linux/include/linux/hmm.h
<<
>>
Prefs
   1/* SPDX-License-Identifier: GPL-2.0-or-later */
   2/*
   3 * Copyright 2013 Red Hat Inc.
   4 *
   5 * Authors: Jérôme Glisse <jglisse@redhat.com>
   6 */
   7/*
   8 * Heterogeneous Memory Management (HMM)
   9 *
  10 * See Documentation/vm/hmm.rst for reasons and overview of what HMM is and it
  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 *
  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
  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>
  63#include <asm/pgtable.h>
  64
  65#ifdef CONFIG_HMM_MIRROR
  66
  67#include <linux/device.h>
  68#include <linux/migrate.h>
  69#include <linux/memremap.h>
  70#include <linux/completion.h>
  71#include <linux/mmu_notifier.h>
  72
  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
  85 */
  86struct hmm {
  87        struct mmu_notifier     mmu_notifier;
  88        spinlock_t              ranges_lock;
  89        struct list_head        ranges;
  90        struct list_head        mirrors;
  91        struct rw_semaphore     mirrors_sem;
  92        wait_queue_head_t       wq;
  93        long                    notifiers;
  94};
  95
  96/*
  97 * hmm_pfn_flag_e - HMM flag enums
  98 *
  99 * Flags:
 100 * HMM_PFN_VALID: pfn is valid. It has, at least, read permission.
 101 * HMM_PFN_WRITE: CPU page table has write permission set
 102 * HMM_PFN_DEVICE_PRIVATE: private device memory (ZONE_DEVICE)
 103 *
 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
 107 * an array in hmm_range.flags with hmm_range.flags[HMM_PFN_VALID] == 1 << 3.
 108 * Same logic apply to all flags. This is the same idea as vm_page_prot in vma
 109 * except that this is per device driver rather than per architecture.
 110 */
 111enum 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:
 122 * HMM_PFN_ERROR: corresponding CPU page table entry points to poisoned memory
 123 * HMM_PFN_NONE: corresponding CPU page table entry is pte_none()
 124 * HMM_PFN_SPECIAL: corresponding CPU page table entry is special; i.e., the
 125 *      result of vmf_insert_pfn() or vm_insert_page(). Therefore, it should not
 126 *      be mirrored by a device, because the entry will never have HMM_PFN_VALID
 127 *      set and the pfn value is undefined.
 128 *
 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.
 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,
 135 * hmm_range.values[HMM_PFN_NONE] if there is no CPU page table entry,
 136 * hmm_range.values[HMM_PFN_SPECIAL] if CPU page table entry is a special one
 137 */
 138enum 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 *
 148 * @hmm: the core HMM structure this range is active against
 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, ...)
 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
 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 */
 161struct hmm_range {
 162        struct hmm              *hmm;
 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;
 169        uint64_t                default_flags;
 170        uint64_t                pfn_flags_mask;
 171        uint8_t                 pfn_shift;
 172        bool                    valid;
 173};
 174
 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)
 179 * Return: true if the range is valid, false otherwise.
 180 */
 181static inline bool hmm_range_wait_until_valid(struct hmm_range *range,
 182                                              unsigned long timeout)
 183{
 184        return wait_event_timeout(range->hmm->wq, range->valid,
 185                                  msecs_to_jiffies(timeout)) != 0;
 186}
 187
 188/*
 189 * hmm_range_valid() - test if a range is valid or not
 190 * @range: range
 191 * Return: true if the range is valid, false otherwise.
 192 */
 193static inline bool hmm_range_valid(struct hmm_range *range)
 194{
 195        return range->valid;
 196}
 197
 198/*
 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
 202 * Return: struct page pointer if entry is a valid, NULL otherwise
 203 *
 204 * If the device entry is valid (ie valid flag set) then return the struct page
 205 * matching the entry value. Otherwise return NULL.
 206 */
 207static inline struct page *hmm_device_entry_to_page(const struct hmm_range *range,
 208                                                    uint64_t entry)
 209{
 210        if (entry == range->values[HMM_PFN_NONE])
 211                return NULL;
 212        if (entry == range->values[HMM_PFN_ERROR])
 213                return NULL;
 214        if (entry == range->values[HMM_PFN_SPECIAL])
 215                return NULL;
 216        if (!(entry & range->flags[HMM_PFN_VALID]))
 217                return NULL;
 218        return pfn_to_page(entry >> range->pfn_shift);
 219}
 220
 221/*
 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
 225 * Return: pfn value if device entry is valid, -1UL otherwise
 226 */
 227static inline unsigned long
 228hmm_device_entry_to_pfn(const struct hmm_range *range, uint64_t pfn)
 229{
 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])
 235                return -1UL;
 236        if (!(pfn & range->flags[HMM_PFN_VALID]))
 237                return -1UL;
 238        return (pfn >> range->pfn_shift);
 239}
 240
 241/*
 242 * hmm_device_entry_from_page() - create a valid device entry for a page
 243 * @range: range use to encode HMM pfn value
 244 * @page: page for which to create the device entry
 245 * Return: valid device entry for the page
 246 */
 247static inline uint64_t hmm_device_entry_from_page(const struct hmm_range *range,
 248                                                  struct page *page)
 249{
 250        return (page_to_pfn(page) << range->pfn_shift) |
 251                range->flags[HMM_PFN_VALID];
 252}
 253
 254/*
 255 * hmm_device_entry_from_pfn() - create a valid device entry value from pfn
 256 * @range: range use to encode HMM pfn value
 257 * @pfn: pfn value for which to create the device entry
 258 * Return: valid device entry for the pfn
 259 */
 260static inline uint64_t hmm_device_entry_from_pfn(const struct hmm_range *range,
 261                                                 unsigned long pfn)
 262{
 263        return (pfn << range->pfn_shift) |
 264                range->flags[HMM_PFN_VALID];
 265}
 266
 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
 315struct hmm_mirror;
 316
 317/*
 318 * struct hmm_mirror_ops - HMM mirror device operations callback
 319 *
 320 * @update: callback to update range on a device
 321 */
 322struct hmm_mirror_ops {
 323        /* release() - release hmm_mirror
 324         *
 325         * @mirror: pointer to struct hmm_mirror
 326         *
 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.
 331         */
 332        void (*release)(struct hmm_mirror *mirror);
 333
 334        /* sync_cpu_device_pagetables() - synchronize page tables
 335         *
 336         * @mirror: pointer to struct hmm_mirror
 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.
 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         */
 350        int (*sync_cpu_device_pagetables)(
 351                struct hmm_mirror *mirror,
 352                const struct mmu_notifier_range *update);
 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 */
 366struct hmm_mirror {
 367        struct hmm                      *hmm;
 368        const struct hmm_mirror_ops     *ops;
 369        struct list_head                list;
 370};
 371
 372int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm);
 373void hmm_mirror_unregister(struct hmm_mirror *mirror);
 374
 375/*
 376 * Please see Documentation/vm/hmm.rst for how to use the range API.
 377 */
 378int hmm_range_register(struct hmm_range *range, struct hmm_mirror *mirror);
 379void hmm_range_unregister(struct hmm_range *range);
 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
 386/* Don't fault in missing PTEs, just snapshot the current state. */
 387#define HMM_FAULT_SNAPSHOT              (1 << 1)
 388
 389long hmm_range_fault(struct hmm_range *range, unsigned int flags);
 390
 391long hmm_range_dma_map(struct hmm_range *range,
 392                       struct device *device,
 393                       dma_addr_t *daddrs,
 394                       unsigned int flags);
 395long hmm_range_dma_unmap(struct hmm_range *range,
 396                         struct device *device,
 397                         dma_addr_t *daddrs,
 398                         bool dirty);
 399
 400/*
 401 * HMM_RANGE_DEFAULT_TIMEOUT - default timeout (ms) when waiting for a range
 402 *
 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.
 406 */
 407#define HMM_RANGE_DEFAULT_TIMEOUT 1000
 408
 409#endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
 410
 411#endif /* LINUX_HMM_H */
 412