blob: e5834082de60702f36a9a4f1c9fc00e725e19961 [file] [log] [blame]
Jérôme Glisse133ff0e2017-09-08 16:11:23 -07001/*
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 *
Jérôme Glissef813f212018-10-30 15:04:06 -070014 * Authors: Jérôme Glisse <jglisse@redhat.com>
Jérôme Glisse133ff0e2017-09-08 16:11:23 -070015 */
16/*
17 * Heterogeneous Memory Management (HMM)
18 *
Mike Rapoportad56b732018-03-21 21:22:47 +020019 * See Documentation/vm/hmm.rst for reasons and overview of what HMM is and it
Jérôme Glisse133ff0e2017-09-08 16:11:23 -070020 * 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>
Dan Williams063a7d12018-12-28 00:39:46 -080072#include <asm/pgtable.h>
Jérôme Glisse133ff0e2017-09-08 16:11:23 -070073
74#if IS_ENABLED(CONFIG_HMM)
75
Jérôme Glisse858b54d2017-09-08 16:12:02 -070076#include <linux/device.h>
Jérôme Glisse4ef589d2017-09-08 16:11:58 -070077#include <linux/migrate.h>
78#include <linux/memremap.h>
79#include <linux/completion.h>
Jérôme Glissea3e0d412019-05-13 17:20:01 -070080#include <linux/mmu_notifier.h>
Jérôme Glisse4ef589d2017-09-08 16:11:58 -070081
Jérôme Glissea3e0d412019-05-13 17:20:01 -070082
83/*
84 * struct hmm - HMM per mm struct
85 *
86 * @mm: mm struct this HMM struct is bound to
87 * @lock: lock protecting ranges list
88 * @ranges: list of range being snapshotted
89 * @mirrors: list of mirrors for this mm
90 * @mmu_notifier: mmu notifier to track updates to CPU page table
91 * @mirrors_sem: read/write semaphore protecting the mirrors list
92 * @wq: wait queue for user waiting on a range invalidation
93 * @notifiers: count of active mmu notifiers
94 * @dead: is the mm dead ?
95 */
96struct hmm {
97 struct mm_struct *mm;
98 struct kref kref;
99 struct mutex lock;
100 struct list_head ranges;
101 struct list_head mirrors;
102 struct mmu_notifier mmu_notifier;
103 struct rw_semaphore mirrors_sem;
104 wait_queue_head_t wq;
105 long notifiers;
106 bool dead;
107};
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700108
109/*
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700110 * hmm_pfn_flag_e - HMM flag enums
111 *
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700112 * Flags:
Jérôme Glisse86586a42018-04-10 16:28:34 -0700113 * HMM_PFN_VALID: pfn is valid. It has, at least, read permission.
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700114 * HMM_PFN_WRITE: CPU page table has write permission set
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700115 * HMM_PFN_DEVICE_PRIVATE: private device memory (ZONE_DEVICE)
116 *
117 * The driver provide a flags array, if driver valid bit for an entry is bit
118 * 3 ie (entry & (1 << 3)) is true if entry is valid then driver must provide
119 * an array in hmm_range.flags with hmm_range.flags[HMM_PFN_VALID] == 1 << 3.
120 * Same logic apply to all flags. This is same idea as vm_page_prot in vma
121 * except that this is per device driver rather than per architecture.
122 */
123enum hmm_pfn_flag_e {
124 HMM_PFN_VALID = 0,
125 HMM_PFN_WRITE,
126 HMM_PFN_DEVICE_PRIVATE,
127 HMM_PFN_FLAG_MAX
128};
129
130/*
131 * hmm_pfn_value_e - HMM pfn special value
132 *
133 * Flags:
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700134 * HMM_PFN_ERROR: corresponding CPU page table entry points to poisoned memory
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700135 * HMM_PFN_NONE: corresponding CPU page table entry is pte_none()
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700136 * HMM_PFN_SPECIAL: corresponding CPU page table entry is special; i.e., the
Matthew Wilcox67fa1662018-10-26 15:04:26 -0700137 * result of vmf_insert_pfn() or vm_insert_page(). Therefore, it should not
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700138 * be mirrored by a device, because the entry will never have HMM_PFN_VALID
139 * set and the pfn value is undefined.
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700140 *
141 * Driver provide entry value for none entry, error entry and special entry,
142 * driver can alias (ie use same value for error and special for instance). It
143 * should not alias none and error or special.
144 *
145 * HMM pfn value returned by hmm_vma_get_pfns() or hmm_vma_fault() will be:
146 * hmm_range.values[HMM_PFN_ERROR] if CPU page table entry is poisonous,
147 * hmm_range.values[HMM_PFN_NONE] if there is no CPU page table
148 * hmm_range.values[HMM_PFN_SPECIAL] if CPU page table entry is a special one
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700149 */
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700150enum hmm_pfn_value_e {
151 HMM_PFN_ERROR,
152 HMM_PFN_NONE,
153 HMM_PFN_SPECIAL,
154 HMM_PFN_VALUE_MAX
155};
156
157/*
158 * struct hmm_range - track invalidation lock on virtual address range
159 *
Jérôme Glisse704f3f22019-05-13 17:19:48 -0700160 * @hmm: the core HMM structure this range is active against
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700161 * @vma: the vm area struct for the range
162 * @list: all range lock are on a list
163 * @start: range virtual start address (inclusive)
164 * @end: range virtual end address (exclusive)
165 * @pfns: array of pfns (big enough for the range)
166 * @flags: pfn flags to match device driver page table
167 * @values: pfn value for some special case (none, special, error, ...)
Jérôme Glisse023a0192019-05-13 17:20:05 -0700168 * @default_flags: default flags for the range (write, read, ... see hmm doc)
169 * @pfn_flags_mask: allows to mask pfn flags so that only default_flags matter
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700170 * @pfn_shifts: pfn shift value (should be <= PAGE_SHIFT)
171 * @valid: pfns array did not change since it has been fill by an HMM function
172 */
173struct hmm_range {
Jérôme Glisse704f3f22019-05-13 17:19:48 -0700174 struct hmm *hmm;
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700175 struct vm_area_struct *vma;
176 struct list_head list;
177 unsigned long start;
178 unsigned long end;
179 uint64_t *pfns;
180 const uint64_t *flags;
181 const uint64_t *values;
Jérôme Glisse023a0192019-05-13 17:20:05 -0700182 uint64_t default_flags;
183 uint64_t pfn_flags_mask;
Jérôme Glisse63d50662019-05-13 17:20:18 -0700184 uint8_t page_shift;
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700185 uint8_t pfn_shift;
186 bool valid;
187};
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700188
189/*
Jérôme Glisse63d50662019-05-13 17:20:18 -0700190 * hmm_range_page_shift() - return the page shift for the range
191 * @range: range being queried
192 * Returns: page shift (page size = 1 << page shift) for the range
193 */
194static inline unsigned hmm_range_page_shift(const struct hmm_range *range)
195{
196 return range->page_shift;
197}
198
199/*
200 * hmm_range_page_size() - return the page size for the range
201 * @range: range being queried
202 * Returns: page size for the range in bytes
203 */
204static inline unsigned long hmm_range_page_size(const struct hmm_range *range)
205{
206 return 1UL << hmm_range_page_shift(range);
207}
208
209/*
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700210 * hmm_range_wait_until_valid() - wait for range to be valid
211 * @range: range affected by invalidation to wait on
212 * @timeout: time out for wait in ms (ie abort wait after that period of time)
213 * Returns: true if the range is valid, false otherwise.
214 */
215static inline bool hmm_range_wait_until_valid(struct hmm_range *range,
216 unsigned long timeout)
217{
218 /* Check if mm is dead ? */
219 if (range->hmm == NULL || range->hmm->dead || range->hmm->mm == NULL) {
220 range->valid = false;
221 return false;
222 }
223 if (range->valid)
224 return true;
225 wait_event_timeout(range->hmm->wq, range->valid || range->hmm->dead,
226 msecs_to_jiffies(timeout));
227 /* Return current valid status just in case we get lucky */
228 return range->valid;
229}
230
231/*
232 * hmm_range_valid() - test if a range is valid or not
233 * @range: range
234 * Returns: true if the range is valid, false otherwise.
235 */
236static inline bool hmm_range_valid(struct hmm_range *range)
237{
238 return range->valid;
239}
240
241/*
Jérôme Glisseff05c0c2018-04-10 16:28:38 -0700242 * hmm_pfn_to_page() - return struct page pointed to by a valid HMM pfn
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700243 * @range: range use to decode HMM pfn value
Jérôme Glisseff05c0c2018-04-10 16:28:38 -0700244 * @pfn: HMM pfn value to get corresponding struct page from
245 * Returns: struct page pointer if pfn is a valid HMM pfn, NULL otherwise
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700246 *
Jérôme Glisseff05c0c2018-04-10 16:28:38 -0700247 * If the HMM pfn is valid (ie valid flag set) then return the struct page
248 * matching the pfn value stored in the HMM pfn. Otherwise return NULL.
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700249 */
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700250static inline struct page *hmm_pfn_to_page(const struct hmm_range *range,
251 uint64_t pfn)
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700252{
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700253 if (pfn == range->values[HMM_PFN_NONE])
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700254 return NULL;
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700255 if (pfn == range->values[HMM_PFN_ERROR])
256 return NULL;
257 if (pfn == range->values[HMM_PFN_SPECIAL])
258 return NULL;
259 if (!(pfn & range->flags[HMM_PFN_VALID]))
260 return NULL;
261 return pfn_to_page(pfn >> range->pfn_shift);
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700262}
263
264/*
Jérôme Glisseff05c0c2018-04-10 16:28:38 -0700265 * hmm_pfn_to_pfn() - return pfn value store in a HMM pfn
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700266 * @range: range use to decode HMM pfn value
Jérôme Glisseff05c0c2018-04-10 16:28:38 -0700267 * @pfn: HMM pfn value to extract pfn from
268 * Returns: pfn value if HMM pfn is valid, -1UL otherwise
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700269 */
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700270static inline unsigned long hmm_pfn_to_pfn(const struct hmm_range *range,
271 uint64_t pfn)
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700272{
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700273 if (pfn == range->values[HMM_PFN_NONE])
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700274 return -1UL;
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700275 if (pfn == range->values[HMM_PFN_ERROR])
276 return -1UL;
277 if (pfn == range->values[HMM_PFN_SPECIAL])
278 return -1UL;
279 if (!(pfn & range->flags[HMM_PFN_VALID]))
280 return -1UL;
281 return (pfn >> range->pfn_shift);
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700282}
283
284/*
Jérôme Glisseff05c0c2018-04-10 16:28:38 -0700285 * hmm_pfn_from_page() - create a valid HMM pfn value from struct page
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700286 * @range: range use to encode HMM pfn value
Jérôme Glisseff05c0c2018-04-10 16:28:38 -0700287 * @page: struct page pointer for which to create the HMM pfn
288 * Returns: valid HMM pfn for the page
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700289 */
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700290static inline uint64_t hmm_pfn_from_page(const struct hmm_range *range,
291 struct page *page)
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700292{
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700293 return (page_to_pfn(page) << range->pfn_shift) |
294 range->flags[HMM_PFN_VALID];
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700295}
296
297/*
Jérôme Glisseff05c0c2018-04-10 16:28:38 -0700298 * hmm_pfn_from_pfn() - create a valid HMM pfn value from pfn
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700299 * @range: range use to encode HMM pfn value
Jérôme Glisseff05c0c2018-04-10 16:28:38 -0700300 * @pfn: pfn value for which to create the HMM pfn
301 * Returns: valid HMM pfn for the pfn
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700302 */
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700303static inline uint64_t hmm_pfn_from_pfn(const struct hmm_range *range,
304 unsigned long pfn)
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700305{
Jérôme Glissef88a1e92018-04-10 16:29:06 -0700306 return (pfn << range->pfn_shift) |
307 range->flags[HMM_PFN_VALID];
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700308}
309
310
Jérôme Glissec0b12402017-09-08 16:11:27 -0700311#if IS_ENABLED(CONFIG_HMM_MIRROR)
312/*
313 * Mirroring: how to synchronize device page table with CPU page table.
314 *
315 * A device driver that is participating in HMM mirroring must always
316 * synchronize with CPU page table updates. For this, device drivers can either
317 * directly use mmu_notifier APIs or they can use the hmm_mirror API. Device
318 * drivers can decide to register one mirror per device per process, or just
319 * one mirror per process for a group of devices. The pattern is:
320 *
321 * int device_bind_address_space(..., struct mm_struct *mm, ...)
322 * {
323 * struct device_address_space *das;
324 *
325 * // Device driver specific initialization, and allocation of das
326 * // which contains an hmm_mirror struct as one of its fields.
327 * ...
328 *
329 * ret = hmm_mirror_register(&das->mirror, mm, &device_mirror_ops);
330 * if (ret) {
331 * // Cleanup on error
332 * return ret;
333 * }
334 *
335 * // Other device driver specific initialization
336 * ...
337 * }
338 *
339 * Once an hmm_mirror is registered for an address space, the device driver
340 * will get callbacks through sync_cpu_device_pagetables() operation (see
341 * hmm_mirror_ops struct).
342 *
343 * Device driver must not free the struct containing the hmm_mirror struct
344 * before calling hmm_mirror_unregister(). The expected usage is to do that when
345 * the device driver is unbinding from an address space.
346 *
347 *
348 * void device_unbind_address_space(struct device_address_space *das)
349 * {
350 * // Device driver specific cleanup
351 * ...
352 *
353 * hmm_mirror_unregister(&das->mirror);
354 *
355 * // Other device driver specific cleanup, and now das can be freed
356 * ...
357 * }
358 */
359
360struct hmm_mirror;
361
362/*
Jérôme Glisse44532d42018-10-30 15:04:24 -0700363 * enum hmm_update_event - type of update
Jérôme Glissec0b12402017-09-08 16:11:27 -0700364 * @HMM_UPDATE_INVALIDATE: invalidate range (no indication as to why)
365 */
Jérôme Glisse44532d42018-10-30 15:04:24 -0700366enum hmm_update_event {
Jérôme Glissec0b12402017-09-08 16:11:27 -0700367 HMM_UPDATE_INVALIDATE,
368};
369
370/*
Jérôme Glisse44532d42018-10-30 15:04:24 -0700371 * struct hmm_update - HMM update informations for callback
372 *
373 * @start: virtual start address of the range to update
374 * @end: virtual end address of the range to update
375 * @event: event triggering the update (what is happening)
376 * @blockable: can the callback block/sleep ?
377 */
378struct hmm_update {
379 unsigned long start;
380 unsigned long end;
381 enum hmm_update_event event;
382 bool blockable;
383};
384
385/*
Jérôme Glissec0b12402017-09-08 16:11:27 -0700386 * struct hmm_mirror_ops - HMM mirror device operations callback
387 *
388 * @update: callback to update range on a device
389 */
390struct hmm_mirror_ops {
Ralph Campbelle1401512018-04-10 16:28:19 -0700391 /* release() - release hmm_mirror
392 *
393 * @mirror: pointer to struct hmm_mirror
394 *
395 * This is called when the mm_struct is being released.
396 * The callback should make sure no references to the mirror occur
397 * after the callback returns.
398 */
399 void (*release)(struct hmm_mirror *mirror);
400
Jérôme Glissec0b12402017-09-08 16:11:27 -0700401 /* sync_cpu_device_pagetables() - synchronize page tables
402 *
403 * @mirror: pointer to struct hmm_mirror
Jérôme Glisse44532d42018-10-30 15:04:24 -0700404 * @update: update informations (see struct hmm_update)
405 * Returns: -EAGAIN if update.blockable false and callback need to
406 * block, 0 otherwise.
Jérôme Glissec0b12402017-09-08 16:11:27 -0700407 *
408 * This callback ultimately originates from mmu_notifiers when the CPU
409 * page table is updated. The device driver must update its page table
410 * in response to this callback. The update argument tells what action
411 * to perform.
412 *
413 * The device driver must not return from this callback until the device
414 * page tables are completely updated (TLBs flushed, etc); this is a
415 * synchronous call.
416 */
Jérôme Glisse44532d42018-10-30 15:04:24 -0700417 int (*sync_cpu_device_pagetables)(struct hmm_mirror *mirror,
418 const struct hmm_update *update);
Jérôme Glissec0b12402017-09-08 16:11:27 -0700419};
420
421/*
422 * struct hmm_mirror - mirror struct for a device driver
423 *
424 * @hmm: pointer to struct hmm (which is unique per mm_struct)
425 * @ops: device driver callback for HMM mirror operations
426 * @list: for list of mirrors of a given mm
427 *
428 * Each address space (mm_struct) being mirrored by a device must register one
429 * instance of an hmm_mirror struct with HMM. HMM will track the list of all
430 * mirrors for each mm_struct.
431 */
432struct hmm_mirror {
433 struct hmm *hmm;
434 const struct hmm_mirror_ops *ops;
435 struct list_head list;
436};
437
438int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm);
439void hmm_mirror_unregister(struct hmm_mirror *mirror);
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700440
441
442/*
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700443 * Please see Documentation/vm/hmm.rst for how to use the range API.
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700444 */
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700445int hmm_range_register(struct hmm_range *range,
446 struct mm_struct *mm,
447 unsigned long start,
Jérôme Glisse63d50662019-05-13 17:20:18 -0700448 unsigned long end,
449 unsigned page_shift);
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700450void hmm_range_unregister(struct hmm_range *range);
Jérôme Glisse25f23a02019-05-13 17:19:55 -0700451long hmm_range_snapshot(struct hmm_range *range);
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700452long hmm_range_fault(struct hmm_range *range, bool block);
Jérôme Glisse74eee182017-09-08 16:11:35 -0700453
454/*
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700455 * HMM_RANGE_DEFAULT_TIMEOUT - default timeout (ms) when waiting for a range
Jérôme Glisse74eee182017-09-08 16:11:35 -0700456 *
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700457 * When waiting for mmu notifiers we need some kind of time out otherwise we
458 * could potentialy wait for ever, 1000ms ie 1s sounds like a long time to
459 * wait already.
Jérôme Glisse74eee182017-09-08 16:11:35 -0700460 */
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700461#define HMM_RANGE_DEFAULT_TIMEOUT 1000
462
463/* This is a temporary helper to avoid merge conflict between trees. */
464static inline bool hmm_vma_range_done(struct hmm_range *range)
465{
466 bool ret = hmm_range_valid(range);
467
468 hmm_range_unregister(range);
469 return ret;
470}
Jérôme Glisse73231612019-05-13 17:19:58 -0700471
472/* This is a temporary helper to avoid merge conflict between trees. */
473static inline int hmm_vma_fault(struct hmm_range *range, bool block)
474{
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700475 long ret;
476
Jérôme Glisse023a0192019-05-13 17:20:05 -0700477 /*
478 * With the old API the driver must set each individual entries with
479 * the requested flags (valid, write, ...). So here we set the mask to
480 * keep intact the entries provided by the driver and zero out the
481 * default_flags.
482 */
483 range->default_flags = 0;
484 range->pfn_flags_mask = -1UL;
485
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700486 ret = hmm_range_register(range, range->vma->vm_mm,
Jérôme Glisse63d50662019-05-13 17:20:18 -0700487 range->start, range->end,
488 PAGE_SHIFT);
Jérôme Glissea3e0d412019-05-13 17:20:01 -0700489 if (ret)
490 return (int)ret;
491
492 if (!hmm_range_wait_until_valid(range, HMM_RANGE_DEFAULT_TIMEOUT)) {
493 /*
494 * The mmap_sem was taken by driver we release it here and
495 * returns -EAGAIN which correspond to mmap_sem have been
496 * drop in the old API.
497 */
498 up_read(&range->vma->vm_mm->mmap_sem);
499 return -EAGAIN;
500 }
501
502 ret = hmm_range_fault(range, block);
503 if (ret <= 0) {
504 if (ret == -EBUSY || !ret) {
505 /* Same as above drop mmap_sem to match old API. */
506 up_read(&range->vma->vm_mm->mmap_sem);
507 ret = -EBUSY;
508 } else if (ret == -EAGAIN)
509 ret = -EBUSY;
510 hmm_range_unregister(range);
511 return ret;
512 }
513 return 0;
Jérôme Glisse73231612019-05-13 17:19:58 -0700514}
Jérôme Glissec0b12402017-09-08 16:11:27 -0700515
Arnd Bergmann9d8a4632018-04-10 16:29:13 -0700516/* Below are for HMM internal use only! Not to be used by device driver! */
517void hmm_mm_destroy(struct mm_struct *mm);
518
519static inline void hmm_mm_init(struct mm_struct *mm)
520{
521 mm->hmm = NULL;
522}
523#else /* IS_ENABLED(CONFIG_HMM_MIRROR) */
524static inline void hmm_mm_destroy(struct mm_struct *mm) {}
525static inline void hmm_mm_init(struct mm_struct *mm) {}
526#endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
Jérôme Glissec0b12402017-09-08 16:11:27 -0700527
Jérôme Glissedf6ad692017-09-08 16:12:24 -0700528#if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC)
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700529struct hmm_devmem;
530
531struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
532 unsigned long addr);
533
534/*
535 * struct hmm_devmem_ops - callback for ZONE_DEVICE memory events
536 *
537 * @free: call when refcount on page reach 1 and thus is no longer use
538 * @fault: call when there is a page fault to unaddressable memory
539 *
540 * Both callback happens from page_free() and page_fault() callback of struct
541 * dev_pagemap respectively. See include/linux/memremap.h for more details on
542 * those.
543 *
544 * The hmm_devmem_ops callback are just here to provide a coherent and
545 * uniq API to device driver and device driver should not register their
546 * own page_free() or page_fault() but rely on the hmm_devmem_ops call-
547 * back.
548 */
549struct hmm_devmem_ops {
550 /*
551 * free() - free a device page
552 * @devmem: device memory structure (see struct hmm_devmem)
553 * @page: pointer to struct page being freed
554 *
555 * Call back occurs whenever a device page refcount reach 1 which
556 * means that no one is holding any reference on the page anymore
557 * (ZONE_DEVICE page have an elevated refcount of 1 as default so
558 * that they are not release to the general page allocator).
559 *
560 * Note that callback has exclusive ownership of the page (as no
561 * one is holding any reference).
562 */
563 void (*free)(struct hmm_devmem *devmem, struct page *page);
564 /*
565 * fault() - CPU page fault or get user page (GUP)
566 * @devmem: device memory structure (see struct hmm_devmem)
567 * @vma: virtual memory area containing the virtual address
568 * @addr: virtual address that faulted or for which there is a GUP
569 * @page: pointer to struct page backing virtual address (unreliable)
570 * @flags: FAULT_FLAG_* (see include/linux/mm.h)
571 * @pmdp: page middle directory
572 * Returns: VM_FAULT_MINOR/MAJOR on success or one of VM_FAULT_ERROR
573 * on error
574 *
575 * The callback occurs whenever there is a CPU page fault or GUP on a
576 * virtual address. This means that the device driver must migrate the
577 * page back to regular memory (CPU accessible).
578 *
579 * The device driver is free to migrate more than one page from the
580 * fault() callback as an optimization. However if device decide to
581 * migrate more than one page it must always priotirize the faulting
582 * address over the others.
583 *
584 * The struct page pointer is only given as an hint to allow quick
585 * lookup of internal device driver data. A concurrent migration
586 * might have already free that page and the virtual address might
587 * not longer be back by it. So it should not be modified by the
588 * callback.
589 *
590 * Note that mmap semaphore is held in read mode at least when this
591 * callback occurs, hence the vma is valid upon callback entry.
592 */
Souptick Joarderb57e622e62019-03-11 23:28:10 -0700593 vm_fault_t (*fault)(struct hmm_devmem *devmem,
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700594 struct vm_area_struct *vma,
595 unsigned long addr,
596 const struct page *page,
597 unsigned int flags,
598 pmd_t *pmdp);
599};
600
601/*
602 * struct hmm_devmem - track device memory
603 *
604 * @completion: completion object for device memory
605 * @pfn_first: first pfn for this resource (set by hmm_devmem_add())
606 * @pfn_last: last pfn for this resource (set by hmm_devmem_add())
607 * @resource: IO resource reserved for this chunk of memory
608 * @pagemap: device page map for that chunk
609 * @device: device to bind resource to
610 * @ops: memory operations callback
611 * @ref: per CPU refcount
Dan Williams063a7d12018-12-28 00:39:46 -0800612 * @page_fault: callback when CPU fault on an unaddressable device page
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700613 *
614 * This an helper structure for device drivers that do not wish to implement
615 * the gory details related to hotplugging new memoy and allocating struct
616 * pages.
617 *
618 * Device drivers can directly use ZONE_DEVICE memory on their own if they
619 * wish to do so.
Dan Williams063a7d12018-12-28 00:39:46 -0800620 *
621 * The page_fault() callback must migrate page back, from device memory to
622 * system memory, so that the CPU can access it. This might fail for various
623 * reasons (device issues, device have been unplugged, ...). When such error
624 * conditions happen, the page_fault() callback must return VM_FAULT_SIGBUS and
625 * set the CPU page table entry to "poisoned".
626 *
627 * Note that because memory cgroup charges are transferred to the device memory,
628 * this should never fail due to memory restrictions. However, allocation
629 * of a regular system page might still fail because we are out of memory. If
630 * that happens, the page_fault() callback must return VM_FAULT_OOM.
631 *
632 * The page_fault() callback can also try to migrate back multiple pages in one
633 * chunk, as an optimization. It must, however, prioritize the faulting address
634 * over all the others.
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700635 */
Souptick Joarderb57e622e62019-03-11 23:28:10 -0700636typedef vm_fault_t (*dev_page_fault_t)(struct vm_area_struct *vma,
Dan Williams063a7d12018-12-28 00:39:46 -0800637 unsigned long addr,
638 const struct page *page,
639 unsigned int flags,
640 pmd_t *pmdp);
641
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700642struct hmm_devmem {
643 struct completion completion;
644 unsigned long pfn_first;
645 unsigned long pfn_last;
646 struct resource *resource;
647 struct device *device;
648 struct dev_pagemap pagemap;
649 const struct hmm_devmem_ops *ops;
650 struct percpu_ref ref;
Dan Williams063a7d12018-12-28 00:39:46 -0800651 dev_page_fault_t page_fault;
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700652};
653
654/*
655 * To add (hotplug) device memory, HMM assumes that there is no real resource
656 * that reserves a range in the physical address space (this is intended to be
657 * use by unaddressable device memory). It will reserve a physical range big
658 * enough and allocate struct page for it.
659 *
660 * The device driver can wrap the hmm_devmem struct inside a private device
Dan Williams58ef15b2018-12-28 00:35:07 -0800661 * driver struct.
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700662 */
663struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
664 struct device *device,
665 unsigned long size);
Jérôme Glissed3df0a42017-09-08 16:12:28 -0700666struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
667 struct device *device,
668 struct resource *res);
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700669
670/*
671 * hmm_devmem_page_set_drvdata - set per-page driver data field
672 *
673 * @page: pointer to struct page
674 * @data: driver data value to set
675 *
676 * Because page can not be on lru we have an unsigned long that driver can use
677 * to store a per page field. This just a simple helper to do that.
678 */
679static inline void hmm_devmem_page_set_drvdata(struct page *page,
680 unsigned long data)
681{
Matthew Wilcox50e7fbc2018-06-07 17:09:01 -0700682 page->hmm_data = data;
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700683}
684
685/*
686 * hmm_devmem_page_get_drvdata - get per page driver data field
687 *
688 * @page: pointer to struct page
689 * Return: driver data value
690 */
Ralph Campbell0bea8032017-11-15 17:34:00 -0800691static inline unsigned long hmm_devmem_page_get_drvdata(const struct page *page)
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700692{
Matthew Wilcox50e7fbc2018-06-07 17:09:01 -0700693 return page->hmm_data;
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700694}
Jérôme Glisse858b54d2017-09-08 16:12:02 -0700695
696
697/*
698 * struct hmm_device - fake device to hang device memory onto
699 *
700 * @device: device struct
701 * @minor: device minor number
702 */
703struct hmm_device {
704 struct device device;
705 unsigned int minor;
706};
707
708/*
709 * A device driver that wants to handle multiple devices memory through a
710 * single fake device can use hmm_device to do so. This is purely a helper and
711 * it is not strictly needed, in order to make use of any HMM functionality.
712 */
713struct hmm_device *hmm_device_new(void *drvdata);
714void hmm_device_put(struct hmm_device *hmm_device);
Jérôme Glissedf6ad692017-09-08 16:12:24 -0700715#endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
Jérôme Glisse6b368cd2017-09-08 16:12:32 -0700716#else /* IS_ENABLED(CONFIG_HMM) */
717static inline void hmm_mm_destroy(struct mm_struct *mm) {}
718static inline void hmm_mm_init(struct mm_struct *mm) {}
Jérôme Glisseb28b08d2018-04-10 16:28:15 -0700719#endif /* IS_ENABLED(CONFIG_HMM) */
Arnd Bergmann9d8a4632018-04-10 16:29:13 -0700720
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700721#endif /* LINUX_HMM_H */