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                 if (ioret == -ENOSPC)
 245                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 246                 else
 247                         set_bit(AS_EIO, &inode->i_mapping->flags);
 248
 249                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 250                     obj->vob_discard_page_warned == 0) {
 251                         obj->vob_discard_page_warned = 1;
 252                         ll_dirty_page_discard_warn(vmpage, ioret);
 253                 }
 254         }
 255 }
 256
 257 static void vvp_page_completion_read(const struct lu_env *env,
 258                                      const struct cl_page_slice *slice,
 259                                      int ioret)
 260 {
 261         struct vvp_page *vpg    = cl2vvp_page(slice);
 262         struct page     *vmpage = vpg->vpg_page;
 263         struct cl_page  *page   = slice->cpl_page;
 264         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 265
 266         LASSERT(PageLocked(vmpage));
 267         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 268
 269         if (vpg->vpg_defer_uptodate)
 270                 ll_ra_count_put(ll_i2sbi(inode), 1);
 271
 272         if (ioret == 0)  {
 273                 if (!vpg->vpg_defer_uptodate)
 274                         cl_page_export(env, page, 1);
 275         } else {
 276                 vpg->vpg_defer_uptodate = 0;
 277         }
 278
 279         if (!page->cp_sync_io)
 280                 unlock_page(vmpage);
 281 }
 282
 283 static void vvp_page_completion_write(const struct lu_env *env,
 284                                       const struct cl_page_slice *slice,
 285                                       int ioret)
 286 {
 287         struct vvp_page *vpg     = cl2vvp_page(slice);
 288         struct cl_page  *pg     = slice->cpl_page;
 289         struct page     *vmpage = vpg->vpg_page;
 290
 291         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 292
 293         /*
 294          * TODO: Actually it makes sense to add the page into oap pending
 295          * list again and so that we don't need to take the page out from
 296          * SoM write pending list, if we just meet a recoverable error,
 297          * -ENOMEM, etc.
 298          * To implement this, we just need to return a non zero value in
 299          * ->cpo_completion method. The underlying transfer should be notified
 300          * and then re-add the page into pending transfer queue.  -jay
 301          */
 302
 303         vpg->vpg_write_queued = 0;
 304         vvp_write_complete(cl2vvp(slice->cpl_obj), vpg);
 305
 306         if (pg->cp_sync_io) {
 307                 LASSERT(PageLocked(vmpage));
 308                 LASSERT(!PageWriteback(vmpage));
 309         } else {
 310                 LASSERT(PageWriteback(vmpage));
 311                 /*
 312                  * Only mark the page error only when it's an async write
 313                  * because applications won't wait for IO to finish.
 314                  */
 315                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 316
 317                 end_page_writeback(vmpage);
 318         }
 319 }
 320
 321 /**
 322  * Implements cl_page_operations::cpo_make_ready() method.
 323  *
 324  * This is called to yank a page from the transfer cache and to send it out as
 325  * a part of transfer. This function try-locks the page. If try-lock failed,
 326  * page is owned by some concurrent IO, and should be skipped (this is bad,
 327  * but hopefully rare situation, as it usually results in transfer being
 328  * shorter than possible).
 329  *
 330  * \retval 0      success, page can be placed into transfer
 331  *
 332  * \retval -EAGAIN page is either used by concurrent IO has been
 333  * truncated. Skip it.
 334  */
 335 static int vvp_page_make_ready(const struct lu_env *env,
 336                                const struct cl_page_slice *slice)
 337 {
 338         struct page *vmpage = cl2vm_page(slice);
 339         struct cl_page *pg = slice->cpl_page;
 340         int result = 0;
 341
 342         lock_page(vmpage);
 343         if (clear_page_dirty_for_io(vmpage)) {
 344                 LASSERT(pg->cp_state == CPS_CACHED);
 345                 /* This actually clears the dirty bit in the radix tree. */
 346                 set_page_writeback(vmpage);
 347                 vvp_write_pending(cl2vvp(slice->cpl_obj), cl2vvp_page(slice));
 348                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 349         } else if (pg->cp_state == CPS_PAGEOUT) {
 350                 /* is it possible for osc_flush_async_page() to already
 351                  * make it ready?
 352                  */
 353                 result = -EALREADY;
 354         } else {
 355                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 356                               pg->cp_state);
 357                 LBUG();
 358         }
 359         unlock_page(vmpage);
 360         return result;
 361 }
 362
 363 static int vvp_page_is_under_lock(const struct lu_env *env,
 364                                   const struct cl_page_slice *slice,
 365                                   struct cl_io *io, pgoff_t *max_index)
 366 {
 367         if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
 368             io->ci_type == CIT_FAULT) {
 369                 struct vvp_io *vio = vvp_env_io(env);
 370
 371                 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED))
 372                         *max_index = CL_PAGE_EOF;
 373         }
 374         return 0;
 375 }
 376
 377 static int vvp_page_print(const struct lu_env *env,
 378                           const struct cl_page_slice *slice,
 379                           void *cookie, lu_printer_t printer)
 380 {
 381         struct vvp_page *vpg = cl2vvp_page(slice);
 382         struct page     *vmpage = vpg->vpg_page;
 383
 384         (*printer)(env, cookie, LUSTRE_VVP_NAME "-page@%p(%d:%d:%d) vm@%p ",
 385                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used,
 386                    vpg->vpg_write_queued, vmpage);
 387         if (vmpage) {
 388                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 389                            (long)vmpage->flags, page_count(vmpage),
 390                            page_mapcount(vmpage), vmpage->private,
 391                            vmpage->index,
 392                            list_empty(&vmpage->lru) ? "not-" : "");
 393         }
 394
 395         (*printer)(env, cookie, "\n");
 396
 397         return 0;
 398 }
 399
 400 static int vvp_page_fail(const struct lu_env *env,
 401                          const struct cl_page_slice *slice)
 402 {
 403         /*
 404          * Cached read?
 405          */
 406         LBUG();
 407
 408         return 0;
 409 }
 410
 411 static const struct cl_page_operations vvp_page_ops = {
 412         .cpo_own           = vvp_page_own,
 413         .cpo_assume     = vvp_page_assume,
 414         .cpo_unassume      = vvp_page_unassume,
 415         .cpo_disown     = vvp_page_disown,
 416         .cpo_discard       = vvp_page_discard,
 417         .cpo_delete     = vvp_page_delete,
 418         .cpo_export     = vvp_page_export,
 419         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 420         .cpo_fini         = vvp_page_fini,
 421         .cpo_print       = vvp_page_print,
 422         .cpo_is_under_lock = vvp_page_is_under_lock,
 423         .io = {
 424                 [CRT_READ] = {
 425                         .cpo_prep       = vvp_page_prep_read,
 426                         .cpo_completion  = vvp_page_completion_read,
 427                         .cpo_make_ready = vvp_page_fail,
 428                 },
 429                 [CRT_WRITE] = {
 430                         .cpo_prep       = vvp_page_prep_write,
 431                         .cpo_completion  = vvp_page_completion_write,
 432                         .cpo_make_ready  = vvp_page_make_ready,
 433                 },
 434         },
 435 };
 436
 437 static int vvp_transient_page_prep(const struct lu_env *env,
 438                                    const struct cl_page_slice *slice,
 439                                    struct cl_io *unused)
 440 {
 441         /* transient page should always be sent. */
 442         return 0;
 443 }
 444
 445 static int vvp_transient_page_own(const struct lu_env *env,
 446                                   const struct cl_page_slice *slice,
 447                                   struct cl_io *unused, int nonblock)
 448 {
 449         return 0;
 450 }
 451
 452 static void vvp_transient_page_assume(const struct lu_env *env,
 453                                       const struct cl_page_slice *slice,
 454                                       struct cl_io *unused)
 455 {
 456 }
 457
 458 static void vvp_transient_page_unassume(const struct lu_env *env,
 459                                         const struct cl_page_slice *slice,
 460                                         struct cl_io *unused)
 461 {
 462 }
 463
 464 static void vvp_transient_page_disown(const struct lu_env *env,
 465                                       const struct cl_page_slice *slice,
 466                                       struct cl_io *unused)
 467 {
 468 }
 469
 470 static void vvp_transient_page_discard(const struct lu_env *env,
 471                                        const struct cl_page_slice *slice,
 472                                        struct cl_io *unused)
 473 {
 474         struct cl_page *page = slice->cpl_page;
 475
 476         /*
 477          * For transient pages, remove it from the radix tree.
 478          */
 479         cl_page_delete(env, page);
 480 }
 481
 482 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 483                                           const struct cl_page_slice *slice)
 484 {
 485         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 486         int     locked;
 487
 488         locked = !inode_trylock(inode);
 489         if (!locked)
 490                 inode_unlock(inode);
 491         return locked ? -EBUSY : -ENODATA;
 492 }
 493
 494 static void
 495 vvp_transient_page_completion(const struct lu_env *env,
 496                               const struct cl_page_slice *slice,
 497                               int ioret)
 498 {
 499 }
 500
 501 static void vvp_transient_page_fini(const struct lu_env *env,
 502                                     struct cl_page_slice *slice)
 503 {
 504         struct vvp_page *vpg = cl2vvp_page(slice);
 505         struct cl_page *clp = slice->cpl_page;
 506         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 507
 508         vvp_page_fini_common(vpg);
 509         clobj->vob_transient_pages--;
 510 }
 511
 512 static const struct cl_page_operations vvp_transient_page_ops = {
 513         .cpo_own           = vvp_transient_page_own,
 514         .cpo_assume     = vvp_transient_page_assume,
 515         .cpo_unassume      = vvp_transient_page_unassume,
 516         .cpo_disown     = vvp_transient_page_disown,
 517         .cpo_discard       = vvp_transient_page_discard,
 518         .cpo_fini         = vvp_transient_page_fini,
 519         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 520         .cpo_print       = vvp_page_print,
 521         .cpo_is_under_lock      = vvp_page_is_under_lock,
 522         .io = {
 523                 [CRT_READ] = {
 524                         .cpo_prep       = vvp_transient_page_prep,
 525                         .cpo_completion  = vvp_transient_page_completion,
 526                 },
 527                 [CRT_WRITE] = {
 528                         .cpo_prep       = vvp_transient_page_prep,
 529                         .cpo_completion  = vvp_transient_page_completion,
 530                 }
 531         }
 532 };
 533
 534 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 535                   struct cl_page *page, pgoff_t index)
 536 {
 537         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 538         struct page     *vmpage = page->cp_vmpage;
 539
 540         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 541
 542         vpg->vpg_page = vmpage;
 543         get_page(vmpage);
 544
 545         INIT_LIST_HEAD(&vpg->vpg_pending_linkage);
 546         if (page->cp_type == CPT_CACHEABLE) {
 547                 /* in cache, decref in vvp_page_delete */
 548                 atomic_inc(&page->cp_ref);
 549                 SetPagePrivate(vmpage);
 550                 vmpage->private = (unsigned long)page;
 551                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 552                                   &vvp_page_ops);
 553         } else {
 554                 struct vvp_object *clobj = cl2vvp(obj);
 555
 556                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 557                                   &vvp_transient_page_ops);
 558                 clobj->vob_transient_pages++;
 559         }
 560         return 0;
 561 }