drivers/md/dm-kcopyd.c

   1 /*
   2  * Copyright (C) 2002 Sistina Software (UK) Limited.
   3  * Copyright (C) 2006 Red Hat GmbH
   4  *
   5  * This file is released under the GPL.
   6  *
   7  * Kcopyd provides a simple interface for copying an area of one
   8  * block-device to one or more other block-devices, with an asynchronous
   9  * completion notification.
  10  */
  11
  12 #include <linux/types.h>
  13 #include <linux/atomic.h>
  14 #include <linux/blkdev.h>
  15 #include <linux/fs.h>
  16 #include <linux/init.h>
  17 #include <linux/list.h>
  18 #include <linux/mempool.h>
  19 #include <linux/module.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/slab.h>
  22 #include <linux/vmalloc.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/mutex.h>
  25 #include <linux/delay.h>
  26 #include <linux/device-mapper.h>
  27 #include <linux/dm-kcopyd.h>
  28
  29 #include "dm-core.h"
  30
  31 #define SUB_JOB_SIZE    128
  32 #define SPLIT_COUNT     8
  33 #define MIN_JOBS        8
  34 #define RESERVE_PAGES   (DIV_ROUND_UP(SUB_JOB_SIZE << SECTOR_SHIFT, PAGE_SIZE))
  35
  36 /*-----------------------------------------------------------------
  37  * Each kcopyd client has its own little pool of preallocated
  38  * pages for kcopyd io.
  39  *---------------------------------------------------------------*/
  40 struct dm_kcopyd_client {
  41         struct page_list *pages;
  42         unsigned nr_reserved_pages;
  43         unsigned nr_free_pages;
  44
  45         struct dm_io_client *io_client;
  46
  47         wait_queue_head_t destroyq;
  48         atomic_t nr_jobs;
  49
  50         mempool_t *job_pool;
  51
  52         struct workqueue_struct *kcopyd_wq;
  53         struct work_struct kcopyd_work;
  54
  55         struct dm_kcopyd_throttle *throttle;
  56
  57 /*
  58  * We maintain three lists of jobs:
  59  *
  60  * i)   jobs waiting for pages
  61  * ii)  jobs that have pages, and are waiting for the io to be issued.
  62  * iii) jobs that have completed.
  63  *
  64  * All three of these are protected by job_lock.
  65  */
  66         spinlock_t job_lock;
  67         struct list_head complete_jobs;
  68         struct list_head io_jobs;
  69         struct list_head pages_jobs;
  70 };
  71
  72 static struct page_list zero_page_list;
  73
  74 static DEFINE_SPINLOCK(throttle_spinlock);
  75
  76 /*
  77  * IO/IDLE accounting slowly decays after (1 << ACCOUNT_INTERVAL_SHIFT) period.
  78  * When total_period >= (1 << ACCOUNT_INTERVAL_SHIFT) the counters are divided
  79  * by 2.
  80  */
  81 #define ACCOUNT_INTERVAL_SHIFT          SHIFT_HZ
  82
  83 /*
  84  * Sleep this number of milliseconds.
  85  *
  86  * The value was decided experimentally.
  87  * Smaller values seem to cause an increased copy rate above the limit.
  88  * The reason for this is unknown but possibly due to jiffies rounding errors
  89  * or read/write cache inside the disk.
  90  */
  91 #define SLEEP_MSEC                      100
  92
  93 /*
  94  * Maximum number of sleep events. There is a theoretical livelock if more
  95  * kcopyd clients do work simultaneously which this limit avoids.
  96  */
  97 #define MAX_SLEEPS                      10
  98
  99 static void io_job_start(struct dm_kcopyd_throttle *t)
 100 {
 101         unsigned throttle, now, difference;
 102         int slept = 0, skew;
 103
 104         if (unlikely(!t))
 105                 return;
 106
 107 try_again:
 108         spin_lock_irq(&throttle_spinlock);
 109
 110         throttle = ACCESS_ONCE(t->throttle);
 111
 112         if (likely(throttle >= 100))
 113                 goto skip_limit;
 114
 115         now = jiffies;
 116         difference = now - t->last_jiffies;
 117         t->last_jiffies = now;
 118         if (t->num_io_jobs)
 119                 t->io_period += difference;
 120         t->total_period += difference;
 121
 122         /*
 123          * Maintain sane values if we got a temporary overflow.
 124          */
 125         if (unlikely(t->io_period > t->total_period))
 126                 t->io_period = t->total_period;
 127
 128         if (unlikely(t->total_period >= (1 << ACCOUNT_INTERVAL_SHIFT))) {
 129                 int shift = fls(t->total_period >> ACCOUNT_INTERVAL_SHIFT);
 130                 t->total_period >>= shift;
 131                 t->io_period >>= shift;
 132         }
 133
 134         skew = t->io_period - throttle * t->total_period / 100;
 135
 136         if (unlikely(skew > 0) && slept < MAX_SLEEPS) {
 137                 slept++;
 138                 spin_unlock_irq(&throttle_spinlock);
 139                 msleep(SLEEP_MSEC);
 140                 goto try_again;
 141         }
 142
 143 skip_limit:
 144         t->num_io_jobs++;
 145
 146         spin_unlock_irq(&throttle_spinlock);
 147 }
 148
 149 static void io_job_finish(struct dm_kcopyd_throttle *t)
 150 {
 151         unsigned long flags;
 152
 153         if (unlikely(!t))
 154                 return;
 155
 156         spin_lock_irqsave(&throttle_spinlock, flags);
 157
 158         t->num_io_jobs--;
 159
 160         if (likely(ACCESS_ONCE(t->throttle) >= 100))
 161                 goto skip_limit;
 162
 163         if (!t->num_io_jobs) {
 164                 unsigned now, difference;
 165
 166                 now = jiffies;
 167                 difference = now - t->last_jiffies;
 168                 t->last_jiffies = now;
 169
 170                 t->io_period += difference;
 171                 t->total_period += difference;
 172
 173                 /*
 174                  * Maintain sane values if we got a temporary overflow.
 175                  */
 176                 if (unlikely(t->io_period > t->total_period))
 177                         t->io_period = t->total_period;
 178         }
 179
 180 skip_limit:
 181         spin_unlock_irqrestore(&throttle_spinlock, flags);
 182 }
 183
 184
 185 static void wake(struct dm_kcopyd_client *kc)
 186 {
 187         queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
 188 }
 189
 190 /*
 191  * Obtain one page for the use of kcopyd.
 192  */
 193 static struct page_list *alloc_pl(gfp_t gfp)
 194 {
 195         struct page_list *pl;
 196
 197         pl = kmalloc(sizeof(*pl), gfp);
 198         if (!pl)
 199                 return NULL;
 200
 201         pl->page = alloc_page(gfp);
 202         if (!pl->page) {
 203                 kfree(pl);
 204                 return NULL;
 205         }
 206
 207         return pl;
 208 }
 209
 210 static void free_pl(struct page_list *pl)
 211 {
 212         __free_page(pl->page);
 213         kfree(pl);
 214 }
 215
 216 /*
 217  * Add the provided pages to a client's free page list, releasing
 218  * back to the system any beyond the reserved_pages limit.
 219  */
 220 static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
 221 {
 222         struct page_list *next;
 223
 224         do {
 225                 next = pl->next;
 226
 227                 if (kc->nr_free_pages >= kc->nr_reserved_pages)
 228                         free_pl(pl);
 229                 else {
 230                         pl->next = kc->pages;
 231                         kc->pages = pl;
 232                         kc->nr_free_pages++;
 233                 }
 234
 235                 pl = next;
 236         } while (pl);
 237 }
 238
 239 static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
 240                             unsigned int nr, struct page_list **pages)
 241 {
 242         struct page_list *pl;
 243
 244         *pages = NULL;
 245
 246         do {
 247                 pl = alloc_pl(__GFP_NOWARN | __GFP_NORETRY | __GFP_KSWAPD_RECLAIM);
 248                 if (unlikely(!pl)) {
 249                         /* Use reserved pages */
 250                         pl = kc->pages;
 251                         if (unlikely(!pl))
 252                                 goto out_of_memory;
 253                         kc->pages = pl->next;
 254                         kc->nr_free_pages--;
 255                 }
 256                 pl->next = *pages;
 257                 *pages = pl;
 258         } while (--nr);
 259
 260         return 0;
 261
 262 out_of_memory:
 263         if (*pages)
 264                 kcopyd_put_pages(kc, *pages);
 265         return -ENOMEM;
 266 }
 267
 268 /*
 269  * These three functions resize the page pool.
 270  */
 271 static void drop_pages(struct page_list *pl)
 272 {
 273         struct page_list *next;
 274
 275         while (pl) {
 276                 next = pl->next;
 277                 free_pl(pl);
 278                 pl = next;
 279         }
 280 }
 281
 282 /*
 283  * Allocate and reserve nr_pages for the use of a specific client.
 284  */
 285 static int client_reserve_pages(struct dm_kcopyd_client *kc, unsigned nr_pages)
 286 {
 287         unsigned i;
 288         struct page_list *pl = NULL, *next;
 289
 290         for (i = 0; i < nr_pages; i++) {
 291                 next = alloc_pl(GFP_KERNEL);
 292                 if (!next) {
 293                         if (pl)
 294                                 drop_pages(pl);
 295                         return -ENOMEM;
 296                 }
 297                 next->next = pl;
 298                 pl = next;
 299         }
 300
 301         kc->nr_reserved_pages += nr_pages;
 302         kcopyd_put_pages(kc, pl);
 303
 304         return 0;
 305 }
 306
 307 static void client_free_pages(struct dm_kcopyd_client *kc)
 308 {
 309         BUG_ON(kc->nr_free_pages != kc->nr_reserved_pages);
 310         drop_pages(kc->pages);
 311         kc->pages = NULL;
 312         kc->nr_free_pages = kc->nr_reserved_pages = 0;
 313 }
 314
 315 /*-----------------------------------------------------------------
 316  * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
 317  * for this reason we use a mempool to prevent the client from
 318  * ever having to do io (which could cause a deadlock).
 319  *---------------------------------------------------------------*/
 320 struct kcopyd_job {
 321         struct dm_kcopyd_client *kc;
 322         struct list_head list;
 323         unsigned long flags;
 324
 325         /*
 326          * Error state of the job.
 327          */
 328         int read_err;
 329         unsigned long write_err;
 330
 331         /*
 332          * Either READ or WRITE
 333          */
 334         int rw;
 335         struct dm_io_region source;
 336
 337         /*
 338          * The destinations for the transfer.
 339          */
 340         unsigned int num_dests;
 341         struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
 342
 343         struct page_list *pages;
 344
 345         /*
 346          * Set this to ensure you are notified when the job has
 347          * completed.  'context' is for callback to use.
 348          */
 349         dm_kcopyd_notify_fn fn;
 350         void *context;
 351
 352         /*
 353          * These fields are only used if the job has been split
 354          * into more manageable parts.
 355          */
 356         struct mutex lock;
 357         atomic_t sub_jobs;
 358         sector_t progress;
 359         sector_t write_offset;
 360
 361         struct kcopyd_job *master_job;
 362 };
 363
 364 static struct kmem_cache *_job_cache;
 365
 366 int __init dm_kcopyd_init(void)
 367 {
 368         _job_cache = kmem_cache_create("kcopyd_job",
 369                                 sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
 370                                 __alignof__(struct kcopyd_job), 0, NULL);
 371         if (!_job_cache)
 372                 return -ENOMEM;
 373
 374         zero_page_list.next = &zero_page_list;
 375         zero_page_list.page = ZERO_PAGE(0);
 376
 377         return 0;
 378 }
 379
 380 void dm_kcopyd_exit(void)
 381 {
 382         kmem_cache_destroy(_job_cache);
 383         _job_cache = NULL;
 384 }
 385
 386 /*
 387  * Functions to push and pop a job onto the head of a given job
 388  * list.
 389  */
 390 static struct kcopyd_job *pop_io_job(struct list_head *jobs,
 391                                      struct dm_kcopyd_client *kc)
 392 {
 393         struct kcopyd_job *job;
 394
 395         /*
 396          * For I/O jobs, pop any read, any write without sequential write
 397          * constraint and sequential writes that are at the right position.
 398          */
 399         list_for_each_entry(job, jobs, list) {
 400                 if (job->rw == READ || !test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags)) {
 401                         list_del(&job->list);
 402                         return job;
 403                 }
 404
 405                 if (job->write_offset == job->master_job->write_offset) {
 406                         job->master_job->write_offset += job->source.count;
 407                         list_del(&job->list);
 408                         return job;
 409                 }
 410         }
 411
 412         return NULL;
 413 }
 414
 415 static struct kcopyd_job *pop(struct list_head *jobs,
 416                               struct dm_kcopyd_client *kc)
 417 {
 418         struct kcopyd_job *job = NULL;
 419         unsigned long flags;
 420
 421         spin_lock_irqsave(&kc->job_lock, flags);
 422
 423         if (!list_empty(jobs)) {
 424                 if (jobs == &kc->io_jobs)
 425                         job = pop_io_job(jobs, kc);
 426                 else {
 427                         job = list_entry(jobs->next, struct kcopyd_job, list);
 428                         list_del(&job->list);
 429                 }
 430         }
 431         spin_unlock_irqrestore(&kc->job_lock, flags);
 432
 433         return job;
 434 }
 435
 436 static void push(struct list_head *jobs, struct kcopyd_job *job)
 437 {
 438         unsigned long flags;
 439         struct dm_kcopyd_client *kc = job->kc;
 440
 441         spin_lock_irqsave(&kc->job_lock, flags);
 442         list_add_tail(&job->list, jobs);
 443         spin_unlock_irqrestore(&kc->job_lock, flags);
 444 }
 445
 446
 447 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
 448 {
 449         unsigned long flags;
 450         struct dm_kcopyd_client *kc = job->kc;
 451
 452         spin_lock_irqsave(&kc->job_lock, flags);
 453         list_add(&job->list, jobs);
 454         spin_unlock_irqrestore(&kc->job_lock, flags);
 455 }
 456
 457 /*
 458  * These three functions process 1 item from the corresponding
 459  * job list.
 460  *
 461  * They return:
 462  * < 0: error
 463  *   0: success
 464  * > 0: can't process yet.
 465  */
 466 static int run_complete_job(struct kcopyd_job *job)
 467 {
 468         void *context = job->context;
 469         int read_err = job->read_err;
 470         unsigned long write_err = job->write_err;
 471         dm_kcopyd_notify_fn fn = job->fn;
 472         struct dm_kcopyd_client *kc = job->kc;
 473
 474         if (job->pages && job->pages != &zero_page_list)
 475                 kcopyd_put_pages(kc, job->pages);
 476         /*
 477          * If this is the master job, the sub jobs have already
 478          * completed so we can free everything.
 479          */
 480         if (job->master_job == job)
 481                 mempool_free(job, kc->job_pool);
 482         fn(read_err, write_err, context);
 483
 484         if (atomic_dec_and_test(&kc->nr_jobs))
 485                 wake_up(&kc->destroyq);
 486
 487         return 0;
 488 }
 489
 490 static void complete_io(unsigned long error, void *context)
 491 {
 492         struct kcopyd_job *job = (struct kcopyd_job *) context;
 493         struct dm_kcopyd_client *kc = job->kc;
 494
 495         io_job_finish(kc->throttle);
 496
 497         if (error) {
 498                 if (op_is_write(job->rw))
 499                         job->write_err |= error;
 500                 else
 501                         job->read_err = 1;
 502
 503                 if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
 504                         push(&kc->complete_jobs, job);
 505                         wake(kc);
 506                         return;
 507                 }
 508         }
 509
 510         if (op_is_write(job->rw))
 511                 push(&kc->complete_jobs, job);
 512
 513         else {
 514                 job->rw = WRITE;
 515                 push(&kc->io_jobs, job);
 516         }
 517
 518         wake(kc);
 519 }
 520
 521 /*
 522  * Request io on as many buffer heads as we can currently get for
 523  * a particular job.
 524  */
 525 static int run_io_job(struct kcopyd_job *job)
 526 {
 527         int r;
 528         struct dm_io_request io_req = {
 529                 .bi_op = job->rw,
 530                 .bi_op_flags = 0,
 531                 .mem.type = DM_IO_PAGE_LIST,
 532                 .mem.ptr.pl = job->pages,
 533                 .mem.offset = 0,
 534                 .notify.fn = complete_io,
 535                 .notify.context = job,
 536                 .client = job->kc->io_client,
 537         };
 538
 539         /*
 540          * If we need to write sequentially and some reads or writes failed,
 541          * no point in continuing.
 542          */
 543         if (test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags) &&
 544             job->master_job->write_err)
 545                 return -EIO;
 546
 547         io_job_start(job->kc->throttle);
 548
 549         if (job->rw == READ)
 550                 r = dm_io(&io_req, 1, &job->source, NULL);
 551         else
 552                 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
 553
 554         return r;
 555 }
 556
 557 static int run_pages_job(struct kcopyd_job *job)
 558 {
 559         int r;
 560         unsigned nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
 561
 562         r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
 563         if (!r) {
 564                 /* this job is ready for io */
 565                 push(&job->kc->io_jobs, job);
 566                 return 0;
 567         }
 568
 569         if (r == -ENOMEM)
 570                 /* can't complete now */
 571                 return 1;
 572
 573         return r;
 574 }
 575
 576 /*
 577  * Run through a list for as long as possible.  Returns the count
 578  * of successful jobs.
 579  */
 580 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
 581                         int (*fn) (struct kcopyd_job *))
 582 {
 583         struct kcopyd_job *job;
 584         int r, count = 0;
 585
 586         while ((job = pop(jobs, kc))) {
 587
 588                 r = fn(job);
 589
 590                 if (r < 0) {
 591                         /* error this rogue job */
 592                         if (op_is_write(job->rw))
 593                                 job->write_err = (unsigned long) -1L;
 594                         else
 595                                 job->read_err = 1;
 596                         push(&kc->complete_jobs, job);
 597                         break;
 598                 }
 599
 600                 if (r > 0) {
 601                         /*
 602                          * We couldn't service this job ATM, so
 603                          * push this job back onto the list.
 604                          */
 605                         push_head(jobs, job);
 606                         break;
 607                 }
 608
 609                 count++;
 610         }
 611
 612         return count;
 613 }
 614
 615 /*
 616  * kcopyd does this every time it's woken up.
 617  */
 618 static void do_work(struct work_struct *work)
 619 {
 620         struct dm_kcopyd_client *kc = container_of(work,
 621                                         struct dm_kcopyd_client, kcopyd_work);
 622         struct blk_plug plug;
 623
 624         /*
 625          * The order that these are called is *very* important.
 626          * complete jobs can free some pages for pages jobs.
 627          * Pages jobs when successful will jump onto the io jobs
 628          * list.  io jobs call wake when they complete and it all
 629          * starts again.
 630          */
 631         blk_start_plug(&plug);
 632         process_jobs(&kc->complete_jobs, kc, run_complete_job);
 633         process_jobs(&kc->pages_jobs, kc, run_pages_job);
 634         process_jobs(&kc->io_jobs, kc, run_io_job);
 635         blk_finish_plug(&plug);
 636 }
 637
 638 /*
 639  * If we are copying a small region we just dispatch a single job
 640  * to do the copy, otherwise the io has to be split up into many
 641  * jobs.
 642  */
 643 static void dispatch_job(struct kcopyd_job *job)
 644 {
 645         struct dm_kcopyd_client *kc = job->kc;
 646         atomic_inc(&kc->nr_jobs);
 647         if (unlikely(!job->source.count))
 648                 push(&kc->complete_jobs, job);
 649         else if (job->pages == &zero_page_list)
 650                 push(&kc->io_jobs, job);
 651         else
 652                 push(&kc->pages_jobs, job);
 653         wake(kc);
 654 }
 655
 656 static void segment_complete(int read_err, unsigned long write_err,
 657                              void *context)
 658 {
 659         /* FIXME: tidy this function */
 660         sector_t progress = 0;
 661         sector_t count = 0;
 662         struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
 663         struct kcopyd_job *job = sub_job->master_job;
 664         struct dm_kcopyd_client *kc = job->kc;
 665
 666         mutex_lock(&job->lock);
 667
 668         /* update the error */
 669         if (read_err)
 670                 job->read_err = 1;
 671
 672         if (write_err)
 673                 job->write_err |= write_err;
 674
 675         /*
 676          * Only dispatch more work if there hasn't been an error.
 677          */
 678         if ((!job->read_err && !job->write_err) ||
 679             test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
 680                 /* get the next chunk of work */
 681                 progress = job->progress;
 682                 count = job->source.count - progress;
 683                 if (count) {
 684                         if (count > SUB_JOB_SIZE)
 685                                 count = SUB_JOB_SIZE;
 686
 687                         job->progress += count;
 688                 }
 689         }
 690         mutex_unlock(&job->lock);
 691
 692         if (count) {
 693                 int i;
 694
 695                 *sub_job = *job;
 696                 sub_job->write_offset = progress;
 697                 sub_job->source.sector += progress;
 698                 sub_job->source.count = count;
 699
 700                 for (i = 0; i < job->num_dests; i++) {
 701                         sub_job->dests[i].sector += progress;
 702                         sub_job->dests[i].count = count;
 703                 }
 704
 705                 sub_job->fn = segment_complete;
 706                 sub_job->context = sub_job;
 707                 dispatch_job(sub_job);
 708
 709         } else if (atomic_dec_and_test(&job->sub_jobs)) {
 710
 711                 /*
 712                  * Queue the completion callback to the kcopyd thread.
 713                  *
 714                  * Some callers assume that all the completions are called
 715                  * from a single thread and don't race with each other.
 716                  *
 717                  * We must not call the callback directly here because this
 718                  * code may not be executing in the thread.
 719                  */
 720                 push(&kc->complete_jobs, job);
 721                 wake(kc);
 722         }
 723 }
 724
 725 /*
 726  * Create some sub jobs to share the work between them.
 727  */
 728 static void split_job(struct kcopyd_job *master_job)
 729 {
 730         int i;
 731
 732         atomic_inc(&master_job->kc->nr_jobs);
 733
 734         atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
 735         for (i = 0; i < SPLIT_COUNT; i++) {
 736                 master_job[i + 1].master_job = master_job;
 737                 segment_complete(0, 0u, &master_job[i + 1]);
 738         }
 739 }
 740
 741 int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
 742                    unsigned int num_dests, struct dm_io_region *dests,
 743                    unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
 744 {
 745         struct kcopyd_job *job;
 746         int i;
 747
 748         /*
 749          * Allocate an array of jobs consisting of one master job
 750          * followed by SPLIT_COUNT sub jobs.
 751          */
 752         job = mempool_alloc(kc->job_pool, GFP_NOIO);
 753
 754         /*
 755          * set up for the read.
 756          */
 757         job->kc = kc;
 758         job->flags = flags;
 759         job->read_err = 0;
 760         job->write_err = 0;
 761
 762         job->num_dests = num_dests;
 763         memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
 764
 765         /*
 766          * If one of the destination is a host-managed zoned block device,
 767          * we need to write sequentially. If one of the destination is a
 768          * host-aware device, then leave it to the caller to choose what to do.
 769          */
 770         if (!test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags)) {
 771                 for (i = 0; i < job->num_dests; i++) {
 772                         if (bdev_zoned_model(dests[i].bdev) == BLK_ZONED_HM) {
 773                                 set_bit(DM_KCOPYD_WRITE_SEQ, &job->flags);
 774                                 break;
 775                         }
 776                 }
 777         }
 778
 779         /*
 780          * If we need to write sequentially, errors cannot be ignored.
 781          */
 782         if (test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags) &&
 783             test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags))
 784                 clear_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags);
 785
 786         if (from) {
 787                 job->source = *from;
 788                 job->pages = NULL;
 789                 job->rw = READ;
 790         } else {
 791                 memset(&job->source, 0, sizeof job->source);
 792                 job->source.count = job->dests[0].count;
 793                 job->pages = &zero_page_list;
 794
 795                 /*
 796                  * Use WRITE ZEROES to optimize zeroing if all dests support it.
 797                  */
 798                 job->rw = REQ_OP_WRITE_ZEROES;
 799                 for (i = 0; i < job->num_dests; i++)
 800                         if (!bdev_write_zeroes_sectors(job->dests[i].bdev)) {
 801                                 job->rw = WRITE;
 802                                 break;
 803                         }
 804         }
 805
 806         job->fn = fn;
 807         job->context = context;
 808         job->master_job = job;
 809         job->write_offset = 0;
 810
 811         if (job->source.count <= SUB_JOB_SIZE)
 812                 dispatch_job(job);
 813         else {
 814                 mutex_init(&job->lock);
 815                 job->progress = 0;
 816                 split_job(job);
 817         }
 818
 819         return 0;
 820 }
 821 EXPORT_SYMBOL(dm_kcopyd_copy);
 822
 823 int dm_kcopyd_zero(struct dm_kcopyd_client *kc,
 824                    unsigned num_dests, struct dm_io_region *dests,
 825                    unsigned flags, dm_kcopyd_notify_fn fn, void *context)
 826 {
 827         return dm_kcopyd_copy(kc, NULL, num_dests, dests, flags, fn, context);
 828 }
 829 EXPORT_SYMBOL(dm_kcopyd_zero);
 830
 831 void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
 832                                  dm_kcopyd_notify_fn fn, void *context)
 833 {
 834         struct kcopyd_job *job;
 835
 836         job = mempool_alloc(kc->job_pool, GFP_NOIO);
 837
 838         memset(job, 0, sizeof(struct kcopyd_job));
 839         job->kc = kc;
 840         job->fn = fn;
 841         job->context = context;
 842         job->master_job = job;
 843
 844         atomic_inc(&kc->nr_jobs);
 845
 846         return job;
 847 }
 848 EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
 849
 850 void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
 851 {
 852         struct kcopyd_job *job = j;
 853         struct dm_kcopyd_client *kc = job->kc;
 854
 855         job->read_err = read_err;
 856         job->write_err = write_err;
 857
 858         push(&kc->complete_jobs, job);
 859         wake(kc);
 860 }
 861 EXPORT_SYMBOL(dm_kcopyd_do_callback);
 862
 863 /*
 864  * Cancels a kcopyd job, eg. someone might be deactivating a
 865  * mirror.
 866  */
 867 #if 0
 868 int kcopyd_cancel(struct kcopyd_job *job, int block)
 869 {
 870         /* FIXME: finish */
 871         return -1;
 872 }
 873 #endif  /*  0  */
 874
 875 /*-----------------------------------------------------------------
 876  * Client setup
 877  *---------------------------------------------------------------*/
 878 struct dm_kcopyd_client *dm_kcopyd_client_create(struct dm_kcopyd_throttle *throttle)
 879 {
 880         int r = -ENOMEM;
 881         struct dm_kcopyd_client *kc;
 882
 883         kc = kmalloc(sizeof(*kc), GFP_KERNEL);
 884         if (!kc)
 885                 return ERR_PTR(-ENOMEM);
 886
 887         spin_lock_init(&kc->job_lock);
 888         INIT_LIST_HEAD(&kc->complete_jobs);
 889         INIT_LIST_HEAD(&kc->io_jobs);
 890         INIT_LIST_HEAD(&kc->pages_jobs);
 891         kc->throttle = throttle;
 892
 893         kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
 894         if (!kc->job_pool)
 895                 goto bad_slab;
 896
 897         INIT_WORK(&kc->kcopyd_work, do_work);
 898         kc->kcopyd_wq = alloc_workqueue("kcopyd", WQ_MEM_RECLAIM, 0);
 899         if (!kc->kcopyd_wq)
 900                 goto bad_workqueue;
 901
 902         kc->pages = NULL;
 903         kc->nr_reserved_pages = kc->nr_free_pages = 0;
 904         r = client_reserve_pages(kc, RESERVE_PAGES);
 905         if (r)
 906                 goto bad_client_pages;
 907
 908         kc->io_client = dm_io_client_create();
 909         if (IS_ERR(kc->io_client)) {
 910                 r = PTR_ERR(kc->io_client);
 911                 goto bad_io_client;
 912         }
 913
 914         init_waitqueue_head(&kc->destroyq);
 915         atomic_set(&kc->nr_jobs, 0);
 916
 917         return kc;
 918
 919 bad_io_client:
 920         client_free_pages(kc);
 921 bad_client_pages:
 922         destroy_workqueue(kc->kcopyd_wq);
 923 bad_workqueue:
 924         mempool_destroy(kc->job_pool);
 925 bad_slab:
 926         kfree(kc);
 927
 928         return ERR_PTR(r);
 929 }
 930 EXPORT_SYMBOL(dm_kcopyd_client_create);
 931
 932 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
 933 {
 934         /* Wait for completion of all jobs submitted by this client. */
 935         wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
 936
 937         BUG_ON(!list_empty(&kc->complete_jobs));
 938         BUG_ON(!list_empty(&kc->io_jobs));
 939         BUG_ON(!list_empty(&kc->pages_jobs));
 940         destroy_workqueue(kc->kcopyd_wq);
 941         dm_io_client_destroy(kc->io_client);
 942         client_free_pages(kc);
 943         mempool_destroy(kc->job_pool);
 944         kfree(kc);
 945 }
 946 EXPORT_SYMBOL(dm_kcopyd_client_destroy);