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