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  *
  31  * Implementation of cl_page for VVP layer.
  32  *
  33  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  34  *   Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LLITE
  38
  39 #include <linux/atomic.h>
  40 #include <linux/bitops.h>
  41 #include <linux/mm.h>
  42 #include <linux/mutex.h>
  43 #include <linux/page-flags.h>
  44 #include <linux/pagemap.h>
  45
  46 #include <libcfs/libcfs.h>
  47 #include "llite_internal.h"
  48 #include "vvp_internal.h"
  49
  50 /*****************************************************************************
  51  *
  52  * Page operations.
  53  *
  54  */
  55
  56 static void vvp_page_fini_common(struct vvp_page *vpg, struct pagevec *pvec)
  57 {
  58         struct page *vmpage = vpg->vpg_page;
  59
  60         LASSERT(vmpage != NULL);
  61         if (pvec) {
  62                 if (!pagevec_add(pvec, vmpage))
  63                         pagevec_release(pvec);
  64         } else {
  65                 put_page(vmpage);
  66         }
  67 }
  68
  69 static void vvp_page_fini(const struct lu_env *env,
  70                           struct cl_page_slice *slice,
  71                           struct pagevec *pvec)
  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, pvec);
  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         ENTRY;
  92
  93         LASSERT(vmpage != NULL);
  94         if (nonblock) {
  95                 if (!trylock_page(vmpage))
  96                         return -EAGAIN;
  97
  98                 if (unlikely(PageWriteback(vmpage))) {
  99                         unlock_page(vmpage);
 100                         return -EAGAIN;
 101                 }
 102
 103                 return 0;
 104         }
 105
 106         lock_page(vmpage);
 107         wait_on_page_writeback(vmpage);
 108
 109         RETURN(0);
 110 }
 111
 112 static void vvp_page_assume(const struct lu_env *env,
 113                             const struct cl_page_slice *slice,
 114                             struct cl_io *unused)
 115 {
 116         struct page *vmpage = cl2vm_page(slice);
 117
 118         LASSERT(vmpage != NULL);
 119         LASSERT(PageLocked(vmpage));
 120         wait_on_page_writeback(vmpage);
 121 }
 122
 123 static void vvp_page_unassume(const struct lu_env *env,
 124                               const struct cl_page_slice *slice,
 125                               struct cl_io *unused)
 126 {
 127         struct page *vmpage = cl2vm_page(slice);
 128
 129         LASSERT(vmpage != NULL);
 130         LASSERT(PageLocked(vmpage));
 131 }
 132
 133 static void vvp_page_disown(const struct lu_env *env,
 134                             const struct cl_page_slice *slice, struct cl_io *io)
 135 {
 136         struct page *vmpage = cl2vm_page(slice);
 137
 138         ENTRY;
 139
 140         LASSERT(vmpage != NULL);
 141         LASSERT(PageLocked(vmpage));
 142
 143         unlock_page(cl2vm_page(slice));
 144
 145         EXIT;
 146 }
 147
 148 static void vvp_page_discard(const struct lu_env *env,
 149                              const struct cl_page_slice *slice,
 150                              struct cl_io *unused)
 151 {
 152         struct page     *vmpage = cl2vm_page(slice);
 153         struct vvp_page *vpg    = cl2vvp_page(slice);
 154
 155         LASSERT(vmpage != NULL);
 156         LASSERT(PageLocked(vmpage));
 157
 158         if (vpg->vpg_defer_uptodate && !vpg->vpg_ra_used && vmpage->mapping)
 159                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 160
 161         generic_error_remove_page(vmpage->mapping, vmpage);
 162 }
 163
 164 static void vvp_page_delete(const struct lu_env *env,
 165                             const struct cl_page_slice *slice)
 166 {
 167         struct page      *vmpage = cl2vm_page(slice);
 168         struct cl_page   *page   = slice->cpl_page;
 169         int refc;
 170
 171         LASSERT(PageLocked(vmpage));
 172         LASSERT((struct cl_page *)vmpage->private == page);
 173
 174
 175         /* Drop the reference count held in vvp_page_init */
 176         refc = atomic_dec_return(&page->cp_ref);
 177         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 178
 179         ClearPagePrivate(vmpage);
 180         vmpage->private = 0;
 181         /*
 182          * Reference from vmpage to cl_page is removed, but the reference back
 183          * is still here. It is removed later in vvp_page_fini().
 184          */
 185 }
 186
 187 static void vvp_page_export(const struct lu_env *env,
 188                             const struct cl_page_slice *slice,
 189                             int uptodate)
 190 {
 191         struct page *vmpage = cl2vm_page(slice);
 192
 193         LASSERT(vmpage != NULL);
 194         LASSERT(PageLocked(vmpage));
 195         if (uptodate)
 196                 SetPageUptodate(vmpage);
 197         else
 198                 ClearPageUptodate(vmpage);
 199 }
 200
 201 static int vvp_page_is_vmlocked(const struct lu_env *env,
 202                                 const struct cl_page_slice *slice)
 203 {
 204         return PageLocked(cl2vm_page(slice)) ? -EBUSY : -ENODATA;
 205 }
 206
 207 static int vvp_page_prep_read(const struct lu_env *env,
 208                               const struct cl_page_slice *slice,
 209                               struct cl_io *unused)
 210 {
 211         ENTRY;
 212         /* Skip the page already marked as PG_uptodate. */
 213         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 214 }
 215
 216 static int vvp_page_prep_write(const struct lu_env *env,
 217                                const struct cl_page_slice *slice,
 218                                struct cl_io *unused)
 219 {
 220         struct page *vmpage = cl2vm_page(slice);
 221         struct cl_page *pg = slice->cpl_page;
 222
 223         LASSERT(PageLocked(vmpage));
 224         LASSERT(!PageDirty(vmpage));
 225
 226         /* ll_writepage path is not a sync write, so need to set page writeback
 227          * flag
 228          */
 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,
 242                              int ioret)
 243 {
 244         struct vvp_object *obj = cl_inode2vvp(inode);
 245
 246         if (ioret == 0) {
 247                 ClearPageError(vmpage);
 248                 obj->vob_discard_page_warned = 0;
 249         } else {
 250                 SetPageError(vmpage);
 251                 if (ioret == -ENOSPC)
 252                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 253                 else
 254                         set_bit(AS_EIO, &inode->i_mapping->flags);
 255
 256                 if ((ioret == -ESHUTDOWN || ioret == -EINTR ||
 257                      ioret == -EIO) && obj->vob_discard_page_warned == 0) {
 258                         obj->vob_discard_page_warned = 1;
 259                         ll_dirty_page_discard_warn(vmpage, ioret);
 260                 }
 261         }
 262 }
 263
 264 static void vvp_page_completion_read(const struct lu_env *env,
 265                                      const struct cl_page_slice *slice,
 266                                      int ioret)
 267 {
 268         struct vvp_page *vpg    = cl2vvp_page(slice);
 269         struct page     *vmpage = vpg->vpg_page;
 270         struct cl_page  *page   = slice->cpl_page;
 271         struct inode    *inode  = vvp_object_inode(page->cp_obj);
 272
 273         ENTRY;
 274         LASSERT(PageLocked(vmpage));
 275         CL_PAGE_HEADER(D_PAGE, env, page, "completing READ with %d\n", ioret);
 276
 277         if (vpg->vpg_defer_uptodate)
 278                 ll_ra_count_put(ll_i2sbi(inode), 1);
 279
 280         if (ioret == 0)  {
 281                 if (!vpg->vpg_defer_uptodate)
 282                         cl_page_export(env, page, 1);
 283         } else if (vpg->vpg_defer_uptodate) {
 284                 vpg->vpg_defer_uptodate = 0;
 285                 if (ioret == -EAGAIN) {
 286                         /* mirror read failed, it needs to destroy the page
 287                          * because subpage would be from wrong osc when trying
 288                          * to read from a new mirror
 289                          */
 290                         generic_error_remove_page(vmpage->mapping, vmpage);
 291                 }
 292         }
 293
 294         if (page->cp_sync_io == NULL)
 295                 unlock_page(vmpage);
 296
 297         EXIT;
 298 }
 299
 300 static void vvp_page_completion_write(const struct lu_env *env,
 301                                       const struct cl_page_slice *slice,
 302                                       int ioret)
 303 {
 304         struct vvp_page *vpg    = cl2vvp_page(slice);
 305         struct cl_page  *pg     = slice->cpl_page;
 306         struct page     *vmpage = vpg->vpg_page;
 307
 308         ENTRY;
 309         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 310
 311         if (pg->cp_sync_io != NULL) {
 312                 LASSERT(PageLocked(vmpage));
 313                 LASSERT(!PageWriteback(vmpage));
 314         } else {
 315                 LASSERT(PageWriteback(vmpage));
 316                 /*
 317                  * Only mark the page error only when it's an async write
 318                  * because applications won't wait for IO to finish.
 319                  */
 320                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 321
 322                 end_page_writeback(vmpage);
 323         }
 324         EXIT;
 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
 352                  * tree.
 353                  */
 354                 set_page_writeback(vmpage);
 355                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 356         } else if (pg->cp_state == CPS_PAGEOUT) {
 357                 /* is it possible for osc_flush_async_page() to already
 358                  * make it ready?
 359                  */
 360                 result = -EALREADY;
 361         } else {
 362                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 363                               pg->cp_state);
 364                 LBUG();
 365         }
 366         unlock_page(vmpage);
 367         RETURN(result);
 368 }
 369
 370 static int vvp_page_print(const struct lu_env *env,
 371                           const struct cl_page_slice *slice,
 372                           void *cookie, lu_printer_t printer)
 373 {
 374         struct vvp_page *vpg    = cl2vvp_page(slice);
 375         struct page     *vmpage = vpg->vpg_page;
 376
 377         (*printer)(env, cookie,
 378                    LUSTRE_VVP_NAME"-page@%p(%d:%d) vm@%p ",
 379                    vpg, vpg->vpg_defer_uptodate, vpg->vpg_ra_used, vmpage);
 380
 381         if (vmpage != NULL) {
 382                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 383                            (long)vmpage->flags, page_count(vmpage),
 384                            page_mapcount(vmpage), vmpage->private,
 385                            page_index(vmpage),
 386                            list_empty(&vmpage->lru) ? "not-" : "");
 387         }
 388
 389         (*printer)(env, cookie, "\n");
 390
 391         return 0;
 392 }
 393
 394 static int vvp_page_fail(const struct lu_env *env,
 395                          const struct cl_page_slice *slice)
 396 {
 397         /*
 398          * Cached read?
 399          */
 400         LBUG();
 401
 402         return 0;
 403 }
 404
 405 static const struct cl_page_operations vvp_page_ops = {
 406         .cpo_own           = vvp_page_own,
 407         .cpo_assume        = vvp_page_assume,
 408         .cpo_unassume      = vvp_page_unassume,
 409         .cpo_disown        = vvp_page_disown,
 410         .cpo_discard       = vvp_page_discard,
 411         .cpo_delete        = vvp_page_delete,
 412         .cpo_export        = vvp_page_export,
 413         .cpo_is_vmlocked   = vvp_page_is_vmlocked,
 414         .cpo_fini          = vvp_page_fini,
 415         .cpo_print         = vvp_page_print,
 416         .io = {
 417                 [CRT_READ] = {
 418                         .cpo_prep       = vvp_page_prep_read,
 419                         .cpo_completion = vvp_page_completion_read,
 420                         .cpo_make_ready = vvp_page_fail,
 421                 },
 422                 [CRT_WRITE] = {
 423                         .cpo_prep       = vvp_page_prep_write,
 424                         .cpo_completion = vvp_page_completion_write,
 425                         .cpo_make_ready = vvp_page_make_ready,
 426                 },
 427         },
 428 };
 429
 430 static void vvp_transient_page_discard(const struct lu_env *env,
 431                                        const struct cl_page_slice *slice,
 432                                        struct cl_io *unused)
 433 {
 434         struct cl_page *page = slice->cpl_page;
 435
 436         /*
 437          * For transient pages, remove it from the radix tree.
 438          */
 439         cl_page_delete(env, page);
 440 }
 441
 442 static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
 443                                           const struct cl_page_slice *slice)
 444 {
 445         return -EBUSY;
 446 }
 447
 448 static void vvp_transient_page_fini(const struct lu_env *env,
 449                                     struct cl_page_slice *slice,
 450                                     struct pagevec *pvec)
 451 {
 452         struct vvp_page *vpg = cl2vvp_page(slice);
 453         struct vvp_object *clobj = cl2vvp(slice->cpl_obj);
 454
 455         vvp_page_fini_common(vpg, pvec);
 456         atomic_dec(&clobj->vob_transient_pages);
 457 }
 458
 459 static const struct cl_page_operations vvp_transient_page_ops = {
 460         .cpo_discard            = vvp_transient_page_discard,
 461         .cpo_fini               = vvp_transient_page_fini,
 462         .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
 463         .cpo_print              = vvp_page_print,
 464 };
 465
 466 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 467                 struct cl_page *page, pgoff_t index)
 468 {
 469         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 470         struct page     *vmpage = page->cp_vmpage;
 471
 472         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 473
 474         vpg->vpg_page = vmpage;
 475         get_page(vmpage);
 476
 477         if (page->cp_type == CPT_CACHEABLE) {
 478                 /* in cache, decref in vvp_page_delete */
 479                 atomic_inc(&page->cp_ref);
 480                 SetPagePrivate(vmpage);
 481                 vmpage->private = (unsigned long)page;
 482                 cl_page_slice_add(page, &vpg->vpg_cl, obj,
 483                                 &vvp_page_ops);
 484         } else {
 485                 struct vvp_object *clobj = cl2vvp(obj);
 486
 487                 cl_page_slice_add(page, &vpg->vpg_cl, obj,
 488                                 &vvp_transient_page_ops);
 489                 atomic_inc(&clobj->vob_transient_pages);
 490         }
 491         return 0;
 492 }