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, 2017, 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, struct pagevec *pvec)
  58 {
  59         struct page *vmpage = vpg->vpg_page;
  60
  61         LASSERT(vmpage != NULL);
  62         if (pvec) {
  63                 if (!pagevec_add(pvec, vmpage))
  64                         pagevec_release(pvec);
  65         } else {
  66                 put_page(vmpage);
  67         }
  68 }
  69
  70 static void vvp_page_fini(const struct lu_env *env,
  71                           struct cl_page_slice *slice,
  72                           struct pagevec *pvec)
  73 {
  74         struct vvp_page *vpg     = cl2vvp_page(slice);
  75         struct page     *vmpage  = vpg->vpg_page;
  76
  77         /*
  78          * vmpage->private was already cleared when page was moved into
  79          * VPG_FREEING state.
  80          */
  81         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  82         vvp_page_fini_common(vpg, pvec);
  83 }
  84
  85 static int vvp_page_own(const struct lu_env *env,
  86                         const struct cl_page_slice *slice, struct cl_io *io,
  87                         int nonblock)
  88 {
  89         struct vvp_page *vpg    = cl2vvp_page(slice);
  90         struct page     *vmpage = vpg->vpg_page;
  91
  92         LASSERT(vmpage != NULL);
  93         if (nonblock) {
  94                 if (!trylock_page(vmpage))
  95                         return -EAGAIN;
  96
  97                 if (unlikely(PageWriteback(vmpage))) {
  98                         unlock_page(vmpage);
  99                         return -EAGAIN;
 100                 }
 101
 102                 return 0;
 103         }
 104
 105         lock_page(vmpage);
 106         wait_on_page_writeback(vmpage);
 107
 108         return 0;
 109 }
 110
 111 static void vvp_page_assume(const struct lu_env *env,
 112                             const struct cl_page_slice *slice,
 113                             struct cl_io *unused)
 114 {
 115         struct page *vmpage = cl2vm_page(slice);
 116
 117         LASSERT(vmpage != NULL);
 118         LASSERT(PageLocked(vmpage));
 119         wait_on_page_writeback(vmpage);
 120 }
 121
 122 static void vvp_page_unassume(const struct lu_env *env,
 123                               const struct cl_page_slice *slice,
 124                               struct cl_io *unused)
 125 {
 126         struct page *vmpage = cl2vm_page(slice);
 127
 128         LASSERT(vmpage != NULL);
 129         LASSERT(PageLocked(vmpage));
 130 }
 131
 132 static void vvp_page_disown(const struct lu_env *env,
 133                             const struct cl_page_slice *slice, struct cl_io *io)
 134 {
 135         struct page *vmpage = cl2vm_page(slice);
 136
 137         LASSERT(vmpage != NULL);
 138         LASSERT(PageLocked(vmpage));
 139
 140         unlock_page(cl2vm_page(slice));
 141 }
 142
 143 static void vvp_page_discard(const struct lu_env *env,
 144                              const struct cl_page_slice *slice,
 145                              struct cl_io *unused)
 146 {
 147         struct page     *vmpage = cl2vm_page(slice);
 148         struct vvp_page *vpg    = cl2vvp_page(slice);
 149
 150         LASSERT(vmpage != NULL);
 151         LASSERT(PageLocked(vmpage));
 152
 153         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used)
 154                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 155
 156         ll_invalidate_page(vmpage);
 157 }
 158
 159 static void vvp_page_delete(const struct lu_env *env,
 160                             const struct cl_page_slice *slice)
 161 {
 162         struct page      *vmpage = cl2vm_page(slice);
 163         struct inode     *inode  = vmpage->mapping->host;
 164         struct cl_object *obj    = slice->cpl_obj;
 165         struct cl_page   *page   = slice->cpl_page;
 166         int refc;
 167
 168         LASSERT(PageLocked(vmpage));
 169         LASSERT((struct cl_page *)vmpage->private == page);
 170         LASSERT(inode == vvp_object_inode(obj));
 171
 172         /* Drop the reference count held in vvp_page_init */
 173         refc = atomic_dec_return(&page->cp_ref);
 174         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 175
 176         ClearPagePrivate(vmpage);
 177         vmpage->private = 0;
 178         /*
 179          * Reference from vmpage to cl_page is removed, but the reference back
 180          * is still here. It is removed later in vvp_page_fini().
 181          */
 182 }
 183
 184 static void vvp_page_export(const struct lu_env *env,
 185                             const struct cl_page_slice *slice,
 186                             int uptodate)
 187 {
 188         struct page *vmpage = cl2vm_page(slice);
 189
 190         LASSERT(vmpage != NULL);
 191         LASSERT(PageLocked(vmpage));
 192         if (uptodate)
 193                 SetPageUptodate(vmpage);
 194         else
 195                 ClearPageUptodate(vmpage);
 196 }
 197
 198 static int vvp_page_is_vmlocked(const struct lu_env *env,
 199                                 const struct cl_page_slice *slice)
 200 {
 201         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 202 }
 203
 204 static int vvp_page_prep_read(const struct lu_env *env,
 205                               const struct cl_page_slice *slice,
 206                               struct cl_io *unused)
 207 {
 208         ENTRY;
 209         /* Skip the page already marked as PG_uptodate. */
 210         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 211 }
 212
 213 static int vvp_page_prep_write(const struct lu_env *env,
 214                                const struct cl_page_slice *slice,
 215                                struct cl_io *unused)
 216 {
 217         struct page *vmpage = cl2vm_page(slice);
 218         struct cl_page *pg = slice->cpl_page;
 219
 220         LASSERT(PageLocked(vmpage));
 221         LASSERT(!PageDirty(vmpage));
 222
 223         /* ll_writepage path is not a sync write, so need to set page writeback
 224          * flag */
 225         if (pg->cp_sync_io == NULL)
 226                 set_page_writeback(vmpage);
 227
 228         return 0;
 229 }
 230
 231 /**
 232  * Handles page transfer errors at VM level.
 233  *
 234  * This takes inode as a separate argument, because inode on which error is to
 235  * be set can be different from \a vmpage inode in case of direct-io.
 236  */
 237 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 238 {
 239         struct vvp_object *obj = cl_inode2vvp(inode);
 240
 241         if (ioret == 0) {
 242                 ClearPageError(vmpage);
 243                 obj->vob_discard_page_warned = 0;
 244         } else {
 245                 SetPageError(vmpage);
 246                 if (ioret == -ENOSPC)
 247                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 248                 else
 249                         set_bit(AS_EIO, &inode->i_mapping->flags);
 250
 251                 if ((ioret == -ESHUTDOWN || ioret == -EINTR ||
 252                      ioret == -EIO) && obj->vob_discard_page_warned == 0) {
 253                         obj->vob_discard_page_warned = 1;
 254                         ll_dirty_page_discard_warn(vmpage, ioret);
 255                 }
 256         }
 257 }
 258
 259 static void vvp_page_completion_read(const struct lu_env *env,
 260                                      const struct cl_page_slice *slice,
 261                                      int ioret)
 262 {
 263         struct vvp_page *vpg    = cl2vvp_page(slice);
 264         struct page     *vmpage = vpg->vpg_page;
 265         struct cl_page  *page   = slice->cpl_page;
 266         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 267         ENTRY;
 268
 269         LASSERT(PageLocked(vmpage));
 270         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 271
 272         if (vpg->vpg_defer_uptodate)
 273                 ll_ra_count_put(ll_i2sbi(inode), 1);
 274
 275         if (ioret == 0)  {
 276                 if (!vpg->vpg_defer_uptodate)
 277                         cl_page_export(env, page, 1);
 278         } else if (vpg->vpg_defer_uptodate) {
 279                 vpg->vpg_defer_uptodate = 0;
 280                 if (ioret == -EWOULDBLOCK) {
 281                         /* mirror read failed, it needs to destroy the page
 282                          * because subpage would be from wrong osc when trying
 283                          * to read from a new mirror */
 284                         ll_invalidate_page(vmpage);
 285                 }
 286         }
 287
 288         if (page->cp_sync_io == NULL)
 289                 unlock_page(vmpage);
 290
 291         EXIT;
 292 }
 293
 294 static void vvp_page_completion_write(const struct lu_env *env,
 295                                       const struct cl_page_slice *slice,
 296                                       int ioret)
 297 {
 298         struct vvp_page *vpg    = cl2vvp_page(slice);
 299         struct cl_page  *pg     = slice->cpl_page;
 300         struct page     *vmpage = vpg->vpg_page;
 301         ENTRY;
 302
 303         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 304
 305         if (pg->cp_sync_io != NULL) {
 306                 LASSERT(PageLocked(vmpage));
 307                 LASSERT(!PageWriteback(vmpage));
 308         } else {
 309                 LASSERT(PageWriteback(vmpage));
 310                 /*
 311                  * Only mark the page error only when it's an async write
 312                  * because applications won't wait for IO to finish.
 313                  */
 314                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 315
 316                 end_page_writeback(vmpage);
 317         }
 318         EXIT;
 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
 346                  * tree. */
 347                 set_page_writeback(vmpage);
 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                 result = -EALREADY;
 353         } else {
 354                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 355                               pg->cp_state);
 356                 LBUG();
 357         }
 358         unlock_page(vmpage);
 359         RETURN(result);
 360 }
 361
 362 static int vvp_page_print(const struct lu_env *env,
 363                           const struct cl_page_slice *slice,
 364                           void *cookie, lu_printer_t printer)
 365 {
 366         struct vvp_page *vpg    = cl2vvp_page(slice);
 367         struct page     *vmpage = vpg->vpg_page;
 368
 369         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d) "
 370                    "vm@%p ",
 371                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used, vmpage);
 372
 373         if (vmpage != NULL) {
 374                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 375                            (long)vmpage->flags, page_count(vmpage),
 376                            page_mapcount(vmpage), vmpage->private,
 377                            page_index(vmpage),
 378                            list_empty(&vmpage->lru) ? "not-" : "");
 379         }
 380
 381         (*printer)(env, cookie, "\n");
 382
 383         return 0;
 384 }
 385
 386 static int vvp_page_fail(const struct lu_env *env,
 387                          const struct cl_page_slice *slice)
 388 {
 389         /*
 390          * Cached read?
 391          */
 392         LBUG();
 393
 394         return 0;
 395 }
 396
 397 static const struct cl_page_operations vvp_page_ops = {
 398         .cpo_own           = vvp_page_own,
 399         .cpo_assume        = vvp_page_assume,
 400         .cpo_unassume      = vvp_page_unassume,
 401         .cpo_disown        = vvp_page_disown,
 402         .cpo_discard       = vvp_page_discard,
 403         .cpo_delete        = vvp_page_delete,
 404         .cpo_export        = vvp_page_export,
 405         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 406         .cpo_fini          = vvp_page_fini,
 407         .cpo_print         = vvp_page_print,
 408         .io = {
 409                 [CRT_READ] = {
 410                         .cpo_prep       = vvp_page_prep_read,
 411                         .cpo_completion = vvp_page_completion_read,
 412                         .cpo_make_ready = vvp_page_fail,
 413                 },
 414                 [CRT_WRITE] = {
 415                         .cpo_prep       = vvp_page_prep_write,
 416                         .cpo_completion = vvp_page_completion_write,
 417                         .cpo_make_ready = vvp_page_make_ready,
 418                 },
 419         },
 420 };
 421
 422 static int vvp_transient_page_prep(const struct lu_env *env,
 423                                    const struct cl_page_slice *slice,
 424                                    struct cl_io *unused)
 425 {
 426         ENTRY;
 427         /* transient page should always be sent. */
 428         RETURN(0);
 429 }
 430
 431 static void vvp_transient_page_verify(const struct cl_page *page)
 432 {
 433 }
 434
 435 static int vvp_transient_page_own(const struct lu_env *env,
 436                                   const struct cl_page_slice *slice,
 437                                   struct cl_io *unused, int nonblock)
 438 {
 439         vvp_transient_page_verify(slice->cpl_page);
 440         return 0;
 441 }
 442
 443 static void vvp_transient_page_assume(const struct lu_env *env,
 444                                       const struct cl_page_slice *slice,
 445                                       struct cl_io *unused)
 446 {
 447         vvp_transient_page_verify(slice->cpl_page);
 448 }
 449
 450 static void vvp_transient_page_unassume(const struct lu_env *env,
 451                                         const struct cl_page_slice *slice,
 452                                         struct cl_io *unused)
 453 {
 454         vvp_transient_page_verify(slice->cpl_page);
 455 }
 456
 457 static void vvp_transient_page_disown(const struct lu_env *env,
 458                                       const struct cl_page_slice *slice,
 459                                       struct cl_io *unused)
 460 {
 461         vvp_transient_page_verify(slice->cpl_page);
 462 }
 463
 464 static void vvp_transient_page_discard(const struct lu_env *env,
 465                                        const struct cl_page_slice *slice,
 466                                        struct cl_io *unused)
 467 {
 468         struct cl_page *page = slice->cpl_page;
 469
 470         vvp_transient_page_verify(slice->cpl_page);
 471
 472         /*
 473          * For transient pages, remove it from the radix tree.
 474          */
 475         cl_page_delete(env, page);
 476 }
 477
 478 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 479                                           const struct cl_page_slice *slice)
 480 {
 481         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 482         int     locked;
 483
 484         locked = !inode_trylock(inode);
 485         if (!locked)
 486                 inode_unlock(inode);
 487         return locked ? -EBUSY : -ENODATA;
 488 }
 489
 490 static void
 491 vvp_transient_page_completion(const struct lu_env *env,
 492                               const struct cl_page_slice *slice,
 493                               int ioret)
 494 {
 495         vvp_transient_page_verify(slice->cpl_page);
 496 }
 497
 498 static void vvp_transient_page_fini(const struct lu_env *env,
 499                                     struct cl_page_slice *slice,
 500                                     struct pagevec *pvec)
 501 {
 502         struct vvp_page *vpg = cl2vvp_page(slice);
 503         struct cl_page *clp = slice->cpl_page;
 504         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 505
 506         vvp_page_fini_common(vpg, pvec);
 507         atomic_dec(&clobj->vob_transient_pages);
 508 }
 509
 510 static const struct cl_page_operations vvp_transient_page_ops = {
 511         .cpo_own                = vvp_transient_page_own,
 512         .cpo_assume             = vvp_transient_page_assume,
 513         .cpo_unassume           = vvp_transient_page_unassume,
 514         .cpo_disown             = vvp_transient_page_disown,
 515         .cpo_discard            = vvp_transient_page_discard,
 516         .cpo_fini               = vvp_transient_page_fini,
 517         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 518         .cpo_print              = vvp_page_print,
 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         if (page->cp_type == CPT_CACHEABLE) {
 543                 /* in cache, decref in vvp_page_delete */
 544                 atomic_inc(&page->cp_ref);
 545                 SetPagePrivate(vmpage);
 546                 vmpage->private = (unsigned long)page;
 547                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 548                                 &vvp_page_ops);
 549         } else {
 550                 struct vvp_object *clobj = cl2vvp(obj);
 551
 552                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 553                                 &vvp_transient_page_ops);
 554                 atomic_inc(&clobj->vob_transient_pages);
 555         }
 556         return 0;
 557 }