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, 2015, 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         ClearPageUptodate(vmpage);
 171         ClearPagePrivate(vmpage);
 172         vmpage->private = 0;
 173         /*
 174          * Reference from vmpage to cl_page is removed, but the reference back
 175          * is still here. It is removed later in vvp_page_fini().
 176          */
 177 }
 178
 179 static void vvp_page_export(const struct lu_env *env,
 180                             const struct cl_page_slice *slice,
 181                             int uptodate)
 182 {
 183         struct page *vmpage = cl2vm_page(slice);
 184
 185         LASSERT(vmpage != NULL);
 186         LASSERT(PageLocked(vmpage));
 187         if (uptodate)
 188                 SetPageUptodate(vmpage);
 189         else
 190                 ClearPageUptodate(vmpage);
 191 }
 192
 193 static int vvp_page_is_vmlocked(const struct lu_env *env,
 194                                 const struct cl_page_slice *slice)
 195 {
 196         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 197 }
 198
 199 static int vvp_page_prep_read(const struct lu_env *env,
 200                               const struct cl_page_slice *slice,
 201                               struct cl_io *unused)
 202 {
 203         ENTRY;
 204         /* Skip the page already marked as PG_uptodate. */
 205         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 206 }
 207
 208 static int vvp_page_prep_write(const struct lu_env *env,
 209                                const struct cl_page_slice *slice,
 210                                struct cl_io *unused)
 211 {
 212         struct page *vmpage = cl2vm_page(slice);
 213
 214         LASSERT(PageLocked(vmpage));
 215         LASSERT(!PageDirty(vmpage));
 216
 217         set_page_writeback(vmpage);
 218
 219         return 0;
 220 }
 221
 222 /**
 223  * Handles page transfer errors at VM level.
 224  *
 225  * This takes inode as a separate argument, because inode on which error is to
 226  * be set can be different from \a vmpage inode in case of direct-io.
 227  */
 228 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage, int ioret)
 229 {
 230         struct vvp_object *obj = cl_inode2vvp(inode);
 231
 232         if (ioret == 0) {
 233                 ClearPageError(vmpage);
 234                 obj->vob_discard_page_warned = 0;
 235         } else {
 236                 SetPageError(vmpage);
 237                 if (ioret == -ENOSPC)
 238                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 239                 else
 240                         set_bit(AS_EIO, &inode->i_mapping->flags);
 241
 242                 if ((ioret == -ESHUTDOWN || ioret == -EINTR) &&
 243                      obj->vob_discard_page_warned == 0) {
 244                         obj->vob_discard_page_warned = 1;
 245                         ll_dirty_page_discard_warn(vmpage, ioret);
 246                 }
 247         }
 248 }
 249
 250 static void vvp_page_completion_read(const struct lu_env *env,
 251                                      const struct cl_page_slice *slice,
 252                                      int ioret)
 253 {
 254         struct vvp_page *vpg    = cl2vvp_page(slice);
 255         struct page     *vmpage = vpg->vpg_page;
 256         struct cl_page  *page   = slice->cpl_page;
 257         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 258         ENTRY;
 259
 260         LASSERT(PageLocked(vmpage));
 261         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 262
 263         if (vpg->vpg_defer_uptodate)
 264                 ll_ra_count_put(ll_i2sbi(inode), 1);
 265
 266         if (ioret == 0)  {
 267                 if (!vpg->vpg_defer_uptodate)
 268                         cl_page_export(env, page, 1);
 269         } else {
 270                 vpg->vpg_defer_uptodate = 0;
 271         }
 272
 273         if (page->cp_sync_io == NULL)
 274                 unlock_page(vmpage);
 275
 276         EXIT;
 277 }
 278
 279 static void vvp_page_completion_write(const struct lu_env *env,
 280                                       const struct cl_page_slice *slice,
 281                                       int ioret)
 282 {
 283         struct vvp_page *vpg    = cl2vvp_page(slice);
 284         struct cl_page  *pg     = slice->cpl_page;
 285         struct page     *vmpage = vpg->vpg_page;
 286         ENTRY;
 287
 288         LASSERT(ergo(pg->cp_sync_io != NULL, PageLocked(vmpage)));
 289         LASSERT(PageWriteback(vmpage));
 290
 291         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 292
 293         /*
 294          * Only mark the page error only when it's an async write because
 295          * applications won't wait for IO to finish.
 296          */
 297         if (pg->cp_sync_io == NULL)
 298                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 299
 300         end_page_writeback(vmpage);
 301         EXIT;
 302 }
 303
 304 /**
 305  * Implements cl_page_operations::cpo_make_ready() method.
 306  *
 307  * This is called to yank a page from the transfer cache and to send it out as
 308  * a part of transfer. This function try-locks the page. If try-lock failed,
 309  * page is owned by some concurrent IO, and should be skipped (this is bad,
 310  * but hopefully rare situation, as it usually results in transfer being
 311  * shorter than possible).
 312  *
 313  * \retval 0      success, page can be placed into transfer
 314  *
 315  * \retval -EAGAIN page is either used by concurrent IO has been
 316  * truncated. Skip it.
 317  */
 318 static int vvp_page_make_ready(const struct lu_env *env,
 319                                const struct cl_page_slice *slice)
 320 {
 321         struct page *vmpage = cl2vm_page(slice);
 322         struct cl_page *pg = slice->cpl_page;
 323         int result = 0;
 324
 325         lock_page(vmpage);
 326         if (clear_page_dirty_for_io(vmpage)) {
 327                 LASSERT(pg->cp_state == CPS_CACHED);
 328                 /* This actually clears the dirty bit in the radix
 329                  * tree. */
 330                 set_page_writeback(vmpage);
 331                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 332         } else if (pg->cp_state == CPS_PAGEOUT) {
 333                 /* is it possible for osc_flush_async_page() to already
 334                  * make it ready? */
 335                 result = -EALREADY;
 336         } else {
 337                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 338                               pg->cp_state);
 339                 LBUG();
 340         }
 341         unlock_page(vmpage);
 342         RETURN(result);
 343 }
 344
 345 static int vvp_page_print(const struct lu_env *env,
 346                           const struct cl_page_slice *slice,
 347                           void *cookie, lu_printer_t printer)
 348 {
 349         struct vvp_page *vpg    = cl2vvp_page(slice);
 350         struct page     *vmpage = vpg->vpg_page;
 351
 352         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d) "
 353                    "vm@%p ",
 354                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used, vmpage);
 355
 356         if (vmpage != NULL) {
 357                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 358                            (long)vmpage->flags, page_count(vmpage),
 359                            page_mapcount(vmpage), vmpage->private,
 360                            page_index(vmpage),
 361                            list_empty(&vmpage->lru) ? "not-" : "");
 362         }
 363
 364         (*printer)(env, cookie, "\n");
 365
 366         return 0;
 367 }
 368
 369 static int vvp_page_fail(const struct lu_env *env,
 370                          const struct cl_page_slice *slice)
 371 {
 372         /*
 373          * Cached read?
 374          */
 375         LBUG();
 376
 377         return 0;
 378 }
 379
 380 static const struct cl_page_operations vvp_page_ops = {
 381         .cpo_own           = vvp_page_own,
 382         .cpo_assume        = vvp_page_assume,
 383         .cpo_unassume      = vvp_page_unassume,
 384         .cpo_disown        = vvp_page_disown,
 385         .cpo_discard       = vvp_page_discard,
 386         .cpo_delete        = vvp_page_delete,
 387         .cpo_export        = vvp_page_export,
 388         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 389         .cpo_fini          = vvp_page_fini,
 390         .cpo_print         = vvp_page_print,
 391         .io = {
 392                 [CRT_READ] = {
 393                         .cpo_prep       = vvp_page_prep_read,
 394                         .cpo_completion = vvp_page_completion_read,
 395                         .cpo_make_ready = vvp_page_fail,
 396                 },
 397                 [CRT_WRITE] = {
 398                         .cpo_prep       = vvp_page_prep_write,
 399                         .cpo_completion = vvp_page_completion_write,
 400                         .cpo_make_ready = vvp_page_make_ready,
 401                 },
 402         },
 403 };
 404
 405 static int vvp_transient_page_prep(const struct lu_env *env,
 406                                    const struct cl_page_slice *slice,
 407                                    struct cl_io *unused)
 408 {
 409         ENTRY;
 410         /* transient page should always be sent. */
 411         RETURN(0);
 412 }
 413
 414 static void vvp_transient_page_verify(const struct cl_page *page)
 415 {
 416 }
 417
 418 static int vvp_transient_page_own(const struct lu_env *env,
 419                                   const struct cl_page_slice *slice,
 420                                   struct cl_io *unused, int nonblock)
 421 {
 422         vvp_transient_page_verify(slice->cpl_page);
 423         return 0;
 424 }
 425
 426 static void vvp_transient_page_assume(const struct lu_env *env,
 427                                       const struct cl_page_slice *slice,
 428                                       struct cl_io *unused)
 429 {
 430         vvp_transient_page_verify(slice->cpl_page);
 431 }
 432
 433 static void vvp_transient_page_unassume(const struct lu_env *env,
 434                                         const struct cl_page_slice *slice,
 435                                         struct cl_io *unused)
 436 {
 437         vvp_transient_page_verify(slice->cpl_page);
 438 }
 439
 440 static void vvp_transient_page_disown(const struct lu_env *env,
 441                                       const struct cl_page_slice *slice,
 442                                       struct cl_io *unused)
 443 {
 444         vvp_transient_page_verify(slice->cpl_page);
 445 }
 446
 447 static void vvp_transient_page_discard(const struct lu_env *env,
 448                                        const struct cl_page_slice *slice,
 449                                        struct cl_io *unused)
 450 {
 451         struct cl_page *page = slice->cpl_page;
 452
 453         vvp_transient_page_verify(slice->cpl_page);
 454
 455         /*
 456          * For transient pages, remove it from the radix tree.
 457          */
 458         cl_page_delete(env, page);
 459 }
 460
 461 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 462                                           const struct cl_page_slice *slice)
 463 {
 464         struct inode    *inode = vvp_object_inode(slice->cpl_obj);
 465         int     locked;
 466
 467         locked = !inode_trylock(inode);
 468         if (!locked)
 469                 inode_unlock(inode);
 470         return locked ? -EBUSY : -ENODATA;
 471 }
 472
 473 static void
 474 vvp_transient_page_completion(const struct lu_env *env,
 475                               const struct cl_page_slice *slice,
 476                               int ioret)
 477 {
 478         vvp_transient_page_verify(slice->cpl_page);
 479 }
 480
 481 static void vvp_transient_page_fini(const struct lu_env *env,
 482                                     struct cl_page_slice *slice)
 483 {
 484         struct vvp_page *vpg = cl2vvp_page(slice);
 485         struct cl_page *clp = slice->cpl_page;
 486         struct vvp_object *clobj = cl2vvp(clp->cp_obj);
 487
 488         vvp_page_fini_common(vpg);
 489         atomic_dec(&clobj->vob_transient_pages);
 490 }
 491
 492 static const struct cl_page_operations vvp_transient_page_ops = {
 493         .cpo_own                = vvp_transient_page_own,
 494         .cpo_assume             = vvp_transient_page_assume,
 495         .cpo_unassume           = vvp_transient_page_unassume,
 496         .cpo_disown             = vvp_transient_page_disown,
 497         .cpo_discard            = vvp_transient_page_discard,
 498         .cpo_fini               = vvp_transient_page_fini,
 499         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 500         .cpo_print              = vvp_page_print,
 501         .io = {
 502                 [CRT_READ] = {
 503                         .cpo_prep       = vvp_transient_page_prep,
 504                         .cpo_completion = vvp_transient_page_completion,
 505                 },
 506                 [CRT_WRITE] = {
 507                         .cpo_prep       = vvp_transient_page_prep,
 508                         .cpo_completion = vvp_transient_page_completion,
 509                 }
 510         }
 511 };
 512
 513 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 514                 struct cl_page *page, pgoff_t index)
 515 {
 516         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 517         struct page     *vmpage = page->cp_vmpage;
 518
 519         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 520
 521         vpg->vpg_page = vmpage;
 522         get_page(vmpage);
 523
 524         if (page->cp_type == CPT_CACHEABLE) {
 525                 /* in cache, decref in vvp_page_delete */
 526                 atomic_inc(&page->cp_ref);
 527                 SetPagePrivate(vmpage);
 528                 vmpage->private = (unsigned long)page;
 529                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 530                                 &vvp_page_ops);
 531         } else {
 532                 struct vvp_object *clobj = cl2vvp(obj);
 533
 534                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 535                                 &vvp_transient_page_ops);
 536                 atomic_inc(&clobj->vob_transient_pages);
 537         }
 538         return 0;
 539 }