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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_page for VVP layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_LLITE
  43
  44 #include <linux/atomic.h>
  45 #include <linux/bitops.h>
  46 #include <linux/mm.h>
  47 #include <linux/mutex.h>
  48 #include <linux/page-flags.h>
  49 #include <linux/pagemap.h>
  50
  51 #include "../include/lustre_lite.h"
  52
  53 #include "llite_internal.h"
  54 #include "vvp_internal.h"
  55
  56 /*****************************************************************************
  57  *
  58  * Page operations.
  59  *
  60  */
  61
  62 static void vvp_page_fini_common(struct vvp_page *vpg)
  63 {
  64         struct page *vmpage = vpg->vpg_page;
  65
  66         LASSERT(vmpage);
  67         put_page(vmpage);
  68 }
  69
  70 static void vvp_page_fini(const struct lu_env *env,
  71                           struct cl_page_slice *slice)
  72 {
  73         struct vvp_page *vpg     = cl2vvp_page(slice);
  74         struct page     *vmpage  = vpg->vpg_page;
  75
  76         /*
  77          * vmpage->private was already cleared when page was moved into
  78          * VPG_FREEING state.
  79          */
  80         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  81         vvp_page_fini_common(vpg);
  82 }
  83
  84 static int vvp_page_own(const struct lu_env *env,
  85                         const struct cl_page_slice *slice, struct cl_io *io,
  86                         int nonblock)
  87 {
  88         struct vvp_page *vpg    = cl2vvp_page(slice);
  89         struct page     *vmpage = vpg->vpg_page;
  90
  91         LASSERT(vmpage);
  92         if (nonblock) {
  93                 if (!trylock_page(vmpage))
  94                         return -EAGAIN;
  95
  96                 if (unlikely(PageWriteback(vmpage))) {
  97                         unlock_page(vmpage);
  98                         return -EAGAIN;
  99                 }
 100
 101                 return 0;
 102         }
 103
 104         lock_page(vmpage);
 105         wait_on_page_writeback(vmpage);
 106
 107         return 0;
 108 }
 109
 110 static void vvp_page_assume(const struct lu_env *env,
 111                             const struct cl_page_slice *slice,
 112                             struct cl_io *unused)
 113 {
 114         struct page *vmpage = cl2vm_page(slice);
 115
 116         LASSERT(vmpage);
 117         LASSERT(PageLocked(vmpage));
 118         wait_on_page_writeback(vmpage);
 119 }
 120
 121 static void vvp_page_unassume(const struct lu_env *env,
 122                               const struct cl_page_slice *slice,
 123                               struct cl_io *unused)
 124 {
 125         struct page *vmpage = cl2vm_page(slice);
 126
 127         LASSERT(vmpage);
 128         LASSERT(PageLocked(vmpage));
 129 }
 130
 131 static void vvp_page_disown(const struct lu_env *env,
 132                             const struct cl_page_slice *slice, struct cl_io *io)
 133 {
 134         struct page *vmpage = cl2vm_page(slice);
 135
 136         LASSERT(vmpage);
 137         LASSERT(PageLocked(vmpage));
 138
 139         unlock_page(cl2vm_page(slice));
 140 }
 141
 142 static void vvp_page_discard(const struct lu_env *env,
 143                              const struct cl_page_slice *slice,
 144                              struct cl_io *unused)
 145 {
 146         struct page        *vmpage  = cl2vm_page(slice);
 147         struct vvp_page      *vpg     = cl2vvp_page(slice);
 148
 149         LASSERT(vmpage);
 150         LASSERT(PageLocked(vmpage));
 151
 152         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used)
 153                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 154
 155         ll_invalidate_page(vmpage);
 156 }
 157
 158 static void vvp_page_delete(const struct lu_env *env,
 159                             const struct cl_page_slice *slice)
 160 {
 161         struct page       *vmpage = cl2vm_page(slice);
 162         struct inode     *inode  = vmpage->mapping->host;
 163         struct cl_object *obj    = slice->cpl_obj;
 164         struct cl_page   *page   = slice->cpl_page;
 165         int refc;
 166
 167         LASSERT(PageLocked(vmpage));
 168         LASSERT((struct cl_page *)vmpage->private == page);
 169         LASSERT(inode == vvp_object_inode(obj));
 170
 171         vvp_write_complete(cl2vvp(obj), cl2vvp_page(slice));
 172
 173         /* Drop the reference count held in vvp_page_init */
 174         refc = atomic_dec_return(&page->cp_ref);
 175         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 176
 177         ClearPageUptodate(vmpage);
 178         ClearPagePrivate(vmpage);
 179         vmpage->private = 0;
 180         /*
 181          * Reference from vmpage to cl_page is removed, but the reference back
 182          * is still here. It is removed later in vvp_page_fini().
 183          */
 184 }
 185
 186 static void vvp_page_export(const struct lu_env *env,
 187                             const struct cl_page_slice *slice,
 188                             int uptodate)
 189 {
 190         struct page *vmpage = cl2vm_page(slice);
 191
 192         LASSERT(vmpage);
 193         LASSERT(PageLocked(vmpage));
 194         if (uptodate)
 195                 SetPageUptodate(vmpage);
 196         else
 197                 ClearPageUptodate(vmpage);
 198 }
 199
 200 static int vvp_page_is_vmlocked(const struct lu_env *env,
 201                                 const struct cl_page_slice *slice)
 202 {
 203         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 204 }
 205
 206 static int vvp_page_prep_read(const struct lu_env *env,
 207                               const struct cl_page_slice *slice,
 208                               struct cl_io *unused)
 209 {
 210         /* Skip the page already marked as PG_uptodate. */
 211         return PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0;
 212 }
 213
 214 static int vvp_page_prep_write(const struct lu_env *env,
 215                                const struct cl_page_slice *slice,
 216                                struct cl_io *unused)
 217 {
 218         struct page *vmpage = cl2vm_page(slice);
 219         struct cl_page *pg = slice->cpl_page;
 220
 221         LASSERT(PageLocked(vmpage));
 222         LASSERT(!PageDirty(vmpage));
 223
 224         /* ll_writepage path is not a sync write, so need to set page writeback
 225          * flag
 226          */
 227         if (!pg->cp_sync_io)
 228                 set_page_writeback(vmpage);
 229
 230         vvp_write_pending(cl2vvp(slice->cpl_obj), cl2vvp_page(slice));
 231
 232         return 0;
 233 }
 234
 235 /**
 236  * Handles page transfer errors at VM level.
 237  *
 238  * This takes inode as a separate argument, because inode on which error is to
 239  * be set can be different from \a vmpage inode in case of direct-io.
 240  */
 241 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 242 {
 243         struct vvp_object *obj = cl_inode2vvp(inode);
 244
 245         if (ioret == 0) {
 246                 ClearPageError(vmpage);
 247                 obj->vob_discard_page_warned = 0;
 248         } else {
 249                 SetPageError(vmpage);
 250                 if (ioret == -ENOSPC)
 251                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 252                 else
 253                         set_bit(AS_EIO, &inode->i_mapping->flags);
 254
 255                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 256                      obj->vob_discard_page_warned == 0) {
 257                         obj->vob_discard_page_warned = 1;
 258                         ll_dirty_page_discard_warn(vmpage, ioret);
 259                 }
 260         }
 261 }
 262
 263 static void vvp_page_completion_read(const struct lu_env *env,
 264                                      const struct cl_page_slice *slice,
 265                                      int ioret)
 266 {
 267         struct vvp_page *vpg    = cl2vvp_page(slice);
 268         struct page     *vmpage = vpg->vpg_page;
 269         struct cl_page  *page   = slice->cpl_page;
 270         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 271
 272         LASSERT(PageLocked(vmpage));
 273         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 274
 275         if (vpg->vpg_defer_uptodate)
 276                 ll_ra_count_put(ll_i2sbi(inode), 1);
 277
 278         if (ioret == 0)  {
 279                 if (!vpg->vpg_defer_uptodate)
 280                         cl_page_export(env, page, 1);
 281         } else {
 282                 vpg->vpg_defer_uptodate = 0;
 283         }
 284
 285         if (!page->cp_sync_io)
 286                 unlock_page(vmpage);
 287 }
 288
 289 static void vvp_page_completion_write(const struct lu_env *env,
 290                                       const struct cl_page_slice *slice,
 291                                       int ioret)
 292 {
 293         struct vvp_page *vpg     = cl2vvp_page(slice);
 294         struct cl_page  *pg     = slice->cpl_page;
 295         struct page     *vmpage = vpg->vpg_page;
 296
 297         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 298
 299         /*
 300          * TODO: Actually it makes sense to add the page into oap pending
 301          * list again and so that we don't need to take the page out from
 302          * SoM write pending list, if we just meet a recoverable error,
 303          * -ENOMEM, etc.
 304          * To implement this, we just need to return a non zero value in
 305          * ->cpo_completion method. The underlying transfer should be notified
 306          * and then re-add the page into pending transfer queue.  -jay
 307          */
 308
 309         vpg->vpg_write_queued = 0;
 310         vvp_write_complete(cl2vvp(slice->cpl_obj), vpg);
 311
 312         if (pg->cp_sync_io) {
 313                 LASSERT(PageLocked(vmpage));
 314                 LASSERT(!PageWriteback(vmpage));
 315         } else {
 316                 LASSERT(PageWriteback(vmpage));
 317                 /*
 318                  * Only mark the page error only when it's an async write
 319                  * because applications won't wait for IO to finish.
 320                  */
 321                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 322
 323                 end_page_writeback(vmpage);
 324         }
 325 }
 326
 327 /**
 328  * Implements cl_page_operations::cpo_make_ready() method.
 329  *
 330  * This is called to yank a page from the transfer cache and to send it out as
 331  * a part of transfer. This function try-locks the page. If try-lock failed,
 332  * page is owned by some concurrent IO, and should be skipped (this is bad,
 333  * but hopefully rare situation, as it usually results in transfer being
 334  * shorter than possible).
 335  *
 336  * \retval 0      success, page can be placed into transfer
 337  *
 338  * \retval -EAGAIN page is either used by concurrent IO has been
 339  * truncated. Skip it.
 340  */
 341 static int vvp_page_make_ready(const struct lu_env *env,
 342                                const struct cl_page_slice *slice)
 343 {
 344         struct page *vmpage = cl2vm_page(slice);
 345         struct cl_page *pg = slice->cpl_page;
 346         int result = 0;
 347
 348         lock_page(vmpage);
 349         if (clear_page_dirty_for_io(vmpage)) {
 350                 LASSERT(pg->cp_state == CPS_CACHED);
 351                 /* This actually clears the dirty bit in the radix tree. */
 352                 set_page_writeback(vmpage);
 353                 vvp_write_pending(cl2vvp(slice->cpl_obj), cl2vvp_page(slice));
 354                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 355         } else if (pg->cp_state == CPS_PAGEOUT) {
 356                 /* is it possible for osc_flush_async_page() to already
 357                  * make it ready?
 358                  */
 359                 result = -EALREADY;
 360         } else {
 361                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 362                               pg->cp_state);
 363                 LBUG();
 364         }
 365         unlock_page(vmpage);
 366         return result;
 367 }
 368
 369 static int vvp_page_is_under_lock(const struct lu_env *env,
 370                                   const struct cl_page_slice *slice,
 371                                   struct cl_io *io, pgoff_t *max_index)
 372 {
 373         if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
 374             io->ci_type == CIT_FAULT) {
 375                 struct vvp_io *vio = vvp_env_io(env);
 376
 377                 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED))
 378                         *max_index = CL_PAGE_EOF;
 379         }
 380         return 0;
 381 }
 382
 383 static int vvp_page_print(const struct lu_env *env,
 384                           const struct cl_page_slice *slice,
 385                           void *cookie, lu_printer_t printer)
 386 {
 387         struct vvp_page *vpg = cl2vvp_page(slice);
 388         struct page     *vmpage = vpg->vpg_page;
 389
 390         (*printer)(env, cookie, LUSTRE_VVP_NAME "-page@%p(%d:%d:%d) vm@%p ",
 391                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used,
 392                    vpg->vpg_write_queued, vmpage);
 393         if (vmpage) {
 394                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 395                            (long)vmpage->flags, page_count(vmpage),
 396                            page_mapcount(vmpage), vmpage->private,
 397                            vmpage->index,
 398                            list_empty(&vmpage->lru) ? "not-" : "");
 399         }
 400
 401         (*printer)(env, cookie, "\n");
 402
 403         return 0;
 404 }
 405
 406 static int vvp_page_fail(const struct lu_env *env,
 407                          const struct cl_page_slice *slice)
 408 {
 409         /*
 410          * Cached read?
 411          */
 412         LBUG();
 413
 414         return 0;
 415 }
 416
 417 static const struct cl_page_operations vvp_page_ops = {
 418         .cpo_own           = vvp_page_own,
 419         .cpo_assume     = vvp_page_assume,
 420         .cpo_unassume      = vvp_page_unassume,
 421         .cpo_disown     = vvp_page_disown,
 422         .cpo_discard       = vvp_page_discard,
 423         .cpo_delete     = vvp_page_delete,
 424         .cpo_export     = vvp_page_export,
 425         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 426         .cpo_fini         = vvp_page_fini,
 427         .cpo_print       = vvp_page_print,
 428         .cpo_is_under_lock = vvp_page_is_under_lock,
 429         .io = {
 430                 [CRT_READ] = {
 431                         .cpo_prep       = vvp_page_prep_read,
 432                         .cpo_completion  = vvp_page_completion_read,
 433                         .cpo_make_ready = vvp_page_fail,
 434                 },
 435                 [CRT_WRITE] = {
 436                         .cpo_prep       = vvp_page_prep_write,
 437                         .cpo_completion  = vvp_page_completion_write,
 438                         .cpo_make_ready  = vvp_page_make_ready,
 439                 },
 440         },
 441 };
 442
 443 static int vvp_transient_page_prep(const struct lu_env *env,
 444                                    const struct cl_page_slice *slice,
 445                                    struct cl_io *unused)
 446 {
 447         /* transient page should always be sent. */
 448         return 0;
 449 }
 450
 451 static void vvp_transient_page_verify(const struct cl_page *page)
 452 {
 453         struct inode *inode = vvp_object_inode(page->cp_obj);
 454
 455         LASSERT(!inode_trylock(inode));
 456 }
 457
 458 static int vvp_transient_page_own(const struct lu_env *env,
 459                                   const struct cl_page_slice *slice,
 460                                   struct cl_io *unused, int nonblock)
 461 {
 462         vvp_transient_page_verify(slice->cpl_page);
 463         return 0;
 464 }
 465
 466 static void vvp_transient_page_assume(const struct lu_env *env,
 467                                       const struct cl_page_slice *slice,
 468                                       struct cl_io *unused)
 469 {
 470         vvp_transient_page_verify(slice->cpl_page);
 471 }
 472
 473 static void vvp_transient_page_unassume(const struct lu_env *env,
 474                                         const struct cl_page_slice *slice,
 475                                         struct cl_io *unused)
 476 {
 477         vvp_transient_page_verify(slice->cpl_page);
 478 }
 479
 480 static void vvp_transient_page_disown(const struct lu_env *env,
 481                                       const struct cl_page_slice *slice,
 482                                       struct cl_io *unused)
 483 {
 484         vvp_transient_page_verify(slice->cpl_page);
 485 }
 486
 487 static void vvp_transient_page_discard(const struct lu_env *env,
 488                                        const struct cl_page_slice *slice,
 489                                        struct cl_io *unused)
 490 {
 491         struct cl_page *page = slice->cpl_page;
 492
 493         vvp_transient_page_verify(slice->cpl_page);
 494
 495         /*
 496          * For transient pages, remove it from the radix tree.
 497          */
 498         cl_page_delete(env, page);
 499 }
 500
 501 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 502                                           const struct cl_page_slice *slice)
 503 {
 504         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 505         int     locked;
 506
 507         locked = !inode_trylock(inode);
 508         if (!locked)
 509                 inode_unlock(inode);
 510         return locked ? -EBUSY : -ENODATA;
 511 }
 512
 513 static void
 514 vvp_transient_page_completion(const struct lu_env *env,
 515                               const struct cl_page_slice *slice,
 516                               int ioret)
 517 {
 518         vvp_transient_page_verify(slice->cpl_page);
 519 }
 520
 521 static void vvp_transient_page_fini(const struct lu_env *env,
 522                                     struct cl_page_slice *slice)
 523 {
 524         struct vvp_page *vpg = cl2vvp_page(slice);
 525         struct cl_page *clp = slice->cpl_page;
 526         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 527
 528         vvp_page_fini_common(vpg);
 529         LASSERT(!inode_trylock(clobj->vob_inode));
 530         clobj->vob_transient_pages--;
 531 }
 532
 533 static const struct cl_page_operations vvp_transient_page_ops = {
 534         .cpo_own           = vvp_transient_page_own,
 535         .cpo_assume     = vvp_transient_page_assume,
 536         .cpo_unassume      = vvp_transient_page_unassume,
 537         .cpo_disown     = vvp_transient_page_disown,
 538         .cpo_discard       = vvp_transient_page_discard,
 539         .cpo_fini         = vvp_transient_page_fini,
 540         .cpo_is_vmlocked   = vvp_transient_page_is_vmlocked,
 541         .cpo_print       = vvp_page_print,
 542         .cpo_is_under_lock      = vvp_page_is_under_lock,
 543         .io = {
 544                 [CRT_READ] = {
 545                         .cpo_prep       = vvp_transient_page_prep,
 546                         .cpo_completion  = vvp_transient_page_completion,
 547                 },
 548                 [CRT_WRITE] = {
 549                         .cpo_prep       = vvp_transient_page_prep,
 550                         .cpo_completion  = vvp_transient_page_completion,
 551                 }
 552         }
 553 };
 554
 555 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 556                 struct cl_page *page, pgoff_t index)
 557 {
 558         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 559         struct page     *vmpage = page->cp_vmpage;
 560
 561         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 562
 563         vpg->vpg_page = vmpage;
 564         get_page(vmpage);
 565
 566         INIT_LIST_HEAD(&vpg->vpg_pending_linkage);
 567         if (page->cp_type == CPT_CACHEABLE) {
 568                 /* in cache, decref in vvp_page_delete */
 569                 atomic_inc(&page->cp_ref);
 570                 SetPagePrivate(vmpage);
 571                 vmpage->private = (unsigned long)page;
 572                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 573                                   &vvp_page_ops);
 574         } else {
 575                 struct vvp_object *clobj = cl2vvp(obj);
 576
 577                 LASSERT(!inode_trylock(clobj->vob_inode));
 578                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 579                                   &vvp_transient_page_ops);
 580                 clobj->vob_transient_pages++;
 581         }
 582         return 0;
 583 }