drivers/md/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 <asm/types.h>
  13 #include <asm/atomic.h>
  14
  15 #include <linux/blkdev.h>
  16 #include <linux/fs.h>
  17 #include <linux/init.h>
  18 #include <linux/list.h>
  19 #include <linux/mempool.h>
  20 #include <linux/module.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/slab.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/workqueue.h>
  25 #include <linux/mutex.h>
  26
  27 #include "kcopyd.h"
  28
  29 static struct workqueue_struct *_kcopyd_wq;
  30 static struct work_struct _kcopyd_work;
  31
  32 static inline void wake(void)
  33 {
  34         queue_work(_kcopyd_wq, &_kcopyd_work);
  35 }
  36
  37 /*-----------------------------------------------------------------
  38  * Each kcopyd client has its own little pool of preallocated
  39  * pages for kcopyd io.
  40  *---------------------------------------------------------------*/
  41 struct kcopyd_client {
  42         struct list_head list;
  43
  44         spinlock_t lock;
  45         struct page_list *pages;
  46         unsigned int nr_pages;
  47         unsigned int nr_free_pages;
  48
  49         struct dm_io_client *io_client;
  50
  51         wait_queue_head_t destroyq;
  52         atomic_t nr_jobs;
  53 };
  54
  55 static struct page_list *alloc_pl(void)
  56 {
  57         struct page_list *pl;
  58
  59         pl = kmalloc(sizeof(*pl), GFP_KERNEL);
  60         if (!pl)
  61                 return NULL;
  62
  63         pl->page = alloc_page(GFP_KERNEL);
  64         if (!pl->page) {
  65                 kfree(pl);
  66                 return NULL;
  67         }
  68
  69         return pl;
  70 }
  71
  72 static void free_pl(struct page_list *pl)
  73 {
  74         __free_page(pl->page);
  75         kfree(pl);
  76 }
  77
  78 static int kcopyd_get_pages(struct kcopyd_client *kc,
  79                             unsigned int nr, struct page_list **pages)
  80 {
  81         struct page_list *pl;
  82
  83         spin_lock(&kc->lock);
  84         if (kc->nr_free_pages < nr) {
  85                 spin_unlock(&kc->lock);
  86                 return -ENOMEM;
  87         }
  88
  89         kc->nr_free_pages -= nr;
  90         for (*pages = pl = kc->pages; --nr; pl = pl->next)
  91                 ;
  92
  93         kc->pages = pl->next;
  94         pl->next = NULL;
  95
  96         spin_unlock(&kc->lock);
  97
  98         return 0;
  99 }
 100
 101 static void kcopyd_put_pages(struct kcopyd_client *kc, struct page_list *pl)
 102 {
 103         struct page_list *cursor;
 104
 105         spin_lock(&kc->lock);
 106         for (cursor = pl; cursor->next; cursor = cursor->next)
 107                 kc->nr_free_pages++;
 108
 109         kc->nr_free_pages++;
 110         cursor->next = kc->pages;
 111         kc->pages = pl;
 112         spin_unlock(&kc->lock);
 113 }
 114
 115 /*
 116  * These three functions resize the page pool.
 117  */
 118 static void drop_pages(struct page_list *pl)
 119 {
 120         struct page_list *next;
 121
 122         while (pl) {
 123                 next = pl->next;
 124                 free_pl(pl);
 125                 pl = next;
 126         }
 127 }
 128
 129 static int client_alloc_pages(struct kcopyd_client *kc, unsigned int nr)
 130 {
 131         unsigned int i;
 132         struct page_list *pl = NULL, *next;
 133
 134         for (i = 0; i < nr; i++) {
 135                 next = alloc_pl();
 136                 if (!next) {
 137                         if (pl)
 138                                 drop_pages(pl);
 139                         return -ENOMEM;
 140                 }
 141                 next->next = pl;
 142                 pl = next;
 143         }
 144
 145         kcopyd_put_pages(kc, pl);
 146         kc->nr_pages += nr;
 147         return 0;
 148 }
 149
 150 static void client_free_pages(struct kcopyd_client *kc)
 151 {
 152         BUG_ON(kc->nr_free_pages != kc->nr_pages);
 153         drop_pages(kc->pages);
 154         kc->pages = NULL;
 155         kc->nr_free_pages = kc->nr_pages = 0;
 156 }
 157
 158 /*-----------------------------------------------------------------
 159  * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
 160  * for this reason we use a mempool to prevent the client from
 161  * ever having to do io (which could cause a deadlock).
 162  *---------------------------------------------------------------*/
 163 struct kcopyd_job {
 164         struct kcopyd_client *kc;
 165         struct list_head list;
 166         unsigned long flags;
 167
 168         /*
 169          * Error state of the job.
 170          */
 171         int read_err;
 172         unsigned int write_err;
 173
 174         /*
 175          * Either READ or WRITE
 176          */
 177         int rw;
 178         struct io_region source;
 179
 180         /*
 181          * The destinations for the transfer.
 182          */
 183         unsigned int num_dests;
 184         struct io_region dests[KCOPYD_MAX_REGIONS];
 185
 186         sector_t offset;
 187         unsigned int nr_pages;
 188         struct page_list *pages;
 189
 190         /*
 191          * Set this to ensure you are notified when the job has
 192          * completed.  'context' is for callback to use.
 193          */
 194         kcopyd_notify_fn fn;
 195         void *context;
 196
 197         /*
 198          * These fields are only used if the job has been split
 199          * into more manageable parts.
 200          */
 201         struct semaphore lock;
 202         atomic_t sub_jobs;
 203         sector_t progress;
 204 };
 205
 206 /* FIXME: this should scale with the number of pages */
 207 #define MIN_JOBS 512
 208
 209 static struct kmem_cache *_job_cache;
 210 static mempool_t *_job_pool;
 211
 212 /*
 213  * We maintain three lists of jobs:
 214  *
 215  * i)   jobs waiting for pages
 216  * ii)  jobs that have pages, and are waiting for the io to be issued.
 217  * iii) jobs that have completed.
 218  *
 219  * All three of these are protected by job_lock.
 220  */
 221 static DEFINE_SPINLOCK(_job_lock);
 222
 223 static LIST_HEAD(_complete_jobs);
 224 static LIST_HEAD(_io_jobs);
 225 static LIST_HEAD(_pages_jobs);
 226
 227 static int jobs_init(void)
 228 {
 229         _job_cache = kmem_cache_create("kcopyd-jobs",
 230                                        sizeof(struct kcopyd_job),
 231                                        __alignof__(struct kcopyd_job),
 232                                        0, NULL, NULL);
 233         if (!_job_cache)
 234                 return -ENOMEM;
 235
 236         _job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
 237         if (!_job_pool) {
 238                 kmem_cache_destroy(_job_cache);
 239                 return -ENOMEM;
 240         }
 241
 242         return 0;
 243 }
 244
 245 static void jobs_exit(void)
 246 {
 247         BUG_ON(!list_empty(&_complete_jobs));
 248         BUG_ON(!list_empty(&_io_jobs));
 249         BUG_ON(!list_empty(&_pages_jobs));
 250
 251         mempool_destroy(_job_pool);
 252         kmem_cache_destroy(_job_cache);
 253         _job_pool = NULL;
 254         _job_cache = NULL;
 255 }
 256
 257 /*
 258  * Functions to push and pop a job onto the head of a given job
 259  * list.
 260  */
 261 static inline struct kcopyd_job *pop(struct list_head *jobs)
 262 {
 263         struct kcopyd_job *job = NULL;
 264         unsigned long flags;
 265
 266         spin_lock_irqsave(&_job_lock, flags);
 267
 268         if (!list_empty(jobs)) {
 269                 job = list_entry(jobs->next, struct kcopyd_job, list);
 270                 list_del(&job->list);
 271         }
 272         spin_unlock_irqrestore(&_job_lock, flags);
 273
 274         return job;
 275 }
 276
 277 static inline void push(struct list_head *jobs, struct kcopyd_job *job)
 278 {
 279         unsigned long flags;
 280
 281         spin_lock_irqsave(&_job_lock, flags);
 282         list_add_tail(&job->list, jobs);
 283         spin_unlock_irqrestore(&_job_lock, flags);
 284 }
 285
 286 /*
 287  * These three functions process 1 item from the corresponding
 288  * job list.
 289  *
 290  * They return:
 291  * < 0: error
 292  *   0: success
 293  * > 0: can't process yet.
 294  */
 295 static int run_complete_job(struct kcopyd_job *job)
 296 {
 297         void *context = job->context;
 298         int read_err = job->read_err;
 299         unsigned int write_err = job->write_err;
 300         kcopyd_notify_fn fn = job->fn;
 301         struct kcopyd_client *kc = job->kc;
 302
 303         kcopyd_put_pages(kc, job->pages);
 304         mempool_free(job, _job_pool);
 305         fn(read_err, write_err, context);
 306
 307         if (atomic_dec_and_test(&kc->nr_jobs))
 308                 wake_up(&kc->destroyq);
 309
 310         return 0;
 311 }
 312
 313 static void complete_io(unsigned long error, void *context)
 314 {
 315         struct kcopyd_job *job = (struct kcopyd_job *) context;
 316
 317         if (error) {
 318                 if (job->rw == WRITE)
 319                         job->write_err |= error;
 320                 else
 321                         job->read_err = 1;
 322
 323                 if (!test_bit(KCOPYD_IGNORE_ERROR, &job->flags)) {
 324                         push(&_complete_jobs, job);
 325                         wake();
 326                         return;
 327                 }
 328         }
 329
 330         if (job->rw == WRITE)
 331                 push(&_complete_jobs, job);
 332
 333         else {
 334                 job->rw = WRITE;
 335                 push(&_io_jobs, job);
 336         }
 337
 338         wake();
 339 }
 340
 341 /*
 342  * Request io on as many buffer heads as we can currently get for
 343  * a particular job.
 344  */
 345 static int run_io_job(struct kcopyd_job *job)
 346 {
 347         int r;
 348         struct dm_io_request io_req = {
 349                 .bi_rw = job->rw,
 350                 .mem.type = DM_IO_PAGE_LIST,
 351                 .mem.ptr.pl = job->pages,
 352                 .mem.offset = job->offset,
 353                 .notify.fn = complete_io,
 354                 .notify.context = job,
 355                 .client = job->kc->io_client,
 356         };
 357
 358         if (job->rw == READ)
 359                 r = dm_io(&io_req, 1, &job->source, NULL);
 360         else
 361                 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
 362
 363         return r;
 364 }
 365
 366 static int run_pages_job(struct kcopyd_job *job)
 367 {
 368         int r;
 369
 370         job->nr_pages = dm_div_up(job->dests[0].count + job->offset,
 371                                   PAGE_SIZE >> 9);
 372         r = kcopyd_get_pages(job->kc, job->nr_pages, &job->pages);
 373         if (!r) {
 374                 /* this job is ready for io */
 375                 push(&_io_jobs, job);
 376                 return 0;
 377         }
 378
 379         if (r == -ENOMEM)
 380                 /* can't complete now */
 381                 return 1;
 382
 383         return r;
 384 }
 385
 386 /*
 387  * Run through a list for as long as possible.  Returns the count
 388  * of successful jobs.
 389  */
 390 static int process_jobs(struct list_head *jobs, int (*fn) (struct kcopyd_job *))
 391 {
 392         struct kcopyd_job *job;
 393         int r, count = 0;
 394
 395         while ((job = pop(jobs))) {
 396
 397                 r = fn(job);
 398
 399                 if (r < 0) {
 400                         /* error this rogue job */
 401                         if (job->rw == WRITE)
 402                                 job->write_err = (unsigned int) -1;
 403                         else
 404                                 job->read_err = 1;
 405                         push(&_complete_jobs, job);
 406                         break;
 407                 }
 408
 409                 if (r > 0) {
 410                         /*
 411                          * We couldn't service this job ATM, so
 412                          * push this job back onto the list.
 413                          */
 414                         push(jobs, job);
 415                         break;
 416                 }
 417
 418                 count++;
 419         }
 420
 421         return count;
 422 }
 423
 424 /*
 425  * kcopyd does this every time it's woken up.
 426  */
 427 static void do_work(struct work_struct *ignored)
 428 {
 429         /*
 430          * The order that these are called is *very* important.
 431          * complete jobs can free some pages for pages jobs.
 432          * Pages jobs when successful will jump onto the io jobs
 433          * list.  io jobs call wake when they complete and it all
 434          * starts again.
 435          */
 436         process_jobs(&_complete_jobs, run_complete_job);
 437         process_jobs(&_pages_jobs, run_pages_job);
 438         process_jobs(&_io_jobs, run_io_job);
 439 }
 440
 441 /*
 442  * If we are copying a small region we just dispatch a single job
 443  * to do the copy, otherwise the io has to be split up into many
 444  * jobs.
 445  */
 446 static void dispatch_job(struct kcopyd_job *job)
 447 {
 448         atomic_inc(&job->kc->nr_jobs);
 449         push(&_pages_jobs, job);
 450         wake();
 451 }
 452
 453 #define SUB_JOB_SIZE 128
 454 static void segment_complete(int read_err,
 455                              unsigned int write_err, void *context)
 456 {
 457         /* FIXME: tidy this function */
 458         sector_t progress = 0;
 459         sector_t count = 0;
 460         struct kcopyd_job *job = (struct kcopyd_job *) context;
 461
 462         down(&job->lock);
 463
 464         /* update the error */
 465         if (read_err)
 466                 job->read_err = 1;
 467
 468         if (write_err)
 469                 job->write_err |= write_err;
 470
 471         /*
 472          * Only dispatch more work if there hasn't been an error.
 473          */
 474         if ((!job->read_err && !job->write_err) ||
 475             test_bit(KCOPYD_IGNORE_ERROR, &job->flags)) {
 476                 /* get the next chunk of work */
 477                 progress = job->progress;
 478                 count = job->source.count - progress;
 479                 if (count) {
 480                         if (count > SUB_JOB_SIZE)
 481                                 count = SUB_JOB_SIZE;
 482
 483                         job->progress += count;
 484                 }
 485         }
 486         up(&job->lock);
 487
 488         if (count) {
 489                 int i;
 490                 struct kcopyd_job *sub_job = mempool_alloc(_job_pool, GFP_NOIO);
 491
 492                 *sub_job = *job;
 493                 sub_job->source.sector += progress;
 494                 sub_job->source.count = count;
 495
 496                 for (i = 0; i < job->num_dests; i++) {
 497                         sub_job->dests[i].sector += progress;
 498                         sub_job->dests[i].count = count;
 499                 }
 500
 501                 sub_job->fn = segment_complete;
 502                 sub_job->context = job;
 503                 dispatch_job(sub_job);
 504
 505         } else if (atomic_dec_and_test(&job->sub_jobs)) {
 506
 507                 /*
 508                  * To avoid a race we must keep the job around
 509                  * until after the notify function has completed.
 510                  * Otherwise the client may try and stop the job
 511                  * after we've completed.
 512                  */
 513                 job->fn(read_err, write_err, job->context);
 514                 mempool_free(job, _job_pool);
 515         }
 516 }
 517
 518 /*
 519  * Create some little jobs that will do the move between
 520  * them.
 521  */
 522 #define SPLIT_COUNT 8
 523 static void split_job(struct kcopyd_job *job)
 524 {
 525         int i;
 526
 527         atomic_set(&job->sub_jobs, SPLIT_COUNT);
 528         for (i = 0; i < SPLIT_COUNT; i++)
 529                 segment_complete(0, 0u, job);
 530 }
 531
 532 int kcopyd_copy(struct kcopyd_client *kc, struct io_region *from,
 533                 unsigned int num_dests, struct io_region *dests,
 534                 unsigned int flags, kcopyd_notify_fn fn, void *context)
 535 {
 536         struct kcopyd_job *job;
 537
 538         /*
 539          * Allocate a new job.
 540          */
 541         job = mempool_alloc(_job_pool, GFP_NOIO);
 542
 543         /*
 544          * set up for the read.
 545          */
 546         job->kc = kc;
 547         job->flags = flags;
 548         job->read_err = 0;
 549         job->write_err = 0;
 550         job->rw = READ;
 551
 552         job->source = *from;
 553
 554         job->num_dests = num_dests;
 555         memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
 556
 557         job->offset = 0;
 558         job->nr_pages = 0;
 559         job->pages = NULL;
 560
 561         job->fn = fn;
 562         job->context = context;
 563
 564         if (job->source.count < SUB_JOB_SIZE)
 565                 dispatch_job(job);
 566
 567         else {
 568                 init_MUTEX(&job->lock);
 569                 job->progress = 0;
 570                 split_job(job);
 571         }
 572
 573         return 0;
 574 }
 575
 576 /*
 577  * Cancels a kcopyd job, eg. someone might be deactivating a
 578  * mirror.
 579  */
 580 #if 0
 581 int kcopyd_cancel(struct kcopyd_job *job, int block)
 582 {
 583         /* FIXME: finish */
 584         return -1;
 585 }
 586 #endif  /*  0  */
 587
 588 /*-----------------------------------------------------------------
 589  * Unit setup
 590  *---------------------------------------------------------------*/
 591 static DEFINE_MUTEX(_client_lock);
 592 static LIST_HEAD(_clients);
 593
 594 static void client_add(struct kcopyd_client *kc)
 595 {
 596         mutex_lock(&_client_lock);
 597         list_add(&kc->list, &_clients);
 598         mutex_unlock(&_client_lock);
 599 }
 600
 601 static void client_del(struct kcopyd_client *kc)
 602 {
 603         mutex_lock(&_client_lock);
 604         list_del(&kc->list);
 605         mutex_unlock(&_client_lock);
 606 }
 607
 608 static DEFINE_MUTEX(kcopyd_init_lock);
 609 static int kcopyd_clients = 0;
 610
 611 static int kcopyd_init(void)
 612 {
 613         int r;
 614
 615         mutex_lock(&kcopyd_init_lock);
 616
 617         if (kcopyd_clients) {
 618                 /* Already initialized. */
 619                 kcopyd_clients++;
 620                 mutex_unlock(&kcopyd_init_lock);
 621                 return 0;
 622         }
 623
 624         r = jobs_init();
 625         if (r) {
 626                 mutex_unlock(&kcopyd_init_lock);
 627                 return r;
 628         }
 629
 630         _kcopyd_wq = create_singlethread_workqueue("kcopyd");
 631         if (!_kcopyd_wq) {
 632                 jobs_exit();
 633                 mutex_unlock(&kcopyd_init_lock);
 634                 return -ENOMEM;
 635         }
 636
 637         kcopyd_clients++;
 638         INIT_WORK(&_kcopyd_work, do_work);
 639         mutex_unlock(&kcopyd_init_lock);
 640         return 0;
 641 }
 642
 643 static void kcopyd_exit(void)
 644 {
 645         mutex_lock(&kcopyd_init_lock);
 646         kcopyd_clients--;
 647         if (!kcopyd_clients) {
 648                 jobs_exit();
 649                 destroy_workqueue(_kcopyd_wq);
 650                 _kcopyd_wq = NULL;
 651         }
 652         mutex_unlock(&kcopyd_init_lock);
 653 }
 654
 655 int kcopyd_client_create(unsigned int nr_pages, struct kcopyd_client **result)
 656 {
 657         int r = 0;
 658         struct kcopyd_client *kc;
 659
 660         r = kcopyd_init();
 661         if (r)
 662                 return r;
 663
 664         kc = kmalloc(sizeof(*kc), GFP_KERNEL);
 665         if (!kc) {
 666                 kcopyd_exit();
 667                 return -ENOMEM;
 668         }
 669
 670         spin_lock_init(&kc->lock);
 671         kc->pages = NULL;
 672         kc->nr_pages = kc->nr_free_pages = 0;
 673         r = client_alloc_pages(kc, nr_pages);
 674         if (r) {
 675                 kfree(kc);
 676                 kcopyd_exit();
 677                 return r;
 678         }
 679
 680         kc->io_client = dm_io_client_create(nr_pages);
 681         if (IS_ERR(kc->io_client)) {
 682                 r = PTR_ERR(kc->io_client);
 683                 client_free_pages(kc);
 684                 kfree(kc);
 685                 kcopyd_exit();
 686                 return r;
 687         }
 688
 689         init_waitqueue_head(&kc->destroyq);
 690         atomic_set(&kc->nr_jobs, 0);
 691
 692         client_add(kc);
 693         *result = kc;
 694         return 0;
 695 }
 696
 697 void kcopyd_client_destroy(struct kcopyd_client *kc)
 698 {
 699         /* Wait for completion of all jobs submitted by this client. */
 700         wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
 701
 702         dm_io_client_destroy(kc->io_client);
 703         client_free_pages(kc);
 704         client_del(kc);
 705         kfree(kc);
 706         kcopyd_exit();
 707 }
 708
 709 EXPORT_SYMBOL(kcopyd_client_create);
 710 EXPORT_SYMBOL(kcopyd_client_destroy);
 711 EXPORT_SYMBOL(kcopyd_copy);