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         ENTRY;
  93
  94         LASSERT(vmpage != NULL);
  95         if (nonblock) {
  96                 if (!trylock_page(vmpage))
  97                         return -EAGAIN;
  98
  99                 if (unlikely(PageWriteback(vmpage))) {
 100                         unlock_page(vmpage);
 101                         return -EAGAIN;
 102                 }
 103
 104                 return 0;
 105         }
 106
 107         lock_page(vmpage);
 108         wait_on_page_writeback(vmpage);
 109
 110         RETURN(0);
 111 }
 112
 113 static void vvp_page_assume(const struct lu_env *env,
 114                             const struct cl_page_slice *slice,
 115                             struct cl_io *unused)
 116 {
 117         struct page *vmpage = cl2vm_page(slice);
 118
 119         LASSERT(vmpage != NULL);
 120         LASSERT(PageLocked(vmpage));
 121         wait_on_page_writeback(vmpage);
 122 }
 123
 124 static void vvp_page_unassume(const struct lu_env *env,
 125                               const struct cl_page_slice *slice,
 126                               struct cl_io *unused)
 127 {
 128         struct page *vmpage = cl2vm_page(slice);
 129
 130         LASSERT(vmpage != NULL);
 131         LASSERT(PageLocked(vmpage));
 132 }
 133
 134 static void vvp_page_disown(const struct lu_env *env,
 135                             const struct cl_page_slice *slice, struct cl_io *io)
 136 {
 137         struct page *vmpage = cl2vm_page(slice);
 138
 139         ENTRY;
 140
 141         LASSERT(vmpage != NULL);
 142         LASSERT(PageLocked(vmpage));
 143
 144         unlock_page(cl2vm_page(slice));
 145
 146         EXIT;
 147 }
 148
 149 static void vvp_page_discard(const struct lu_env *env,
 150                              const struct cl_page_slice *slice,
 151                              struct cl_io *unused)
 152 {
 153         struct page     *vmpage = cl2vm_page(slice);
 154         struct vvp_page *vpg    = cl2vvp_page(slice);
 155
 156         LASSERT(vmpage != NULL);
 157         LASSERT(PageLocked(vmpage));
 158
 159         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used && vmpage->mapping)
 160                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 161
 162         generic_error_remove_page(vmpage->mapping, vmpage);
 163 }
 164
 165 static void vvp_page_delete(const struct lu_env *env,
 166                             const struct cl_page_slice *slice)
 167 {
 168         struct page      *vmpage = cl2vm_page(slice);
 169         struct cl_page   *page   = slice->cpl_page;
 170         int refc;
 171
 172         LASSERT(PageLocked(vmpage));
 173         LASSERT((struct cl_page *)vmpage->private == page);
 174
 175
 176         /* Drop the reference count held in vvp_page_init */
 177         refc = atomic_dec_return(&page->cp_ref);
 178         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 179
 180         ClearPagePrivate(vmpage);
 181         vmpage->private = 0;
 182         /*
 183          * Reference from vmpage to cl_page is removed, but the reference back
 184          * is still here. It is removed later in vvp_page_fini().
 185          */
 186 }
 187
 188 static void vvp_page_export(const struct lu_env *env,
 189                             const struct cl_page_slice *slice,
 190                             int uptodate)
 191 {
 192         struct page *vmpage = cl2vm_page(slice);
 193
 194         LASSERT(vmpage != NULL);
 195         LASSERT(PageLocked(vmpage));
 196         if (uptodate)
 197                 SetPageUptodate(vmpage);
 198         else
 199                 ClearPageUptodate(vmpage);
 200 }
 201
 202 static int vvp_page_is_vmlocked(const struct lu_env *env,
 203                                 const struct cl_page_slice *slice)
 204 {
 205         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 206 }
 207
 208 static int vvp_page_prep_read(const struct lu_env *env,
 209                               const struct cl_page_slice *slice,
 210                               struct cl_io *unused)
 211 {
 212         ENTRY;
 213         /* Skip the page already marked as PG_uptodate. */
 214         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 215 }
 216
 217 static int vvp_page_prep_write(const struct lu_env *env,
 218                                const struct cl_page_slice *slice,
 219                                struct cl_io *unused)
 220 {
 221         struct page *vmpage = cl2vm_page(slice);
 222         struct cl_page *pg = slice->cpl_page;
 223
 224         LASSERT(PageLocked(vmpage));
 225         LASSERT(!PageDirty(vmpage));
 226
 227         /* ll_writepage path is not a sync write, so need to set page writeback
 228          * flag */
 229         if (pg->cp_sync_io == NULL)
 230                 set_page_writeback(vmpage);
 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                      ioret == -EIO) && 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         ENTRY;
 272
 273         LASSERT(PageLocked(vmpage));
 274         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 275
 276         if (vpg->vpg_defer_uptodate)
 277                 ll_ra_count_put(ll_i2sbi(inode), 1);
 278
 279         if (ioret == 0)  {
 280                 if (!vpg->vpg_defer_uptodate)
 281                         cl_page_export(env, page, 1);
 282         } else if (vpg->vpg_defer_uptodate) {
 283                 vpg->vpg_defer_uptodate = 0;
 284                 if (ioret == -EWOULDBLOCK) {
 285                         /* mirror read failed, it needs to destroy the page
 286                          * because subpage would be from wrong osc when trying
 287                          * to read from a new mirror */
 288                         generic_error_remove_page(vmpage->mapping, vmpage);
 289                 }
 290         }
 291
 292         if (page->cp_sync_io == NULL)
 293                 unlock_page(vmpage);
 294
 295         EXIT;
 296 }
 297
 298 static void vvp_page_completion_write(const struct lu_env *env,
 299                                       const struct cl_page_slice *slice,
 300                                       int ioret)
 301 {
 302         struct vvp_page *vpg    = cl2vvp_page(slice);
 303         struct cl_page  *pg     = slice->cpl_page;
 304         struct page     *vmpage = vpg->vpg_page;
 305         ENTRY;
 306
 307         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 308
 309         if (pg->cp_sync_io != NULL) {
 310                 LASSERT(PageLocked(vmpage));
 311                 LASSERT(!PageWriteback(vmpage));
 312         } else {
 313                 LASSERT(PageWriteback(vmpage));
 314                 /*
 315                  * Only mark the page error only when it's an async write
 316                  * because applications won't wait for IO to finish.
 317                  */
 318                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 319
 320                 end_page_writeback(vmpage);
 321         }
 322         EXIT;
 323 }
 324
 325 /**
 326  * Implements cl_page_operations::cpo_make_ready() method.
 327  *
 328  * This is called to yank a page from the transfer cache and to send it out as
 329  * a part of transfer. This function try-locks the page. If try-lock failed,
 330  * page is owned by some concurrent IO, and should be skipped (this is bad,
 331  * but hopefully rare situation, as it usually results in transfer being
 332  * shorter than possible).
 333  *
 334  * \retval 0      success, page can be placed into transfer
 335  *
 336  * \retval -EAGAIN page is either used by concurrent IO has been
 337  * truncated. Skip it.
 338  */
 339 static int vvp_page_make_ready(const struct lu_env *env,
 340                                const struct cl_page_slice *slice)
 341 {
 342         struct page *vmpage = cl2vm_page(slice);
 343         struct cl_page *pg = slice->cpl_page;
 344         int result = 0;
 345
 346         lock_page(vmpage);
 347         if (clear_page_dirty_for_io(vmpage)) {
 348                 LASSERT(pg->cp_state == CPS_CACHED);
 349                 /* This actually clears the dirty bit in the radix
 350                  * tree. */
 351                 set_page_writeback(vmpage);
 352                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 353         } else if (pg->cp_state == CPS_PAGEOUT) {
 354                 /* is it possible for osc_flush_async_page() to already
 355                  * make it ready? */
 356                 result = -EALREADY;
 357         } else {
 358                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 359                               pg->cp_state);
 360                 LBUG();
 361         }
 362         unlock_page(vmpage);
 363         RETURN(result);
 364 }
 365
 366 static int vvp_page_print(const struct lu_env *env,
 367                           const struct cl_page_slice *slice,
 368                           void *cookie, lu_printer_t printer)
 369 {
 370         struct vvp_page *vpg    = cl2vvp_page(slice);
 371         struct page     *vmpage = vpg->vpg_page;
 372
 373         (*printer)(env, cookie, LUSTRE_VVP_NAME"-page@%p(%d:%d) "
 374                    "vm@%p ",
 375                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used, vmpage);
 376
 377         if (vmpage != NULL) {
 378                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 379                            (long)vmpage->flags, page_count(vmpage),
 380                            page_mapcount(vmpage), vmpage->private,
 381                            page_index(vmpage),
 382                            list_empty(&vmpage->lru) ? "not-" : "");
 383         }
 384
 385         (*printer)(env, cookie, "\n");
 386
 387         return 0;
 388 }
 389
 390 static int vvp_page_fail(const struct lu_env *env,
 391                          const struct cl_page_slice *slice)
 392 {
 393         /*
 394          * Cached read?
 395          */
 396         LBUG();
 397
 398         return 0;
 399 }
 400
 401 static const struct cl_page_operations vvp_page_ops = {
 402         .cpo_own           = vvp_page_own,
 403         .cpo_assume        = vvp_page_assume,
 404         .cpo_unassume      = vvp_page_unassume,
 405         .cpo_disown        = vvp_page_disown,
 406         .cpo_discard       = vvp_page_discard,
 407         .cpo_delete        = vvp_page_delete,
 408         .cpo_export        = vvp_page_export,
 409         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 410         .cpo_fini          = vvp_page_fini,
 411         .cpo_print         = vvp_page_print,
 412         .io = {
 413                 [CRT_READ] = {
 414                         .cpo_prep       = vvp_page_prep_read,
 415                         .cpo_completion = vvp_page_completion_read,
 416                         .cpo_make_ready = vvp_page_fail,
 417                 },
 418                 [CRT_WRITE] = {
 419                         .cpo_prep       = vvp_page_prep_write,
 420                         .cpo_completion = vvp_page_completion_write,
 421                         .cpo_make_ready = vvp_page_make_ready,
 422                 },
 423         },
 424 };
 425
 426 static void vvp_transient_page_discard(const struct lu_env *env,
 427                                        const struct cl_page_slice *slice,
 428                                        struct cl_io *unused)
 429 {
 430         struct cl_page *page = slice->cpl_page;
 431
 432         /*
 433          * For transient pages, remove it from the radix tree.
 434          */
 435         cl_page_delete(env, page);
 436 }
 437
 438 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 439                                           const struct cl_page_slice *slice)
 440 {
 441         return -EBUSY;
 442 }
 443
 444 static void vvp_transient_page_fini(const struct lu_env *env,
 445                                     struct cl_page_slice *slice,
 446                                     struct pagevec *pvec)
 447 {
 448         struct vvp_page *vpg = cl2vvp_page(slice);
 449         struct vvp_object *clobj = cl2vvp(slice->cpl_obj);
 450
 451         vvp_page_fini_common(vpg, pvec);
 452         atomic_dec(&clobj->vob_transient_pages);
 453 }
 454
 455 static const struct cl_page_operations vvp_transient_page_ops = {
 456         .cpo_discard            = vvp_transient_page_discard,
 457         .cpo_fini               = vvp_transient_page_fini,
 458         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 459         .cpo_print              = vvp_page_print,
 460 };
 461
 462 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 463                 struct cl_page *page, pgoff_t index)
 464 {
 465         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 466         struct page     *vmpage = page->cp_vmpage;
 467
 468         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 469
 470         vpg->vpg_page = vmpage;
 471         get_page(vmpage);
 472
 473         if (page->cp_type == CPT_CACHEABLE) {
 474                 /* in cache, decref in vvp_page_delete */
 475                 atomic_inc(&page->cp_ref);
 476                 SetPagePrivate(vmpage);
 477                 vmpage->private = (unsigned long)page;
 478                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 479                                 &vvp_page_ops);
 480         } else {
 481                 struct vvp_object *clobj = cl2vvp(obj);
 482
 483                 cl_page_slice_add(page, &vpg->vpg_cl, obj, index,
 484                                 &vvp_transient_page_ops);
 485                 atomic_inc(&clobj->vob_transient_pages);
 486         }
 487         return 0;
 488 }