (ret) {
245 * // Cleanup on error
246 * return ret;
247 * }
248 *
249 * // Other device driver specific initialization
250 * ...
251 * }
252 *
253 * Once an hmm_mirror is registered for an address space, the device driver
254 * will get callbacks through sync_cpu_device_pagetables() operation (see
255 * hmm_mirror_ops struct).
256 *
257 * Device driver must not free the struct containing the hmm_mirror struct
258 * before calling hmm_mirror_unregister(). The expected usage is to do that when
259 * the device driver is unbinding from an address space.
260 *
261 *
262 * void device_unbind_address_space(struct device_address_space *das)
263 * {
264 * // Device driver specific cleanup
265 * ...
266 *
267 * hmm_mirror_unregister(&das->mirror);
268 *
269 * // Other device driver specific cleanup, and now das can be freed
270 * ...
271 * }
272 */
273
274struct hmm_mirror;
275
276/*
277 * enum hmm_update_type - type of update
278 * @HMM_UPDATE_INVALIDATE: invalidate range (no indication as to why)
279 */
280enum hmm_update_type {
281 HMM_UPDATE_INVALIDATE,
282};
283
284/*
285 * struct hmm_mirror_ops - HMM mirror device operations callback
286 *
287 * @update: callback to update range on a device
288 */
289struct hmm_mirror_ops {
290 /* release() - release hmm_mirror
291 *
292 * @mirror: pointer to struct hmm_mirror
293 *
294 * This is called when the mm_struct is being released.
295 * The callback should make sure no references to the mirror occur
296 * after the callback returns.
297 */
298 void (*release)(struct hmm_mirror *mirror);
299
300 /* sync_cpu_device_pagetables() - synchronize page tables
301 *
302 * @mirror: pointer to struct hmm_mirror
303 * @update_type: type of update that occurred to the CPU page table
304 * @start: virtual start address of the range to update
305 * @end: virtual end address of the range to update
306 *
307 * This callback ultimately originates from mmu_notifiers when the CPU
308 * page table is updated. The device driver must update its page table
309 * in response to this callback. The update argument tells what action
310 * to perform.
311 *
312 * The device driver must not return from this callback until the device
313 * page tables are completely updated (TLBs flushed, etc); this is a
314 * synchronous call.
315 */
316 void (*sync_cpu_device_pagetables)(struct hmm_mirror *mirror,
317 enum hmm_update_type update_type,
318 unsigned long start,
319 unsigned long end);
320};
321
322/*
323 * struct hmm_mirror - mirror struct for a device driver
324 *
325 * @hmm: pointer to struct hmm (which is unique per mm_struct)
326 * @ops: device driver callback for HMM mirror operations
327 * @list: for list of mirrors of a given mm
328 *
329 * Each address space (mm_struct) being mirrored by a device must register one
330 * instance of an hmm_mirror struct with HMM. HMM will track the list of all
331 * mirrors for each mm_struct.
332 */
333struct hmm_mirror {
334 struct hmm *hmm;
335 const struct hmm_mirror_ops *ops;
336 struct list_head list;
337};
338
339int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm);
340void hmm_mirror_unregister(struct hmm_mirror *mirror);
341
342
343/*
344 * To snapshot the CPU page table, call hmm_vma_get_pfns(), then take a device
345 * driver lock that serializes device page table updates, then call
346 * hmm_vma_range_done(), to check if the snapshot is still valid. The same
347 * device driver page table update lock must also be used in the
348 * hmm_mirror_ops.sync_cpu_device_pagetables() callback, so that CPU page
349 * table invalidation serializes on it.
350 *
351 * YOU MUST CALL hmm_vma_range_done() ONCE AND ONLY ONCE EACH TIME YOU CALL
352 * hmm_vma_get_pfns() WITHOUT ERROR !
353 *
354 * IF YOU DO NOT FOLLOW THE ABOVE RULE THE SNAPSHOT CONTENT MIGHT BE INVALID !
355 */
356int hmm_vma_get_pfns(struct hmm_range *range);
357bool hmm_vma_range_done(struct hmm_range *range);
358
359
360/*
361 * Fault memory on behalf of device driver. Unlike handle_mm_fault(), this will
362 * not migrate any device memory back to system memory. The HMM pfn array will
363 * be updated with the fault result and current snapshot of the CPU page table
364 * for the range.
365 *
366 * The mmap_sem must be taken in read mode before entering and it might be
367 * dropped by the function if the block argument is false. In that case, the
368 * function returns -EAGAIN.
369 *
370 * Return value does not reflect if the fault was successful for every single
371 * address or not. Therefore, the caller must to inspect the HMM pfn array to
372 * determine fault status for each address.
373 *
374 * Trying to fault inside an invalid vma will result in -EINVAL.
375 *
376 * See the function description in mm/hmm.c for further documentation.
377 */
378int hmm_vma_fault(struct hmm_range *range, bool block);
379
380/* Below are for HMM internal use only! Not to be used by device driver! */
381void hmm_mm_destroy(struct mm_struct *mm);
382
383static inline void hmm_mm_init(struct mm_struct *mm)
384{
385 mm->hmm = NULL;
386}
387#else /* IS_ENABLED(CONFIG_HMM_MIRROR) */
388static inline void hmm_mm_destroy(struct mm_struct *mm) {}
389static inline void hmm_mm_init(struct mm_struct *mm) {}
390#endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
391
392#if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC)
393struct hmm_devmem;
394
395struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
396 unsigned long addr);
397
398/*
399 * struct hmm_devmem_ops - callback for ZONE_DEVICE memory events
400 *
401 * @free: call when refcount on page reach 1 and thus is no longer use
402 * @fault: call when there is a page fault to unaddressable memory
403 *
404 * Both callback happens from page_free() and page_fault() callback of struct
405 * dev_pagemap respectively. See include/linux/memremap.h for more details on
406 * those.
407 *
408 * The hmm_devmem_ops callback are just here to provide a coherent and
409 * uniq API to device driver and device driver should not register their
410 * own page_free() or page_fault() but rely on the hmm_devmem_ops call-
411 * back.
412 */
413struct hmm_devmem_ops {
414 /*
415 * free() - free a device page
416 * @devmem: device memory structure (see struct hmm_devmem)
417 * @page: pointer to struct page being freed
418 *
419 * Call back occurs whenever a device page refcount reach 1 which
420 * means that no one is holding any reference on the page anymore
421 * (ZONE_DEVICE page have an elevated refcount of 1 as default so
422 * that they are not release to the general page allocator).
423 *
424 * Note that callback has exclusive ownership of the page (as no
425 * one is holding any reference).
426 */
427 void (*free)(struct hmm_devmem *devmem, struct page *page);
428 /*
429 * fault() - CPU page fault or get user page (GUP)
430 * @devmem: device memory structure (see struct hmm_devmem)
431 * @vma: virtual memory area containing the virtual address
432 * @addr: virtual address that faulted or for which there is a GUP
433 * @page: pointer to struct page backing virtual address (unreliable)
434 * @flags: FAULT_FLAG_* (see include/linux/mm.h)
435 * @pmdp: page middle directory
436 * Returns: VM_FAULT_MINOR/MAJOR on success or one of VM_FAULT_ERROR
437 * on error
438 *
439 * The callback occurs whenever there is a CPU page fault or GUP on a
440 * virtual address. This means that the device driver must migrate the
441 * page back to regular memory (CPU accessible).
442 *
443 * The device driver is free to migrate more than one page from the
444 * fault() callback as an optimization. However if device decide to
445 * migrate more than one page it must always priotirize the faulting
446 * address over the others.
447 *
448 * The struct page pointer is only given as an hint to allow quick
449 * lookup of internal device driver data. A concurrent migration
450 * might have already free that page and the virtual address might
451 * not longer be back by it. So it should not be modified by the
452 * callback.
453 *
454 * Note that mmap semaphore is held in read mode at least when this
455 * callback occurs, hence the vma is valid upon callback entry.
456 */
457 int (*fault)(struct hmm_devmem *devmem,
458 struct vm_area_struct *vma,
459 unsigned long addr,
460 const struct page *page,
461 unsigned int flags,
462 pmd_t *pmdp);
463};
464
465/*
466 * struct hmm_devmem - track device memory
467 *
468 * @completion: completion object for device memory
469 * @pfn_first: first pfn for this resource (set by hmm_devmem_add())
470 * @pfn_last: last pfn for this resource (set by hmm_devmem_add())
471 * @resource: IO resource reserved for this chunk of memory
472 * @pagemap: device page map for that chunk
473 * @device: device to bind resource to
474 * @ops: memory operations callback
475 * @ref: per CPU refcount
476 *
477 * This an helper structure for device drivers that do not wish to implement
478 * the gory details related to hotplugging new memoy and allocating struct
479 * pages.
480 *
481 * Device drivers can directly use ZONE_DEVICE memory on their own if they
482 * wish to do so.
483 */
484struct hmm_devmem {
485 struct completion completion;
486 unsigned long pfn_first;
487 unsigned long pfn_last;
488 struct resource *resource;
489 struct device *device;
490 struct dev_pagemap pagemap;
491 const struct hmm_devmem_ops *ops;
492 struct percpu_ref ref;
493};
494
495/*
496 * To add (hotplug) device memory, HMM assumes that there is no real resource
497 * that reserves a range in the physical address space (this is intended to be
498 * use by unaddressable device memory). It will reserve a physical range big
499 * enough and allocate struct page for it.
500 *
501 * The device driver can wrap the hmm_devmem struct inside a private device
502 * driver struct. The device driver must call hmm_devmem_remove() before the
503 * device goes away and before freeing the hmm_devmem struct memory.
504 */
505struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
506 struct device *device,
507 unsigned long size);
508struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
509 struct device *device,
510 struct resource *res);
511void hmm_devmem_remove(struct hmm_devmem *devmem);
512
513/*
514 * hmm_devmem_page_set_drvdata - set per-page driver data field
515 *
516 * @page: pointer to struct page
517 * @data: driver data value to set
518 *
519 * Because page can not be on lru we have an unsigned long that driver can use
520 * to store a per page field. This just a simple helper to do that.
521 */
522static inline void hmm_devmem_page_set_drvdata(struct page *page,
523 unsigned long data)
524{
525 page->hmm_data = data;
526}
527
528/*
529 * hmm_devmem_page_get_drvdata - get per page driver data field
530 *
531 * @page: pointer to struct page
532 * Return: driver data value
533 */
534static inline unsigned long hmm_devmem_page_get_drvdata(const struct page *page)
535{
536 return page->hmm_data;
537}
538
539
540/*
541 * struct hmm_device - fake device to hang device memory onto
542 *
543 * @device: device struct
544 * @minor: device minor number
545 */
546struct hmm_device {
547 struct device device;
548 unsigned int minor;
549};
550
551/*
552 * A device driver that wants to handle multiple devices memory through a
553 * single fake device can use hmm_device to do so. This is purely a helper and
554 * it is not strictly needed, in order to make use of any HMM functionality.
555 */
556struct hmm_device *hmm_device_new(void *drvdata);
557void hmm_device_put(struct hmm_device *hmm_device);
558#endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
559#else /* IS_ENABLED(CONFIG_HMM) */
560static inline void hmm_mm_destroy(struct mm_struct *mm) {}
561static inline void hmm_mm_init(struct mm_struct *mm) {}
562#endif /* IS_ENABLED(CONFIG_HMM) */
563
564#endif /* LINUX_HMM_H */
565