2 * Copyright 2013 Red Hat Inc.
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.
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.
14 * Authors: Jérôme Glisse <jglisse@redhat.com>
17 * Heterogeneous Memory Management (HMM)
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.
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.
31 * HMM address space mirroring API:
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.
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.
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.
54 * HMM migration to and from device memory:
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.
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.
71 #include <linux/kconfig.h>
72 #include <asm/pgtable.h>
74 #if IS_ENABLED(CONFIG_HMM)
76 #include <linux/device.h>
77 #include <linux/migrate.h>
78 #include <linux/memremap.h>
79 #include <linux/completion.h>
80 #include <linux/mmu_notifier.h>
84 * struct hmm - HMM per mm struct
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 ?
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;
110 * hmm_pfn_flag_e - HMM flag enums
113 * HMM_PFN_VALID: pfn is valid. It has, at least, read permission.
114 * HMM_PFN_WRITE: CPU page table has write permission set
115 * HMM_PFN_DEVICE_PRIVATE: private device memory (ZONE_DEVICE)
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.
123 enum hmm_pfn_flag_e {
126 HMM_PFN_DEVICE_PRIVATE,
131 * hmm_pfn_value_e - HMM pfn special value
134 * HMM_PFN_ERROR: corresponding CPU page table entry points to poisoned memory
135 * HMM_PFN_NONE: corresponding CPU page table entry is pte_none()
136 * HMM_PFN_SPECIAL: corresponding CPU page table entry is special; i.e., the
137 * result of vmf_insert_pfn() or vm_insert_page(). Therefore, it should not
138 * be mirrored by a device, because the entry will never have HMM_PFN_VALID
139 * set and the pfn value is undefined.
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.
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
150 enum hmm_pfn_value_e {
158 * struct hmm_range - track invalidation lock on virtual address range
160 * @hmm: the core HMM structure this range is active against
161 * @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, ...)
168 * @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
170 * @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
175 struct vm_area_struct *vma;
176 struct list_head list;
180 const uint64_t *flags;
181 const uint64_t *values;
182 uint64_t default_flags;
183 uint64_t pfn_flags_mask;
190 * 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
194 static inline unsigned hmm_range_page_shift(const struct hmm_range *range)
196 return range->page_shift;
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
204 static inline unsigned long hmm_range_page_size(const struct hmm_range *range)
206 return 1UL << hmm_range_page_shift(range);
210 * 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.
215 static inline bool hmm_range_wait_until_valid(struct hmm_range *range,
216 unsigned long timeout)
218 /* Check if mm is dead ? */
219 if (range->hmm == NULL || range->hmm->dead || range->hmm->mm == NULL) {
220 range->valid = false;
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 */
232 * hmm_range_valid() - test if a range is valid or not
234 * Returns: true if the range is valid, false otherwise.
236 static inline bool hmm_range_valid(struct hmm_range *range)
242 * hmm_pfn_to_page() - return struct page pointed to by a valid HMM pfn
243 * @range: range use to decode HMM pfn value
244 * @pfn: HMM pfn value to get corresponding struct page from
245 * Returns: struct page pointer if pfn is a valid HMM pfn, NULL otherwise
247 * 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.
250 static inline struct page *hmm_pfn_to_page(const struct hmm_range *range,
253 if (pfn == range->values[HMM_PFN_NONE])
255 if (pfn == range->values[HMM_PFN_ERROR])
257 if (pfn == range->values[HMM_PFN_SPECIAL])
259 if (!(pfn & range->flags[HMM_PFN_VALID]))
261 return pfn_to_page(pfn >> range->pfn_shift);
265 * hmm_pfn_to_pfn() - return pfn value store in a HMM pfn
266 * @range: range use to decode HMM pfn value
267 * @pfn: HMM pfn value to extract pfn from
268 * Returns: pfn value if HMM pfn is valid, -1UL otherwise
270 static inline unsigned long hmm_pfn_to_pfn(const struct hmm_range *range,
273 if (pfn == range->values[HMM_PFN_NONE])
275 if (pfn == range->values[HMM_PFN_ERROR])
277 if (pfn == range->values[HMM_PFN_SPECIAL])
279 if (!(pfn & range->flags[HMM_PFN_VALID]))
281 return (pfn >> range->pfn_shift);
285 * hmm_pfn_from_page() - create a valid HMM pfn value from struct page
286 * @range: range use to encode HMM pfn value
287 * @page: struct page pointer for which to create the HMM pfn
288 * Returns: valid HMM pfn for the page
290 static inline uint64_t hmm_pfn_from_page(const struct hmm_range *range,
293 return (page_to_pfn(page) << range->pfn_shift) |
294 range->flags[HMM_PFN_VALID];
298 * hmm_pfn_from_pfn() - create a valid HMM pfn value from pfn
299 * @range: range use to encode HMM pfn value
300 * @pfn: pfn value for which to create the HMM pfn
301 * Returns: valid HMM pfn for the pfn
303 static inline uint64_t hmm_pfn_from_pfn(const struct hmm_range *range,
306 return (pfn << range->pfn_shift) |
307 range->flags[HMM_PFN_VALID];
311 #if IS_ENABLED(CONFIG_HMM_MIRROR)
313 * Mirroring: how to synchronize device page table with CPU page table.
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:
321 * int device_bind_address_space(..., struct mm_struct *mm, ...)
323 * struct device_address_space *das;
325 * // Device driver specific initialization, and allocation of das
326 * // which contains an hmm_mirror struct as one of its fields.
329 * ret = hmm_mirror_register(&das->mirror, mm, &device_mirror_ops);
331 * // Cleanup on error
335 * // Other device driver specific initialization
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).
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.
348 * void device_unbind_address_space(struct device_address_space *das)
350 * // Device driver specific cleanup
353 * hmm_mirror_unregister(&das->mirror);
355 * // Other device driver specific cleanup, and now das can be freed
363 * enum hmm_update_event - type of update
364 * @HMM_UPDATE_INVALIDATE: invalidate range (no indication as to why)
366 enum hmm_update_event {
367 HMM_UPDATE_INVALIDATE,
371 * struct hmm_update - HMM update informations for callback
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 ?
381 enum hmm_update_event event;
386 * struct hmm_mirror_ops - HMM mirror device operations callback
388 * @update: callback to update range on a device
390 struct hmm_mirror_ops {
391 /* release() - release hmm_mirror
393 * @mirror: pointer to struct hmm_mirror
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.
399 void (*release)(struct hmm_mirror *mirror);
401 /* sync_cpu_device_pagetables() - synchronize page tables
403 * @mirror: pointer to struct hmm_mirror
404 * @update: update informations (see struct hmm_update)
405 * Returns: -EAGAIN if update.blockable false and callback need to
406 * block, 0 otherwise.
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
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
417 int (*sync_cpu_device_pagetables)(struct hmm_mirror *mirror,
418 const struct hmm_update *update);
422 * struct hmm_mirror - mirror struct for a device driver
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
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.
434 const struct hmm_mirror_ops *ops;
435 struct list_head list;
438 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm);
439 void hmm_mirror_unregister(struct hmm_mirror *mirror);
442 * hmm_mirror_mm_is_alive() - test if mm is still alive
443 * @mirror: the HMM mm mirror for which we want to lock the mmap_sem
444 * Returns: false if the mm is dead, true otherwise
446 * This is an optimization it will not accurately always return -EINVAL if the
447 * mm is dead ie there can be false negative (process is being kill but HMM is
448 * not yet inform of that). It is only intented to be use to optimize out case
449 * where driver is about to do something time consuming and it would be better
450 * to skip it if the mm is dead.
452 static inline bool hmm_mirror_mm_is_alive(struct hmm_mirror *mirror)
454 struct mm_struct *mm;
456 if (!mirror || !mirror->hmm)
458 mm = READ_ONCE(mirror->hmm->mm);
459 if (mirror->hmm->dead || !mm)
467 * Please see Documentation/vm/hmm.rst for how to use the range API.
469 int hmm_range_register(struct hmm_range *range,
470 struct mm_struct *mm,
473 unsigned page_shift);
474 void hmm_range_unregister(struct hmm_range *range);
475 long hmm_range_snapshot(struct hmm_range *range);
476 long hmm_range_fault(struct hmm_range *range, bool block);
479 * HMM_RANGE_DEFAULT_TIMEOUT - default timeout (ms) when waiting for a range
481 * When waiting for mmu notifiers we need some kind of time out otherwise we
482 * could potentialy wait for ever, 1000ms ie 1s sounds like a long time to
485 #define HMM_RANGE_DEFAULT_TIMEOUT 1000
487 /* This is a temporary helper to avoid merge conflict between trees. */
488 static inline bool hmm_vma_range_done(struct hmm_range *range)
490 bool ret = hmm_range_valid(range);
492 hmm_range_unregister(range);
496 /* This is a temporary helper to avoid merge conflict between trees. */
497 static inline int hmm_vma_fault(struct hmm_range *range, bool block)
502 * With the old API the driver must set each individual entries with
503 * the requested flags (valid, write, ...). So here we set the mask to
504 * keep intact the entries provided by the driver and zero out the
507 range->default_flags = 0;
508 range->pfn_flags_mask = -1UL;
510 ret = hmm_range_register(range, range->vma->vm_mm,
511 range->start, range->end,
516 if (!hmm_range_wait_until_valid(range, HMM_RANGE_DEFAULT_TIMEOUT)) {
518 * The mmap_sem was taken by driver we release it here and
519 * returns -EAGAIN which correspond to mmap_sem have been
520 * drop in the old API.
522 up_read(&range->vma->vm_mm->mmap_sem);
526 ret = hmm_range_fault(range, block);
528 if (ret == -EBUSY || !ret) {
529 /* Same as above drop mmap_sem to match old API. */
530 up_read(&range->vma->vm_mm->mmap_sem);
532 } else if (ret == -EAGAIN)
534 hmm_range_unregister(range);
540 /* Below are for HMM internal use only! Not to be used by device driver! */
541 void hmm_mm_destroy(struct mm_struct *mm);
543 static inline void hmm_mm_init(struct mm_struct *mm)
547 #else /* IS_ENABLED(CONFIG_HMM_MIRROR) */
548 static inline void hmm_mm_destroy(struct mm_struct *mm) {}
549 static inline void hmm_mm_init(struct mm_struct *mm) {}
550 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
552 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC)
555 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
559 * struct hmm_devmem_ops - callback for ZONE_DEVICE memory events
561 * @free: call when refcount on page reach 1 and thus is no longer use
562 * @fault: call when there is a page fault to unaddressable memory
564 * Both callback happens from page_free() and page_fault() callback of struct
565 * dev_pagemap respectively. See include/linux/memremap.h for more details on
568 * The hmm_devmem_ops callback are just here to provide a coherent and
569 * uniq API to device driver and device driver should not register their
570 * own page_free() or page_fault() but rely on the hmm_devmem_ops call-
573 struct hmm_devmem_ops {
575 * free() - free a device page
576 * @devmem: device memory structure (see struct hmm_devmem)
577 * @page: pointer to struct page being freed
579 * Call back occurs whenever a device page refcount reach 1 which
580 * means that no one is holding any reference on the page anymore
581 * (ZONE_DEVICE page have an elevated refcount of 1 as default so
582 * that they are not release to the general page allocator).
584 * Note that callback has exclusive ownership of the page (as no
585 * one is holding any reference).
587 void (*free)(struct hmm_devmem *devmem, struct page *page);
589 * fault() - CPU page fault or get user page (GUP)
590 * @devmem: device memory structure (see struct hmm_devmem)
591 * @vma: virtual memory area containing the virtual address
592 * @addr: virtual address that faulted or for which there is a GUP
593 * @page: pointer to struct page backing virtual address (unreliable)
594 * @flags: FAULT_FLAG_* (see include/linux/mm.h)
595 * @pmdp: page middle directory
596 * Returns: VM_FAULT_MINOR/MAJOR on success or one of VM_FAULT_ERROR
599 * The callback occurs whenever there is a CPU page fault or GUP on a
600 * virtual address. This means that the device driver must migrate the
601 * page back to regular memory (CPU accessible).
603 * The device driver is free to migrate more than one page from the
604 * fault() callback as an optimization. However if device decide to
605 * migrate more than one page it must always priotirize the faulting
606 * address over the others.
608 * The struct page pointer is only given as an hint to allow quick
609 * lookup of internal device driver data. A concurrent migration
610 * might have already free that page and the virtual address might
611 * not longer be back by it. So it should not be modified by the
614 * Note that mmap semaphore is held in read mode at least when this
615 * callback occurs, hence the vma is valid upon callback entry.
617 vm_fault_t (*fault)(struct hmm_devmem *devmem,
618 struct vm_area_struct *vma,
620 const struct page *page,
626 * struct hmm_devmem - track device memory
628 * @completion: completion object for device memory
629 * @pfn_first: first pfn for this resource (set by hmm_devmem_add())
630 * @pfn_last: last pfn for this resource (set by hmm_devmem_add())
631 * @resource: IO resource reserved for this chunk of memory
632 * @pagemap: device page map for that chunk
633 * @device: device to bind resource to
634 * @ops: memory operations callback
635 * @ref: per CPU refcount
636 * @page_fault: callback when CPU fault on an unaddressable device page
638 * This an helper structure for device drivers that do not wish to implement
639 * the gory details related to hotplugging new memoy and allocating struct
642 * Device drivers can directly use ZONE_DEVICE memory on their own if they
645 * The page_fault() callback must migrate page back, from device memory to
646 * system memory, so that the CPU can access it. This might fail for various
647 * reasons (device issues, device have been unplugged, ...). When such error
648 * conditions happen, the page_fault() callback must return VM_FAULT_SIGBUS and
649 * set the CPU page table entry to "poisoned".
651 * Note that because memory cgroup charges are transferred to the device memory,
652 * this should never fail due to memory restrictions. However, allocation
653 * of a regular system page might still fail because we are out of memory. If
654 * that happens, the page_fault() callback must return VM_FAULT_OOM.
656 * The page_fault() callback can also try to migrate back multiple pages in one
657 * chunk, as an optimization. It must, however, prioritize the faulting address
658 * over all the others.
660 typedef vm_fault_t (*dev_page_fault_t)(struct vm_area_struct *vma,
662 const struct page *page,
667 struct completion completion;
668 unsigned long pfn_first;
669 unsigned long pfn_last;
670 struct resource *resource;
671 struct device *device;
672 struct dev_pagemap pagemap;
673 const struct hmm_devmem_ops *ops;
674 struct percpu_ref ref;
675 dev_page_fault_t page_fault;
679 * To add (hotplug) device memory, HMM assumes that there is no real resource
680 * that reserves a range in the physical address space (this is intended to be
681 * use by unaddressable device memory). It will reserve a physical range big
682 * enough and allocate struct page for it.
684 * The device driver can wrap the hmm_devmem struct inside a private device
687 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
688 struct device *device,
690 struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
691 struct device *device,
692 struct resource *res);
695 * hmm_devmem_page_set_drvdata - set per-page driver data field
697 * @page: pointer to struct page
698 * @data: driver data value to set
700 * Because page can not be on lru we have an unsigned long that driver can use
701 * to store a per page field. This just a simple helper to do that.
703 static inline void hmm_devmem_page_set_drvdata(struct page *page,
706 page->hmm_data = data;
710 * hmm_devmem_page_get_drvdata - get per page driver data field
712 * @page: pointer to struct page
713 * Return: driver data value
715 static inline unsigned long hmm_devmem_page_get_drvdata(const struct page *page)
717 return page->hmm_data;
722 * struct hmm_device - fake device to hang device memory onto
724 * @device: device struct
725 * @minor: device minor number
728 struct device device;
733 * A device driver that wants to handle multiple devices memory through a
734 * single fake device can use hmm_device to do so. This is purely a helper and
735 * it is not strictly needed, in order to make use of any HMM functionality.
737 struct hmm_device *hmm_device_new(void *drvdata);
738 void hmm_device_put(struct hmm_device *hmm_device);
739 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
740 #else /* IS_ENABLED(CONFIG_HMM) */
741 static inline void hmm_mm_destroy(struct mm_struct *mm) {}
742 static inline void hmm_mm_init(struct mm_struct *mm) {}
743 #endif /* IS_ENABLED(CONFIG_HMM) */
745 #endif /* LINUX_HMM_H */