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, 2016, 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 <libcfs/libcfs.h>
  48 #include "llite_internal.h"
  49 #include "vvp_internal.h"
  50
  51 /*****************************************************************************
  52  *
  53  * Page operations.
  54  *
  55  */
  56
  57 static void vvp_page_fini_common(struct vvp_page *vpg)
  58 {
  59         struct page *vmpage = vpg->vpg_page;
  60
  61         LASSERT(vmpage != NULL);
  62         put_page(vmpage);
  63 }
  64
  65 static void vvp_page_fini(const struct lu_env *env,
  66                           struct cl_page_slice *slice)
  67 {
  68         struct vvp_page *vpg     = cl2vvp_page(slice);
  69         struct page     *vmpage  = vpg->vpg_page;
  70
  71         /*
  72          * vmpage->private was already cleared when page was moved into
  73          * VPG_FREEING state.
  74          */
  75         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  76         vvp_page_fini_common(vpg);
  77 }
  78
  79 static int vvp_page_own(const struct lu_env *env,
  80                         const struct cl_page_slice *slice, struct cl_io *io,
  81                         int nonblock)
  82 {
  83         struct vvp_page *vpg    = cl2vvp_page(slice);
  84         struct page     *vmpage = vpg->vpg_page;
  85
  86         LASSERT(vmpage != NULL);
  87         if (nonblock) {
  88                 if (!trylock_page(vmpage))
  89                         return -EAGAIN;
  90
  91                 if (unlikely(PageWriteback(vmpage))) {
  92                         unlock_page(vmpage);
  93                         return -EAGAIN;
  94                 }
  95
  96                 return 0;
  97         }
  98
  99         lock_page(vmpage);
 100         wait_on_page_writeback(vmpage);
 101
 102         return 0;
 103 }
 104
 105 static void vvp_page_assume(const struct lu_env *env,
 106                             const struct cl_page_slice *slice,
 107                             struct cl_io *unused)
 108 {
 109         struct page *vmpage = cl2vm_page(slice);
 110
 111         LASSERT(vmpage != NULL);
 112         LASSERT(PageLocked(vmpage));
 113         wait_on_page_writeback(vmpage);
 114 }
 115
 116 static void vvp_page_unassume(const struct lu_env *env,
 117                               const struct cl_page_slice *slice,
 118                               struct cl_io *unused)
 119 {
 120         struct page *vmpage = cl2vm_page(slice);
 121
 122         LASSERT(vmpage != NULL);
 123         LASSERT(PageLocked(vmpage));
 124 }
 125
 126 static void vvp_page_disown(const struct lu_env *env,
 127                             const struct cl_page_slice *slice, struct cl_io *io)
 128 {
 129         struct page *vmpage = cl2vm_page(slice);
 130
 131         LASSERT(vmpage != NULL);
 132         LASSERT(PageLocked(vmpage));
 133
 134         unlock_page(cl2vm_page(slice));
 135 }
 136
 137 static void vvp_page_discard(const struct lu_env *env,
 138                              const struct cl_page_slice *slice,
 139                              struct cl_io *unused)
 140 {
 141         struct page     *vmpage = cl2vm_page(slice);
 142         struct vvp_page *vpg    = cl2vvp_page(slice);
 143
 144         LASSERT(vmpage != NULL);
 145         LASSERT(PageLocked(vmpage));
 146
 147         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used)
 148                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 149
 150         ll_invalidate_page(vmpage);
 151 }
 152
 153 static void vvp_page_delete(const struct lu_env *env,
 154                             const struct cl_page_slice *slice)
 155 {
 156         struct page      *vmpage = cl2vm_page(slice);
 157         struct inode     *inode  = vmpage->mapping->host;
 158         struct cl_object *obj    = slice->cpl_obj;
 159         struct cl_page   *page   = slice->cpl_page;
 160         int refc;
 161
 162         LASSERT(PageLocked(vmpage));
 163         LASSERT((struct cl_page *)vmpage->private == page);
 164         LASSERT(inode == vvp_object_inode(obj));
 165
 166         /* Drop the reference count held in vvp_page_init */
 167         refc = atomic_dec_return(&page->cp_ref);
 168         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 169
 170         ClearPagePrivate(vmpage);
 171         vmpage->private = 0;
 172         /*
 173          * Reference from vmpage to cl_page is removed, but the reference back
 174          * is still here. It is removed later in vvp_page_fini().
 175          */
 176 }
 177
 178 static void vvp_page_export(const struct lu_env *env,
 179                             const struct cl_page_slice *slice,
 180                             int uptodate)
 181 {
 182         struct page *vmpage = cl2vm_page(slice);
 183
 184         LASSERT(vmpage != NULL);
 185         LASSERT(PageLocked(vmpage));
 186         if (uptodate)
 187                 SetPageUptodate(vmpage);
 188         else
 189                 ClearPageUptodate(vmpage);
 190 }
 191
 192 static int vvp_page_is_vmlocked(const struct lu_env *env,
 193                                 const struct cl_page_slice *slice)
 194 {
 195         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 196 }
 197
 198 static int vvp_page_prep_read(const struct lu_env *env,
 199                               const struct cl_page_slice *slice,
 200                               struct cl_io *unused)
 201 {
 202         ENTRY;
 203         /* Skip the page already marked as PG_uptodate. */
 204         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 205 }
 206
 207 static int vvp_page_prep_write(const struct lu_env *env,
 208                                const struct cl_page_slice *slice,
 209                                struct cl_io *unused)
 210 {
 211         struct page *vmpage = cl2vm_page(slice);
 212         struct cl_page *pg = slice->cpl_page;
 213
 214         LASSERT(PageLocked(vmpage));
 215         LASSERT(!PageDirty(vmpage));
 216
 217         /* ll_writepage path is not a sync write, so need to set page writeback
 218          * flag */
 219         if (pg->cp_sync_io == NULL)
 220                 set_page_writeback(vmpage);
 221
 222         return 0;
 223 }
 224
 225 /**
 226  * Handles page transfer errors at VM level.
 227  *
 228  * This takes inode as a separate argument, because inode on which error is to
 229  * be set can be different from \a vmpage inode in case of direct-io.
 230  */
 231 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 232 {
 233         struct vvp_object *obj = cl_inode2vvp(inode);
 234
 235         if (ioret == 0) {
 236                 ClearPageError(vmpage);
 237                 obj->vob_discard_page_warned = 0;
 238         } else {
 239                 SetPageError(vmpage);
 240                 if (ioret == -ENOSPC)
 241                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 242                 else
 243                         set_bit(AS_EIO, &inode->i_mapping->flags);
 244
 245                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 246                      obj->vob_discard_page_warned == 0) {
 247                         obj->vob_discard_page_warned = 1;
 248                         ll_dirty_page_discard_warn(vmpage, ioret);
 249                 }
 250         }
 251 }
 252
 253 static void vvp_page_completion_read(const struct lu_env *env,
 254                                      const struct cl_page_slice *slice,
 255                                      int ioret)
 256 {
 257         struct vvp_page *vpg    = cl2vvp_page(slice);
 258         struct page     *vmpage = vpg->vpg_page;
 259         struct cl_page  *page   = slice->cpl_page;
 260         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 261         ENTRY;
 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 if (vpg->vpg_defer_uptodate) {
 273                 vpg->vpg_defer_uptodate = 0;
 274                 if (ioret == -EWOULDBLOCK) {
 275                         /* mirror read failed, it needs to destroy the page
 276                          * because subpage would be from wrong osc when trying
 277                          * to read from a new mirror */
 278                         ll_invalidate_page(vmpage);
 279                 }
 280         }
 281
 282         if (page->cp_sync_io == NULL)
 283                 unlock_page(vmpage);
 284
 285         EXIT;
 286 }
 287
 288 static void vvp_page_completion_write(const struct lu_env *env,
 289                                       const struct cl_page_slice *slice,
 290                                       int ioret)
 291 {
 292         struct vvp_page *vpg    = cl2vvp_page(slice);
 293         struct cl_page  *pg     = slice->cpl_page;
 294         struct page     *vmpage = vpg->vpg_page;
 295         ENTRY;
 296
 297         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 298
 299         if (pg->cp_sync_io != NULL) {
 300                 LASSERT(PageLocked(vmpage));
 301                 LASSERT(!PageWriteback(vmpage));
 302         } else {
 303                 LASSERT(PageWriteback(vmpage));
 304                 /*
 305                  * Only mark the page error only when it's an async write
 306                  * because applications won't wait for IO to finish.
 307                  */
 308                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 309
 310                 end_page_writeback(vmpage);
 311         }
 312         EXIT;
 313 }
 314
 315 /**
 316  * Implements cl_page_operations::cpo_make_ready() method.
 317  *
 318  * This is called to yank a page from the transfer cache and to send it out as
 319  * a part of transfer. This function try-locks the page. If try-lock failed,
 320  * page is owned by some concurrent IO, and should be skipped (this is bad,
 321  * but hopefully rare situation, as it usually results in transfer being
 322  * shorter than possible).
 323  *
 324  * \retval 0      success, page can be placed into transfer
 325  *
 326  * \retval -EAGAIN page is either used by concurrent IO has been
 327  * truncated. Skip it.
 328  */
 329 static int vvp_page_make_ready(const struct lu_env *env,
 330                                const struct cl_page_slice *slice)
 331 {
 332         struct page *vmpage = cl2vm_page(slice);
 333         struct cl_page *pg = slice->cpl_page;
 334         int result = 0;
 335
 336         lock_page(vmpage);
 337         if (clear_page_dirty_for_io(vmpage)) {
 338                 LASSERT(pg->cp_state == CPS_CACHED);
 339                 /* This actually clears the dirty bit in the radix
 340                  * tree. */
 341                 set_page_writeback(vmpage);
 342                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 343         } else if (pg->cp_state == CPS_PAGEOUT) {
 344                 /* is it possible for osc_flush_async_page() to already
 345                  * make it ready? */
 346                 result = -EALREADY;
 347         } else {
 348                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 349                               pg->cp_state);
 350                 LBUG();
 351         }
 352         unlock_page(vmpage);
 353         RETURN(result);
 354 }
 355
 356 static int vvp_page_print(const struct lu_env *env,
 357                           const struct cl_page_slice *slice,
 358                           void *cookie, lu_printer_t printer)
 359 {
 360         struct vvp_page *vpg    = cl2vvp_page(slice);
 361         struct page     *vmpage = vpg->vpg_page;
 362
 363         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d) "
 364                    "vm@%p ",
 365                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used, vmpage);
 366
 367         if (vmpage != NULL) {
 368                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 369                            (long)vmpage->flags, page_count(vmpage),
 370                            page_mapcount(vmpage), vmpage->private,
 371                            page_index(vmpage),
 372                            list_empty(&vmpage->lru) ? "not-" : "");
 373         }
 374
 375         (*printer)(env, cookie, "\n");
 376
 377         return 0;
 378 }
 379
 380 static int vvp_page_fail(const struct lu_env *env,
 381                          const struct cl_page_slice *slice)
 382 {
 383         /*
 384          * Cached read?
 385          */
 386         LBUG();
 387
 388         return 0;
 389 }
 390
 391 static const struct cl_page_operations vvp_page_ops = {
 392         .cpo_own           = vvp_page_own,
 393         .cpo_assume        = vvp_page_assume,
 394         .cpo_unassume      = vvp_page_unassume,
 395         .cpo_disown        = vvp_page_disown,
 396         .cpo_discard       = vvp_page_discard,
 397         .cpo_delete        = vvp_page_delete,
 398         .cpo_export        = vvp_page_export,
 399         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 400         .cpo_fini          = vvp_page_fini,
 401         .cpo_print         = vvp_page_print,
 402         .io = {
 403                 [CRT_READ] = {
 404                         .cpo_prep       = vvp_page_prep_read,
 405                         .cpo_completion = vvp_page_completion_read,
 406                         .cpo_make_ready = vvp_page_fail,
 407                 },
 408                 [CRT_WRITE] = {
 409                         .cpo_prep       = vvp_page_prep_write,
 410                         .cpo_completion = vvp_page_completion_write,
 411                         .cpo_make_ready = vvp_page_make_ready,
 412                 },
 413         },
 414 };
 415
 416 static int vvp_transient_page_prep(const struct lu_env *env,
 417                                    const struct cl_page_slice *slice,
 418                                    struct cl_io *unused)
 419 {
 420         ENTRY;
 421         /* transient page should always be sent. */
 422         RETURN(0);
 423 }
 424
 425 static void vvp_transient_page_verify(const struct cl_page *page)
 426 {
 427 }
 428
 429 static int vvp_transient_page_own(const struct lu_env *env,
 430                                   const struct cl_page_slice *slice,
 431                                   struct cl_io *unused, int nonblock)
 432 {
 433         vvp_transient_page_verify(slice->cpl_page);
 434         return 0;
 435 }
 436
 437 static void vvp_transient_page_assume(const struct lu_env *env,
 438                                       const struct cl_page_slice *slice,
 439                                       struct cl_io *unused)
 440 {
 441         vvp_transient_page_verify(slice->cpl_page);
 442 }
 443
 444 static void vvp_transient_page_unassume(const struct lu_env *env,
 445                                         const struct cl_page_slice *slice,
 446                                         struct cl_io *unused)
 447 {
 448         vvp_transient_page_verify(slice->cpl_page);
 449 }
 450
 451 static void vvp_transient_page_disown(const struct lu_env *env,
 452                                       const struct cl_page_slice *slice,
 453                                       struct cl_io *unused)
 454 {
 455         vvp_transient_page_verify(slice->cpl_page);
 456 }
 457
 458 static void vvp_transient_page_discard(const struct lu_env *env,
 459                                        const struct cl_page_slice *slice,
 460                                        struct cl_io *unused)
 461 {
 462         struct cl_page *page = slice->cpl_page;
 463
 464         vvp_transient_page_verify(slice->cpl_page);
 465
 466         /*
 467          * For transient pages, remove it from the radix tree.
 468          */
 469         cl_page_delete(env, page);
 470 }
 471
 472 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 473                                           const struct cl_page_slice *slice)
 474 {
 475         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 476         int     locked;
 477
 478         locked = !inode_trylock(inode);
 479         if (!locked)
 480                 inode_unlock(inode);
 481         return locked ? -EBUSY : -ENODATA;
 482 }
 483
 484 static void
 485 vvp_transient_page_completion(const struct lu_env *env,
 486                               const struct cl_page_slice *slice,
 487                               int ioret)
 488 {
 489         vvp_transient_page_verify(slice->cpl_page);
 490 }
 491
 492 static void vvp_transient_page_fini(const struct lu_env *env,
 493                                     struct cl_page_slice *slice)
 494 {
 495         struct vvp_page *vpg = cl2vvp_page(slice);
 496         struct cl_page *clp = slice->cpl_page;
 497         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 498
 499         vvp_page_fini_common(vpg);
 500         atomic_dec(&clobj->vob_transient_pages);
 501 }
 502
 503 static const struct cl_page_operations vvp_transient_page_ops = {
 504         .cpo_own                = vvp_transient_page_own,
 505         .cpo_assume             = vvp_transient_page_assume,
 506         .cpo_unassume           = vvp_transient_page_unassume,
 507         .cpo_disown             = vvp_transient_page_disown,
 508         .cpo_discard            = vvp_transient_page_discard,
 509         .cpo_fini               = vvp_transient_page_fini,
 510         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 511         .cpo_print              = vvp_page_print,
 512         .io = {
 513                 [CRT_READ] = {
 514                         .cpo_prep       = vvp_transient_page_prep,
 515                         .cpo_completion = vvp_transient_page_completion,
 516                 },
 517                 [CRT_WRITE] = {
 518                         .cpo_prep       = vvp_transient_page_prep,
 519                         .cpo_completion = vvp_transient_page_completion,
 520                 }
 521         }
 522 };
 523
 524 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 525                 struct cl_page *page, pgoff_t index)
 526 {
 527         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 528         struct page     *vmpage = page->cp_vmpage;
 529
 530         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 531
 532         vpg->vpg_page = vmpage;
 533         get_page(vmpage);
 534
 535         if (page->cp_type == CPT_CACHEABLE) {
 536                 /* in cache, decref in vvp_page_delete */
 537                 atomic_inc(&page->cp_ref);
 538                 SetPagePrivate(vmpage);
 539                 vmpage->private = (unsigned long)page;
 540                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 541                                 &vvp_page_ops);
 542         } else {
 543                 struct vvp_object *clobj = cl2vvp(obj);
 544
 545                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 546                                 &vvp_transient_page_ops);
 547                 atomic_inc(&clobj->vob_transient_pages);
 548         }
 549         return 0;
 550 }