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 && vmpage->mapping)
 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 cl_page   *page   = slice->cpl_page;
 164         int refc;
 165
 166         LASSERT(PageLocked(vmpage));
 167         LASSERT((struct cl_page *)vmpage->private == page);
 168
 169
 170         /* Drop the reference count held in vvp_page_init */
 171         refc = atomic_dec_return(&page->cp_ref);
 172         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 173
 174         ClearPagePrivate(vmpage);
 175         vmpage->private = 0;
 176         /*
 177          * Reference from vmpage to cl_page is removed, but the reference back
 178          * is still here. It is removed later in vvp_page_fini().
 179          */
 180 }
 181
 182 static void vvp_page_export(const struct lu_env *env,
 183                             const struct cl_page_slice *slice,
 184                             int uptodate)
 185 {
 186         struct page *vmpage = cl2vm_page(slice);
 187
 188         LASSERT(vmpage != NULL);
 189         LASSERT(PageLocked(vmpage));
 190         if (uptodate)
 191                 SetPageUptodate(vmpage);
 192         else
 193                 ClearPageUptodate(vmpage);
 194 }
 195
 196 static int vvp_page_is_vmlocked(const struct lu_env *env,
 197                                 const struct cl_page_slice *slice)
 198 {
 199         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 200 }
 201
 202 static int vvp_page_prep_read(const struct lu_env *env,
 203                               const struct cl_page_slice *slice,
 204                               struct cl_io *unused)
 205 {
 206         ENTRY;
 207         /* Skip the page already marked as PG_uptodate. */
 208         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 209 }
 210
 211 static int vvp_page_prep_write(const struct lu_env *env,
 212                                const struct cl_page_slice *slice,
 213                                struct cl_io *unused)
 214 {
 215         struct page *vmpage = cl2vm_page(slice);
 216         struct cl_page *pg = slice->cpl_page;
 217
 218         LASSERT(PageLocked(vmpage));
 219         LASSERT(!PageDirty(vmpage));
 220
 221         /* ll_writepage path is not a sync write, so need to set page writeback
 222          * flag */
 223         if (pg->cp_sync_io == NULL)
 224                 set_page_writeback(vmpage);
 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                      ioret == -EIO) && 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         ENTRY;
 266
 267         LASSERT(PageLocked(vmpage));
 268         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 269
 270         if (vpg->vpg_defer_uptodate)
 271                 ll_ra_count_put(ll_i2sbi(inode), 1);
 272
 273         if (ioret == 0)  {
 274                 if (!vpg->vpg_defer_uptodate)
 275                         cl_page_export(env, page, 1);
 276         } else if (vpg->vpg_defer_uptodate) {
 277                 vpg->vpg_defer_uptodate = 0;
 278                 if (ioret == -EWOULDBLOCK) {
 279                         /* mirror read failed, it needs to destroy the page
 280                          * because subpage would be from wrong osc when trying
 281                          * to read from a new mirror */
 282                         ll_invalidate_page(vmpage);
 283                 }
 284         }
 285
 286         if (page->cp_sync_io == NULL)
 287                 unlock_page(vmpage);
 288
 289         EXIT;
 290 }
 291
 292 static void vvp_page_completion_write(const struct lu_env *env,
 293                                       const struct cl_page_slice *slice,
 294                                       int ioret)
 295 {
 296         struct vvp_page *vpg    = cl2vvp_page(slice);
 297         struct cl_page  *pg     = slice->cpl_page;
 298         struct page     *vmpage = vpg->vpg_page;
 299         ENTRY;
 300
 301         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 302
 303         if (pg->cp_sync_io != NULL) {
 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         EXIT;
 317 }
 318
 319 /**
 320  * Implements cl_page_operations::cpo_make_ready() method.
 321  *
 322  * This is called to yank a page from the transfer cache and to send it out as
 323  * a part of transfer. This function try-locks the page. If try-lock failed,
 324  * page is owned by some concurrent IO, and should be skipped (this is bad,
 325  * but hopefully rare situation, as it usually results in transfer being
 326  * shorter than possible).
 327  *
 328  * \retval 0      success, page can be placed into transfer
 329  *
 330  * \retval -EAGAIN page is either used by concurrent IO has been
 331  * truncated. Skip it.
 332  */
 333 static int vvp_page_make_ready(const struct lu_env *env,
 334                                const struct cl_page_slice *slice)
 335 {
 336         struct page *vmpage = cl2vm_page(slice);
 337         struct cl_page *pg = slice->cpl_page;
 338         int result = 0;
 339
 340         lock_page(vmpage);
 341         if (clear_page_dirty_for_io(vmpage)) {
 342                 LASSERT(pg->cp_state == CPS_CACHED);
 343                 /* This actually clears the dirty bit in the radix
 344                  * tree. */
 345                 set_page_writeback(vmpage);
 346                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 347         } else if (pg->cp_state == CPS_PAGEOUT) {
 348                 /* is it possible for osc_flush_async_page() to already
 349                  * make it ready? */
 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_print(const struct lu_env *env,
 361                           const struct cl_page_slice *slice,
 362                           void *cookie, lu_printer_t printer)
 363 {
 364         struct vvp_page *vpg    = cl2vvp_page(slice);
 365         struct page     *vmpage = vpg->vpg_page;
 366
 367         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d) "
 368                    "vm@%p ",
 369                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used, vmpage);
 370
 371         if (vmpage != NULL) {
 372                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 373                            (long)vmpage->flags, page_count(vmpage),
 374                            page_mapcount(vmpage), vmpage->private,
 375                            page_index(vmpage),
 376                            list_empty(&vmpage->lru) ? "not-" : "");
 377         }
 378
 379         (*printer)(env, cookie, "\n");
 380
 381         return 0;
 382 }
 383
 384 static int vvp_page_fail(const struct lu_env *env,
 385                          const struct cl_page_slice *slice)
 386 {
 387         /*
 388          * Cached read?
 389          */
 390         LBUG();
 391
 392         return 0;
 393 }
 394
 395 static const struct cl_page_operations vvp_page_ops = {
 396         .cpo_own           = vvp_page_own,
 397         .cpo_assume        = vvp_page_assume,
 398         .cpo_unassume      = vvp_page_unassume,
 399         .cpo_disown        = vvp_page_disown,
 400         .cpo_discard       = vvp_page_discard,
 401         .cpo_delete        = vvp_page_delete,
 402         .cpo_export        = vvp_page_export,
 403         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 404         .cpo_fini          = vvp_page_fini,
 405         .cpo_print         = vvp_page_print,
 406         .io = {
 407                 [CRT_READ] = {
 408                         .cpo_prep       = vvp_page_prep_read,
 409                         .cpo_completion = vvp_page_completion_read,
 410                         .cpo_make_ready = vvp_page_fail,
 411                 },
 412                 [CRT_WRITE] = {
 413                         .cpo_prep       = vvp_page_prep_write,
 414                         .cpo_completion = vvp_page_completion_write,
 415                         .cpo_make_ready = vvp_page_make_ready,
 416                 },
 417         },
 418 };
 419
 420 static int vvp_transient_page_prep(const struct lu_env *env,
 421                                    const struct cl_page_slice *slice,
 422                                    struct cl_io *unused)
 423 {
 424         ENTRY;
 425         /* transient page should always be sent. */
 426         RETURN(0);
 427 }
 428
 429 static void vvp_transient_page_verify(const struct cl_page *page)
 430 {
 431 }
 432
 433 static int vvp_transient_page_own(const struct lu_env *env,
 434                                   const struct cl_page_slice *slice,
 435                                   struct cl_io *unused, int nonblock)
 436 {
 437         vvp_transient_page_verify(slice->cpl_page);
 438         return 0;
 439 }
 440
 441 static void vvp_transient_page_assume(const struct lu_env *env,
 442                                       const struct cl_page_slice *slice,
 443                                       struct cl_io *unused)
 444 {
 445         vvp_transient_page_verify(slice->cpl_page);
 446 }
 447
 448 static void vvp_transient_page_unassume(const struct lu_env *env,
 449                                         const struct cl_page_slice *slice,
 450                                         struct cl_io *unused)
 451 {
 452         vvp_transient_page_verify(slice->cpl_page);
 453 }
 454
 455 static void vvp_transient_page_disown(const struct lu_env *env,
 456                                       const struct cl_page_slice *slice,
 457                                       struct cl_io *unused)
 458 {
 459         vvp_transient_page_verify(slice->cpl_page);
 460 }
 461
 462 static void vvp_transient_page_discard(const struct lu_env *env,
 463                                        const struct cl_page_slice *slice,
 464                                        struct cl_io *unused)
 465 {
 466         struct cl_page *page = slice->cpl_page;
 467
 468         vvp_transient_page_verify(slice->cpl_page);
 469
 470         /*
 471          * For transient pages, remove it from the radix tree.
 472          */
 473         cl_page_delete(env, page);
 474 }
 475
 476 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 477                                           const struct cl_page_slice *slice)
 478 {
 479         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 480         int     locked;
 481
 482         locked = !inode_trylock(inode);
 483         if (!locked)
 484                 inode_unlock(inode);
 485         return locked ? -EBUSY : -ENODATA;
 486 }
 487
 488 static void
 489 vvp_transient_page_completion(const struct lu_env *env,
 490                               const struct cl_page_slice *slice,
 491                               int ioret)
 492 {
 493         vvp_transient_page_verify(slice->cpl_page);
 494 }
 495
 496 static void vvp_transient_page_fini(const struct lu_env *env,
 497                                     struct cl_page_slice *slice,
 498                                     struct pagevec *pvec)
 499 {
 500         struct vvp_page *vpg = cl2vvp_page(slice);
 501         struct cl_page *clp = slice->cpl_page;
 502         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 503
 504         vvp_page_fini_common(vpg, pvec);
 505         atomic_dec(&clobj->vob_transient_pages);
 506 }
 507
 508 static const struct cl_page_operations vvp_transient_page_ops = {
 509         .cpo_own                = vvp_transient_page_own,
 510         .cpo_assume             = vvp_transient_page_assume,
 511         .cpo_unassume           = vvp_transient_page_unassume,
 512         .cpo_disown             = vvp_transient_page_disown,
 513         .cpo_discard            = vvp_transient_page_discard,
 514         .cpo_fini               = vvp_transient_page_fini,
 515         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 516         .cpo_print              = vvp_page_print,
 517         .io = {
 518                 [CRT_READ] = {
 519                         .cpo_prep       = vvp_transient_page_prep,
 520                         .cpo_completion = vvp_transient_page_completion,
 521                 },
 522                 [CRT_WRITE] = {
 523                         .cpo_prep       = vvp_transient_page_prep,
 524                         .cpo_completion = vvp_transient_page_completion,
 525                 }
 526         }
 527 };
 528
 529 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 530                 struct cl_page *page, pgoff_t index)
 531 {
 532         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 533         struct page     *vmpage = page->cp_vmpage;
 534
 535         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 536
 537         vpg->vpg_page = vmpage;
 538         get_page(vmpage);
 539
 540         if (page->cp_type == CPT_CACHEABLE) {
 541                 /* in cache, decref in vvp_page_delete */
 542                 atomic_inc(&page->cp_ref);
 543                 SetPagePrivate(vmpage);
 544                 vmpage->private = (unsigned long)page;
 545                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 546                                 &vvp_page_ops);
 547         } else {
 548                 struct vvp_object *clobj = cl2vvp(obj);
 549
 550                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 551                                 &vvp_transient_page_ops);
 552                 atomic_inc(&clobj->vob_transient_pages);
 553         }
 554         return 0;
 555 }