drivers/staging/lustre/lustre/llite/vvp_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for VVP layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_LLITE
  39
  40 #include <linux/atomic.h>
  41 #include <linux/bitops.h>
  42 #include <linux/mm.h>
  43 #include <linux/mutex.h>
  44 #include <linux/page-flags.h>
  45 #include <linux/pagemap.h>
  46
  47 #include "llite_internal.h"
  48 #include "vvp_internal.h"
  49
  50 /*****************************************************************************
  51  *
  52  * Page operations.
  53  *
  54  */
  55
  56 static void vvp_page_fini_common(struct vvp_page *vpg)
  57 {
  58         struct page *vmpage = vpg->vpg_page;
  59
  60         LASSERT(vmpage);
  61         put_page(vmpage);
  62 }
  63
  64 static void vvp_page_fini(const struct lu_env *env,
  65                           struct cl_page_slice *slice)
  66 {
  67         struct vvp_page *vpg     = cl2vvp_page(slice);
  68         struct page     *vmpage  = vpg->vpg_page;
  69
  70         /*
  71          * vmpage->private was already cleared when page was moved into
  72          * VPG_FREEING state.
  73          */
  74         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  75         vvp_page_fini_common(vpg);
  76 }
  77
  78 static int vvp_page_own(const struct lu_env *env,
  79                         const struct cl_page_slice *slice, struct cl_io *io,
  80                         int nonblock)
  81 {
  82         struct vvp_page *vpg    = cl2vvp_page(slice);
  83         struct page     *vmpage = vpg->vpg_page;
  84
  85         LASSERT(vmpage);
  86         if (nonblock) {
  87                 if (!trylock_page(vmpage))
  88                         return -EAGAIN;
  89
  90                 if (unlikely(PageWriteback(vmpage))) {
  91                         unlock_page(vmpage);
  92                         return -EAGAIN;
  93                 }
  94
  95                 return 0;
  96         }
  97
  98         lock_page(vmpage);
  99         wait_on_page_writeback(vmpage);
 100
 101         return 0;
 102 }
 103
 104 static void vvp_page_assume(const struct lu_env *env,
 105                             const struct cl_page_slice *slice,
 106                             struct cl_io *unused)
 107 {
 108         struct page *vmpage = cl2vm_page(slice);
 109
 110         LASSERT(vmpage);
 111         LASSERT(PageLocked(vmpage));
 112         wait_on_page_writeback(vmpage);
 113 }
 114
 115 static void vvp_page_unassume(const struct lu_env *env,
 116                               const struct cl_page_slice *slice,
 117                               struct cl_io *unused)
 118 {
 119         struct page *vmpage = cl2vm_page(slice);
 120
 121         LASSERT(vmpage);
 122         LASSERT(PageLocked(vmpage));
 123 }
 124
 125 static void vvp_page_disown(const struct lu_env *env,
 126                             const struct cl_page_slice *slice, struct cl_io *io)
 127 {
 128         struct page *vmpage = cl2vm_page(slice);
 129
 130         LASSERT(vmpage);
 131         LASSERT(PageLocked(vmpage));
 132
 133         unlock_page(cl2vm_page(slice));
 134 }
 135
 136 static void vvp_page_discard(const struct lu_env *env,
 137                              const struct cl_page_slice *slice,
 138                              struct cl_io *unused)
 139 {
 140         struct page        *vmpage  = cl2vm_page(slice);
 141         struct vvp_page      *vpg     = cl2vvp_page(slice);
 142
 143         LASSERT(vmpage);
 144         LASSERT(PageLocked(vmpage));
 145
 146         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used)
 147                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 148
 149         ll_invalidate_page(vmpage);
 150 }
 151
 152 static void vvp_page_delete(const struct lu_env *env,
 153                             const struct cl_page_slice *slice)
 154 {
 155         struct page       *vmpage = cl2vm_page(slice);
 156         struct inode     *inode  = vmpage->mapping->host;
 157         struct cl_object *obj    = slice->cpl_obj;
 158         struct cl_page   *page   = slice->cpl_page;
 159         int refc;
 160
 161         LASSERT(PageLocked(vmpage));
 162         LASSERT((struct cl_page *)vmpage->private == page);
 163         LASSERT(inode == vvp_object_inode(obj));
 164
 165         vvp_write_complete(cl2vvp(obj), cl2vvp_page(slice));
 166
 167         /* Drop the reference count held in vvp_page_init */
 168         refc = atomic_dec_return(&page->cp_ref);
 169         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 170
 171         ClearPageUptodate(vmpage);
 172         ClearPagePrivate(vmpage);
 173         vmpage->private = 0;
 174         /*
 175          * Reference from vmpage to cl_page is removed, but the reference back
 176          * is still here. It is removed later in vvp_page_fini().
 177          */
 178 }
 179
 180 static void vvp_page_export(const struct lu_env *env,
 181                             const struct cl_page_slice *slice,
 182                             int uptodate)
 183 {
 184         struct page *vmpage = cl2vm_page(slice);
 185
 186         LASSERT(vmpage);
 187         LASSERT(PageLocked(vmpage));
 188         if (uptodate)
 189                 SetPageUptodate(vmpage);
 190         else
 191                 ClearPageUptodate(vmpage);
 192 }
 193
 194 static int vvp_page_is_vmlocked(const struct lu_env *env,
 195                                 const struct cl_page_slice *slice)
 196 {
 197         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 198 }
 199
 200 static int vvp_page_prep_read(const struct lu_env *env,
 201                               const struct cl_page_slice *slice,
 202                               struct cl_io *unused)
 203 {
 204         /* Skip the page already marked as PG_uptodate. */
 205         return PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0;
 206 }
 207
 208 static int vvp_page_prep_write(const struct lu_env *env,
 209                                const struct cl_page_slice *slice,
 210                                struct cl_io *unused)
 211 {
 212         struct page *vmpage = cl2vm_page(slice);
 213         struct cl_page *pg = slice->cpl_page;
 214
 215         LASSERT(PageLocked(vmpage));
 216         LASSERT(!PageDirty(vmpage));
 217
 218         /* ll_writepage path is not a sync write, so need to set page writeback
 219          * flag
 220          */
 221         if (!pg->cp_sync_io)
 222                 set_page_writeback(vmpage);
 223
 224         vvp_write_pending(cl2vvp(slice->cpl_obj), cl2vvp_page(slice));
 225
 226         return 0;
 227 }
 228
 229 /**
 230  * Handles page transfer errors at VM level.
 231  *
 232  * This takes inode as a separate argument, because inode on which error is to
 233  * be set can be different from \a vmpage inode in case of direct-io.
 234  */
 235 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 236 {
 237         struct vvp_object *obj = cl_inode2vvp(inode);
 238
 239         if (ioret == 0) {
 240                 ClearPageError(vmpage);
 241                 obj->vob_discard_page_warned = 0;
 242         } else {
 243                 SetPageError(vmpage);
 244                 mapping_set_error(inode->i_mapping, ioret);
 245
 246                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 247                     obj->vob_discard_page_warned == 0) {
 248                         obj->vob_discard_page_warned = 1;
 249                         ll_dirty_page_discard_warn(vmpage, ioret);
 250                 }
 251         }
 252 }
 253
 254 static void vvp_page_completion_read(const struct lu_env *env,
 255                                      const struct cl_page_slice *slice,
 256                                      int ioret)
 257 {
 258         struct vvp_page *vpg    = cl2vvp_page(slice);
 259         struct page     *vmpage = vpg->vpg_page;
 260         struct cl_page  *page   = slice->cpl_page;
 261         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 262
 263         LASSERT(PageLocked(vmpage));
 264         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 265
 266         if (vpg->vpg_defer_uptodate)
 267                 ll_ra_count_put(ll_i2sbi(inode), 1);
 268
 269         if (ioret == 0)  {
 270                 if (!vpg->vpg_defer_uptodate)
 271                         cl_page_export(env, page, 1);
 272         } else {
 273                 vpg->vpg_defer_uptodate = 0;
 274         }
 275
 276         if (!page->cp_sync_io)
 277                 unlock_page(vmpage);
 278 }
 279
 280 static void vvp_page_completion_write(const struct lu_env *env,
 281                                       const struct cl_page_slice *slice,
 282                                       int ioret)
 283 {
 284         struct vvp_page *vpg     = cl2vvp_page(slice);
 285         struct cl_page  *pg     = slice->cpl_page;
 286         struct page     *vmpage = vpg->vpg_page;
 287
 288         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 289
 290         /*
 291          * TODO: Actually it makes sense to add the page into oap pending
 292          * list again and so that we don't need to take the page out from
 293          * SoM write pending list, if we just meet a recoverable error,
 294          * -ENOMEM, etc.
 295          * To implement this, we just need to return a non zero value in
 296          * ->cpo_completion method. The underlying transfer should be notified
 297          * and then re-add the page into pending transfer queue.  -jay
 298          */
 299
 300         vpg->vpg_write_queued = 0;
 301         vvp_write_complete(cl2vvp(slice->cpl_obj), vpg);
 302
 303         if (pg->cp_sync_io) {
 304                 LASSERT(PageLocked(vmpage));
 305                 LASSERT(!PageWriteback(vmpage));
 306         } else {
 307                 LASSERT(PageWriteback(vmpage));
 308                 /*
 309                  * Only mark the page error only when it's an async write
 310                  * because applications won't wait for IO to finish.
 311                  */
 312                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 313
 314                 end_page_writeback(vmpage);
 315         }
 316 }
 317
 318 /**
 319  * Implements cl_page_operations::cpo_make_ready() method.
 320  *
 321  * This is called to yank a page from the transfer cache and to send it out as
 322  * a part of transfer. This function try-locks the page. If try-lock failed,
 323  * page is owned by some concurrent IO, and should be skipped (this is bad,
 324  * but hopefully rare situation, as it usually results in transfer being
 325  * shorter than possible).
 326  *
 327  * \retval 0      success, page can be placed into transfer
 328  *
 329  * \retval -EAGAIN page is either used by concurrent IO has been
 330  * truncated. Skip it.
 331  */
 332 static int vvp_page_make_ready(const struct lu_env *env,
 333                                const struct cl_page_slice *slice)
 334 {
 335         struct page *vmpage = cl2vm_page(slice);
 336         struct cl_page *pg = slice->cpl_page;
 337         int result = 0;
 338
 339         lock_page(vmpage);
 340         if (clear_page_dirty_for_io(vmpage)) {
 341                 LASSERT(pg->cp_state == CPS_CACHED);
 342                 /* This actually clears the dirty bit in the radix tree. */
 343                 set_page_writeback(vmpage);
 344                 vvp_write_pending(cl2vvp(slice->cpl_obj), cl2vvp_page(slice));
 345                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 346         } else if (pg->cp_state == CPS_PAGEOUT) {
 347                 /* is it possible for osc_flush_async_page() to already
 348                  * make it ready?
 349                  */
 350                 result = -EALREADY;
 351         } else {
 352                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 353                               pg->cp_state);
 354                 LBUG();
 355         }
 356         unlock_page(vmpage);
 357         return result;
 358 }
 359
 360 static int vvp_page_is_under_lock(const struct lu_env *env,
 361                                   const struct cl_page_slice *slice,
 362                                   struct cl_io *io, pgoff_t *max_index)
 363 {
 364         if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
 365             io->ci_type == CIT_FAULT) {
 366                 struct vvp_io *vio = vvp_env_io(env);
 367
 368                 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED))
 369                         *max_index = CL_PAGE_EOF;
 370         }
 371         return 0;
 372 }
 373
 374 static int vvp_page_print(const struct lu_env *env,
 375                           const struct cl_page_slice *slice,
 376                           void *cookie, lu_printer_t printer)
 377 {
 378         struct vvp_page *vpg = cl2vvp_page(slice);
 379         struct page     *vmpage = vpg->vpg_page;
 380
 381         (*printer)(env, cookie, LUSTRE_VVP_NAME "-page@%p(%d:%d:%d) vm@%p ",
 382                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used,
 383                    vpg->vpg_write_queued, vmpage);
 384         if (vmpage) {
 385                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 386                            (long)vmpage->flags, page_count(vmpage),
 387                            page_mapcount(vmpage), vmpage->private,
 388                            vmpage->index,
 389                            list_empty(&vmpage->lru) ? "not-" : "");
 390         }
 391
 392         (*printer)(env, cookie, "\n");
 393
 394         return 0;
 395 }
 396
 397 static int vvp_page_fail(const struct lu_env *env,
 398                          const struct cl_page_slice *slice)
 399 {
 400         /*
 401          * Cached read?
 402          */
 403         LBUG();
 404
 405         return 0;
 406 }
 407
 408 static const struct cl_page_operations vvp_page_ops = {
 409         .cpo_own           = vvp_page_own,
 410         .cpo_assume     = vvp_page_assume,
 411         .cpo_unassume      = vvp_page_unassume,
 412         .cpo_disown     = vvp_page_disown,
 413         .cpo_discard       = vvp_page_discard,
 414         .cpo_delete     = vvp_page_delete,
 415         .cpo_export     = vvp_page_export,
 416         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 417         .cpo_fini         = vvp_page_fini,
 418         .cpo_print       = vvp_page_print,
 419         .cpo_is_under_lock = vvp_page_is_under_lock,
 420         .io = {
 421                 [CRT_READ] = {
 422                         .cpo_prep       = vvp_page_prep_read,
 423                         .cpo_completion  = vvp_page_completion_read,
 424                         .cpo_make_ready = vvp_page_fail,
 425                 },
 426                 [CRT_WRITE] = {
 427                         .cpo_prep       = vvp_page_prep_write,
 428                         .cpo_completion  = vvp_page_completion_write,
 429                         .cpo_make_ready  = vvp_page_make_ready,
 430                 },
 431         },
 432 };
 433
 434 static int vvp_transient_page_prep(const struct lu_env *env,
 435                                    const struct cl_page_slice *slice,
 436                                    struct cl_io *unused)
 437 {
 438         /* transient page should always be sent. */
 439         return 0;
 440 }
 441
 442 static int vvp_transient_page_own(const struct lu_env *env,
 443                                   const struct cl_page_slice *slice,
 444                                   struct cl_io *unused, int nonblock)
 445 {
 446         return 0;
 447 }
 448
 449 static void vvp_transient_page_assume(const struct lu_env *env,
 450                                       const struct cl_page_slice *slice,
 451                                       struct cl_io *unused)
 452 {
 453 }
 454
 455 static void vvp_transient_page_unassume(const struct lu_env *env,
 456                                         const struct cl_page_slice *slice,
 457                                         struct cl_io *unused)
 458 {
 459 }
 460
 461 static void vvp_transient_page_disown(const struct lu_env *env,
 462                                       const struct cl_page_slice *slice,
 463                                       struct cl_io *unused)
 464 {
 465 }
 466
 467 static void vvp_transient_page_discard(const struct lu_env *env,
 468                                        const struct cl_page_slice *slice,
 469                                        struct cl_io *unused)
 470 {
 471         struct cl_page *page = slice->cpl_page;
 472
 473         /*
 474          * For transient pages, remove it from the radix tree.
 475          */
 476         cl_page_delete(env, page);
 477 }
 478
 479 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 480                                           const struct cl_page_slice *slice)
 481 {
 482         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 483         int     locked;
 484
 485         locked = !inode_trylock(inode);
 486         if (!locked)
 487                 inode_unlock(inode);
 488         return locked ? -EBUSY : -ENODATA;
 489 }
 490
 491 static void
 492 vvp_transient_page_completion(const struct lu_env *env,
 493                               const struct cl_page_slice *slice,
 494                               int ioret)
 495 {
 496 }
 497
 498 static void vvp_transient_page_fini(const struct lu_env *env,
 499                                     struct cl_page_slice *slice)
 500 {
 501         struct vvp_page *vpg = cl2vvp_page(slice);
 502         struct cl_page *clp = slice->cpl_page;
 503         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 504
 505         vvp_page_fini_common(vpg);
 506         atomic_dec(&clobj->vob_transient_pages);
 507 }
 508
 509 static const struct cl_page_operations vvp_transient_page_ops = {
 510         .cpo_own           = vvp_transient_page_own,
 511         .cpo_assume     = vvp_transient_page_assume,
 512         .cpo_unassume      = vvp_transient_page_unassume,
 513         .cpo_disown     = vvp_transient_page_disown,
 514         .cpo_discard       = vvp_transient_page_discard,
 515         .cpo_fini         = vvp_transient_page_fini,
 516         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 517         .cpo_print       = vvp_page_print,
 518         .cpo_is_under_lock      = vvp_page_is_under_lock,
 519         .io = {
 520                 [CRT_READ] = {
 521                         .cpo_prep       = vvp_transient_page_prep,
 522                         .cpo_completion  = vvp_transient_page_completion,
 523                 },
 524                 [CRT_WRITE] = {
 525                         .cpo_prep       = vvp_transient_page_prep,
 526                         .cpo_completion  = vvp_transient_page_completion,
 527                 }
 528         }
 529 };
 530
 531 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 532                   struct cl_page *page, pgoff_t index)
 533 {
 534         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 535         struct page     *vmpage = page->cp_vmpage;
 536
 537         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 538
 539         vpg->vpg_page = vmpage;
 540         get_page(vmpage);
 541
 542         INIT_LIST_HEAD(&vpg->vpg_pending_linkage);
 543         if (page->cp_type == CPT_CACHEABLE) {
 544                 /* in cache, decref in vvp_page_delete */
 545                 atomic_inc(&page->cp_ref);
 546                 SetPagePrivate(vmpage);
 547                 vmpage->private = (unsigned long)page;
 548                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 549                                   &vvp_page_ops);
 550         } else {
 551                 struct vvp_object *clobj = cl2vvp(obj);
 552
 553                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 554                                   &vvp_transient_page_ops);
 555                 atomic_inc(&clobj->vob_transient_pages);
 556         }
 557         return 0;
 558 }