Commit | Line | Data |
---|---|---|
c942fddf | 1 | /* SPDX-License-Identifier: GPL-2.0-or-later */ |
133ff0ea JG |
2 | /* |
3 | * Copyright 2013 Red Hat Inc. | |
4 | * | |
f813f219 | 5 | * Authors: Jérôme Glisse <jglisse@redhat.com> |
133ff0ea JG |
6 | */ |
7 | /* | |
8 | * Heterogeneous Memory Management (HMM) | |
9 | * | |
ad56b738 | 10 | * See Documentation/vm/hmm.rst for reasons and overview of what HMM is and it |
133ff0ea JG |
11 | * is for. Here we focus on the HMM API description, with some explanation of |
12 | * the underlying implementation. | |
13 | * | |
14 | * Short description: HMM provides a set of helpers to share a virtual address | |
15 | * space between CPU and a device, so that the device can access any valid | |
16 | * address of the process (while still obeying memory protection). HMM also | |
17 | * provides helpers to migrate process memory to device memory, and back. Each | |
18 | * set of functionality (address space mirroring, and migration to and from | |
19 | * device memory) can be used independently of the other. | |
20 | * | |
21 | * | |
22 | * HMM address space mirroring API: | |
23 | * | |
085ea250 RC |
24 | * Use HMM address space mirroring if you want to mirror a range of the CPU |
25 | * page tables of a process into a device page table. Here, "mirror" means "keep | |
133ff0ea JG |
26 | * synchronized". Prerequisites: the device must provide the ability to write- |
27 | * protect its page tables (at PAGE_SIZE granularity), and must be able to | |
28 | * recover from the resulting potential page faults. | |
29 | * | |
30 | * HMM guarantees that at any point in time, a given virtual address points to | |
31 | * either the same memory in both CPU and device page tables (that is: CPU and | |
32 | * device page tables each point to the same pages), or that one page table (CPU | |
33 | * or device) points to no entry, while the other still points to the old page | |
34 | * for the address. The latter case happens when the CPU page table update | |
35 | * happens first, and then the update is mirrored over to the device page table. | |
36 | * This does not cause any issue, because the CPU page table cannot start | |
37 | * pointing to a new page until the device page table is invalidated. | |
38 | * | |
39 | * HMM uses mmu_notifiers to monitor the CPU page tables, and forwards any | |
40 | * updates to each device driver that has registered a mirror. It also provides | |
41 | * some API calls to help with taking a snapshot of the CPU page table, and to | |
42 | * synchronize with any updates that might happen concurrently. | |
43 | * | |
44 | * | |
45 | * HMM migration to and from device memory: | |
46 | * | |
47 | * HMM provides a set of helpers to hotplug device memory as ZONE_DEVICE, with | |
48 | * a new MEMORY_DEVICE_PRIVATE type. This provides a struct page for each page | |
49 | * of the device memory, and allows the device driver to manage its memory | |
50 | * using those struct pages. Having struct pages for device memory makes | |
51 | * migration easier. Because that memory is not addressable by the CPU it must | |
52 | * never be pinned to the device; in other words, any CPU page fault can always | |
53 | * cause the device memory to be migrated (copied/moved) back to regular memory. | |
54 | * | |
55 | * A new migrate helper (migrate_vma()) has been added (see mm/migrate.c) that | |
56 | * allows use of a device DMA engine to perform the copy operation between | |
57 | * regular system memory and device memory. | |
58 | */ | |
59 | #ifndef LINUX_HMM_H | |
60 | #define LINUX_HMM_H | |
61 | ||
62 | #include <linux/kconfig.h> | |
063a7d1d | 63 | #include <asm/pgtable.h> |
133ff0ea | 64 | |
43535b0a | 65 | #ifdef CONFIG_HMM_MIRROR |
133ff0ea | 66 | |
858b54da | 67 | #include <linux/device.h> |
4ef589dc JG |
68 | #include <linux/migrate.h> |
69 | #include <linux/memremap.h> | |
70 | #include <linux/completion.h> | |
a3e0d41c | 71 | #include <linux/mmu_notifier.h> |
4ef589dc | 72 | |
a3e0d41c JG |
73 | |
74 | /* | |
75 | * struct hmm - HMM per mm struct | |
76 | * | |
77 | * @mm: mm struct this HMM struct is bound to | |
78 | * @lock: lock protecting ranges list | |
79 | * @ranges: list of range being snapshotted | |
80 | * @mirrors: list of mirrors for this mm | |
81 | * @mmu_notifier: mmu notifier to track updates to CPU page table | |
82 | * @mirrors_sem: read/write semaphore protecting the mirrors list | |
83 | * @wq: wait queue for user waiting on a range invalidation | |
84 | * @notifiers: count of active mmu notifiers | |
a3e0d41c JG |
85 | */ |
86 | struct hmm { | |
c7d8b782 | 87 | struct mmu_notifier mmu_notifier; |
5a136b4a | 88 | spinlock_t ranges_lock; |
a3e0d41c JG |
89 | struct list_head ranges; |
90 | struct list_head mirrors; | |
a3e0d41c JG |
91 | struct rw_semaphore mirrors_sem; |
92 | wait_queue_head_t wq; | |
93 | long notifiers; | |
a3e0d41c | 94 | }; |
133ff0ea JG |
95 | |
96 | /* | |
f88a1e90 JG |
97 | * hmm_pfn_flag_e - HMM flag enums |
98 | * | |
133ff0ea | 99 | * Flags: |
86586a41 | 100 | * HMM_PFN_VALID: pfn is valid. It has, at least, read permission. |
133ff0ea | 101 | * HMM_PFN_WRITE: CPU page table has write permission set |
f88a1e90 JG |
102 | * HMM_PFN_DEVICE_PRIVATE: private device memory (ZONE_DEVICE) |
103 | * | |
085ea250 RC |
104 | * The driver provides a flags array for mapping page protections to device |
105 | * PTE bits. If the driver valid bit for an entry is bit 3, | |
106 | * i.e., (entry & (1 << 3)), then the driver must provide | |
f88a1e90 | 107 | * an array in hmm_range.flags with hmm_range.flags[HMM_PFN_VALID] == 1 << 3. |
085ea250 | 108 | * Same logic apply to all flags. This is the same idea as vm_page_prot in vma |
f88a1e90 JG |
109 | * except that this is per device driver rather than per architecture. |
110 | */ | |
111 | enum hmm_pfn_flag_e { | |
112 | HMM_PFN_VALID = 0, | |
113 | HMM_PFN_WRITE, | |
114 | HMM_PFN_DEVICE_PRIVATE, | |
115 | HMM_PFN_FLAG_MAX | |
116 | }; | |
117 | ||
118 | /* | |
119 | * hmm_pfn_value_e - HMM pfn special value | |
120 | * | |
121 | * Flags: | |
da4c3c73 | 122 | * HMM_PFN_ERROR: corresponding CPU page table entry points to poisoned memory |
f88a1e90 | 123 | * HMM_PFN_NONE: corresponding CPU page table entry is pte_none() |
da4c3c73 | 124 | * HMM_PFN_SPECIAL: corresponding CPU page table entry is special; i.e., the |
67fa1666 | 125 | * result of vmf_insert_pfn() or vm_insert_page(). Therefore, it should not |
da4c3c73 JG |
126 | * be mirrored by a device, because the entry will never have HMM_PFN_VALID |
127 | * set and the pfn value is undefined. | |
f88a1e90 | 128 | * |
085ea250 RC |
129 | * Driver provides values for none entry, error entry, and special entry. |
130 | * Driver can alias (i.e., use same value) error and special, but | |
131 | * it should not alias none with error or special. | |
f88a1e90 JG |
132 | * |
133 | * HMM pfn value returned by hmm_vma_get_pfns() or hmm_vma_fault() will be: | |
134 | * hmm_range.values[HMM_PFN_ERROR] if CPU page table entry is poisonous, | |
085ea250 | 135 | * hmm_range.values[HMM_PFN_NONE] if there is no CPU page table entry, |
f88a1e90 | 136 | * hmm_range.values[HMM_PFN_SPECIAL] if CPU page table entry is a special one |
133ff0ea | 137 | */ |
f88a1e90 JG |
138 | enum hmm_pfn_value_e { |
139 | HMM_PFN_ERROR, | |
140 | HMM_PFN_NONE, | |
141 | HMM_PFN_SPECIAL, | |
142 | HMM_PFN_VALUE_MAX | |
143 | }; | |
144 | ||
145 | /* | |
146 | * struct hmm_range - track invalidation lock on virtual address range | |
147 | * | |
704f3f2c | 148 | * @hmm: the core HMM structure this range is active against |
f88a1e90 JG |
149 | * @vma: the vm area struct for the range |
150 | * @list: all range lock are on a list | |
151 | * @start: range virtual start address (inclusive) | |
152 | * @end: range virtual end address (exclusive) | |
153 | * @pfns: array of pfns (big enough for the range) | |
154 | * @flags: pfn flags to match device driver page table | |
155 | * @values: pfn value for some special case (none, special, error, ...) | |
023a019a JG |
156 | * @default_flags: default flags for the range (write, read, ... see hmm doc) |
157 | * @pfn_flags_mask: allows to mask pfn flags so that only default_flags matter | |
f88a1e90 JG |
158 | * @pfn_shifts: pfn shift value (should be <= PAGE_SHIFT) |
159 | * @valid: pfns array did not change since it has been fill by an HMM function | |
160 | */ | |
161 | struct hmm_range { | |
704f3f2c | 162 | struct hmm *hmm; |
f88a1e90 JG |
163 | struct list_head list; |
164 | unsigned long start; | |
165 | unsigned long end; | |
166 | uint64_t *pfns; | |
167 | const uint64_t *flags; | |
168 | const uint64_t *values; | |
023a019a JG |
169 | uint64_t default_flags; |
170 | uint64_t pfn_flags_mask; | |
f88a1e90 JG |
171 | uint8_t pfn_shift; |
172 | bool valid; | |
173 | }; | |
133ff0ea | 174 | |
a3e0d41c JG |
175 | /* |
176 | * hmm_range_wait_until_valid() - wait for range to be valid | |
177 | * @range: range affected by invalidation to wait on | |
178 | * @timeout: time out for wait in ms (ie abort wait after that period of time) | |
085ea250 | 179 | * Return: true if the range is valid, false otherwise. |
a3e0d41c JG |
180 | */ |
181 | static inline bool hmm_range_wait_until_valid(struct hmm_range *range, | |
182 | unsigned long timeout) | |
183 | { | |
378a6040 JG |
184 | return wait_event_timeout(range->hmm->wq, range->valid, |
185 | msecs_to_jiffies(timeout)) != 0; | |
a3e0d41c JG |
186 | } |
187 | ||
188 | /* | |
189 | * hmm_range_valid() - test if a range is valid or not | |
190 | * @range: range | |
085ea250 | 191 | * Return: true if the range is valid, false otherwise. |
a3e0d41c JG |
192 | */ |
193 | static inline bool hmm_range_valid(struct hmm_range *range) | |
194 | { | |
195 | return range->valid; | |
196 | } | |
197 | ||
133ff0ea | 198 | /* |
391aab11 JG |
199 | * hmm_device_entry_to_page() - return struct page pointed to by a device entry |
200 | * @range: range use to decode device entry value | |
201 | * @entry: device entry value to get corresponding struct page from | |
085ea250 | 202 | * Return: struct page pointer if entry is a valid, NULL otherwise |
133ff0ea | 203 | * |
391aab11 JG |
204 | * If the device entry is valid (ie valid flag set) then return the struct page |
205 | * matching the entry value. Otherwise return NULL. | |
133ff0ea | 206 | */ |
391aab11 JG |
207 | static inline struct page *hmm_device_entry_to_page(const struct hmm_range *range, |
208 | uint64_t entry) | |
133ff0ea | 209 | { |
391aab11 | 210 | if (entry == range->values[HMM_PFN_NONE]) |
f88a1e90 | 211 | return NULL; |
391aab11 | 212 | if (entry == range->values[HMM_PFN_ERROR]) |
f88a1e90 | 213 | return NULL; |
391aab11 | 214 | if (entry == range->values[HMM_PFN_SPECIAL]) |
133ff0ea | 215 | return NULL; |
391aab11 | 216 | if (!(entry & range->flags[HMM_PFN_VALID])) |
f88a1e90 | 217 | return NULL; |
391aab11 | 218 | return pfn_to_page(entry >> range->pfn_shift); |
133ff0ea JG |
219 | } |
220 | ||
221 | /* | |
391aab11 JG |
222 | * hmm_device_entry_to_pfn() - return pfn value store in a device entry |
223 | * @range: range use to decode device entry value | |
224 | * @entry: device entry to extract pfn from | |
085ea250 | 225 | * Return: pfn value if device entry is valid, -1UL otherwise |
133ff0ea | 226 | */ |
391aab11 JG |
227 | static inline unsigned long |
228 | hmm_device_entry_to_pfn(const struct hmm_range *range, uint64_t pfn) | |
133ff0ea | 229 | { |
f88a1e90 JG |
230 | if (pfn == range->values[HMM_PFN_NONE]) |
231 | return -1UL; | |
232 | if (pfn == range->values[HMM_PFN_ERROR]) | |
233 | return -1UL; | |
234 | if (pfn == range->values[HMM_PFN_SPECIAL]) | |
133ff0ea | 235 | return -1UL; |
f88a1e90 JG |
236 | if (!(pfn & range->flags[HMM_PFN_VALID])) |
237 | return -1UL; | |
238 | return (pfn >> range->pfn_shift); | |
133ff0ea JG |
239 | } |
240 | ||
241 | /* | |
391aab11 | 242 | * hmm_device_entry_from_page() - create a valid device entry for a page |
f88a1e90 | 243 | * @range: range use to encode HMM pfn value |
391aab11 | 244 | * @page: page for which to create the device entry |
085ea250 | 245 | * Return: valid device entry for the page |
133ff0ea | 246 | */ |
391aab11 JG |
247 | static inline uint64_t hmm_device_entry_from_page(const struct hmm_range *range, |
248 | struct page *page) | |
133ff0ea | 249 | { |
f88a1e90 JG |
250 | return (page_to_pfn(page) << range->pfn_shift) | |
251 | range->flags[HMM_PFN_VALID]; | |
133ff0ea JG |
252 | } |
253 | ||
254 | /* | |
391aab11 | 255 | * hmm_device_entry_from_pfn() - create a valid device entry value from pfn |
f88a1e90 | 256 | * @range: range use to encode HMM pfn value |
391aab11 | 257 | * @pfn: pfn value for which to create the device entry |
085ea250 | 258 | * Return: valid device entry for the pfn |
133ff0ea | 259 | */ |
391aab11 JG |
260 | static inline uint64_t hmm_device_entry_from_pfn(const struct hmm_range *range, |
261 | unsigned long pfn) | |
133ff0ea | 262 | { |
f88a1e90 JG |
263 | return (pfn << range->pfn_shift) | |
264 | range->flags[HMM_PFN_VALID]; | |
133ff0ea JG |
265 | } |
266 | ||
c0b12405 JG |
267 | /* |
268 | * Mirroring: how to synchronize device page table with CPU page table. | |
269 | * | |
270 | * A device driver that is participating in HMM mirroring must always | |
271 | * synchronize with CPU page table updates. For this, device drivers can either | |
272 | * directly use mmu_notifier APIs or they can use the hmm_mirror API. Device | |
273 | * drivers can decide to register one mirror per device per process, or just | |
274 | * one mirror per process for a group of devices. The pattern is: | |
275 | * | |
276 | * int device_bind_address_space(..., struct mm_struct *mm, ...) | |
277 | * { | |
278 | * struct device_address_space *das; | |
279 | * | |
280 | * // Device driver specific initialization, and allocation of das | |
281 | * // which contains an hmm_mirror struct as one of its fields. | |
282 | * ... | |
283 | * | |
284 | * ret = hmm_mirror_register(&das->mirror, mm, &device_mirror_ops); | |
285 | * if (ret) { | |
286 | * // Cleanup on error | |
287 | * return ret; | |
288 | * } | |
289 | * | |
290 | * // Other device driver specific initialization | |
291 | * ... | |
292 | * } | |
293 | * | |
294 | * Once an hmm_mirror is registered for an address space, the device driver | |
295 | * will get callbacks through sync_cpu_device_pagetables() operation (see | |
296 | * hmm_mirror_ops struct). | |
297 | * | |
298 | * Device driver must not free the struct containing the hmm_mirror struct | |
299 | * before calling hmm_mirror_unregister(). The expected usage is to do that when | |
300 | * the device driver is unbinding from an address space. | |
301 | * | |
302 | * | |
303 | * void device_unbind_address_space(struct device_address_space *das) | |
304 | * { | |
305 | * // Device driver specific cleanup | |
306 | * ... | |
307 | * | |
308 | * hmm_mirror_unregister(&das->mirror); | |
309 | * | |
310 | * // Other device driver specific cleanup, and now das can be freed | |
311 | * ... | |
312 | * } | |
313 | */ | |
314 | ||
315 | struct hmm_mirror; | |
316 | ||
c0b12405 JG |
317 | /* |
318 | * struct hmm_mirror_ops - HMM mirror device operations callback | |
319 | * | |
320 | * @update: callback to update range on a device | |
321 | */ | |
322 | struct hmm_mirror_ops { | |
e1401513 RC |
323 | /* release() - release hmm_mirror |
324 | * | |
325 | * @mirror: pointer to struct hmm_mirror | |
326 | * | |
2076e5c0 RC |
327 | * This is called when the mm_struct is being released. The callback |
328 | * must ensure that all access to any pages obtained from this mirror | |
329 | * is halted before the callback returns. All future access should | |
330 | * fault. | |
e1401513 RC |
331 | */ |
332 | void (*release)(struct hmm_mirror *mirror); | |
333 | ||
c0b12405 JG |
334 | /* sync_cpu_device_pagetables() - synchronize page tables |
335 | * | |
336 | * @mirror: pointer to struct hmm_mirror | |
1f961807 RC |
337 | * @update: update information (see struct mmu_notifier_range) |
338 | * Return: -EAGAIN if mmu_notifier_range_blockable(update) is false | |
339 | * and callback needs to block, 0 otherwise. | |
c0b12405 JG |
340 | * |
341 | * This callback ultimately originates from mmu_notifiers when the CPU | |
342 | * page table is updated. The device driver must update its page table | |
343 | * in response to this callback. The update argument tells what action | |
344 | * to perform. | |
345 | * | |
346 | * The device driver must not return from this callback until the device | |
347 | * page tables are completely updated (TLBs flushed, etc); this is a | |
348 | * synchronous call. | |
349 | */ | |
1f961807 RC |
350 | int (*sync_cpu_device_pagetables)( |
351 | struct hmm_mirror *mirror, | |
352 | const struct mmu_notifier_range *update); | |
c0b12405 JG |
353 | }; |
354 | ||
355 | /* | |
356 | * struct hmm_mirror - mirror struct for a device driver | |
357 | * | |
358 | * @hmm: pointer to struct hmm (which is unique per mm_struct) | |
359 | * @ops: device driver callback for HMM mirror operations | |
360 | * @list: for list of mirrors of a given mm | |
361 | * | |
362 | * Each address space (mm_struct) being mirrored by a device must register one | |
363 | * instance of an hmm_mirror struct with HMM. HMM will track the list of all | |
364 | * mirrors for each mm_struct. | |
365 | */ | |
366 | struct hmm_mirror { | |
367 | struct hmm *hmm; | |
368 | const struct hmm_mirror_ops *ops; | |
369 | struct list_head list; | |
370 | }; | |
371 | ||
372 | int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm); | |
373 | void hmm_mirror_unregister(struct hmm_mirror *mirror); | |
da4c3c73 | 374 | |
da4c3c73 | 375 | /* |
a3e0d41c | 376 | * Please see Documentation/vm/hmm.rst for how to use the range API. |
da4c3c73 | 377 | */ |
fac555ac | 378 | int hmm_range_register(struct hmm_range *range, struct hmm_mirror *mirror); |
a3e0d41c | 379 | void hmm_range_unregister(struct hmm_range *range); |
9a4903e4 CH |
380 | |
381 | /* | |
382 | * Retry fault if non-blocking, drop mmap_sem and return -EAGAIN in that case. | |
383 | */ | |
384 | #define HMM_FAULT_ALLOW_RETRY (1 << 0) | |
385 | ||
d45d464b CH |
386 | /* Don't fault in missing PTEs, just snapshot the current state. */ |
387 | #define HMM_FAULT_SNAPSHOT (1 << 1) | |
388 | ||
9a4903e4 CH |
389 | long hmm_range_fault(struct hmm_range *range, unsigned int flags); |
390 | ||
55c0ece8 JG |
391 | long hmm_range_dma_map(struct hmm_range *range, |
392 | struct device *device, | |
393 | dma_addr_t *daddrs, | |
9a4903e4 | 394 | unsigned int flags); |
55c0ece8 | 395 | long hmm_range_dma_unmap(struct hmm_range *range, |
55c0ece8 JG |
396 | struct device *device, |
397 | dma_addr_t *daddrs, | |
398 | bool dirty); | |
74eee180 JG |
399 | |
400 | /* | |
a3e0d41c | 401 | * HMM_RANGE_DEFAULT_TIMEOUT - default timeout (ms) when waiting for a range |
74eee180 | 402 | * |
a3e0d41c JG |
403 | * When waiting for mmu notifiers we need some kind of time out otherwise we |
404 | * could potentialy wait for ever, 1000ms ie 1s sounds like a long time to | |
405 | * wait already. | |
74eee180 | 406 | */ |
a3e0d41c JG |
407 | #define HMM_RANGE_DEFAULT_TIMEOUT 1000 |
408 | ||
9d8a463a | 409 | #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */ |
c0b12405 | 410 | |
133ff0ea | 411 | #endif /* LINUX_HMM_H */ |