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 static void vvp_page_fini(const struct lu_env *env,
  56                           struct cl_page_slice *slice,
  57                           struct pagevec *pvec)
  58 {
  59         struct vvp_page *vpg     = cl2vvp_page(slice);
  60         struct page     *vmpage  = vpg->vpg_page;
  61
  62         /*
  63          * vmpage->private was already cleared when page was moved into
  64          * VPG_FREEING state.
  65          */
  66         LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
  67         LASSERT(vmpage != NULL);
  68         if (pvec) {
  69                 if (!pagevec_add(pvec, vmpage))
  70                         pagevec_release(pvec);
  71         } else {
  72                 put_page(vmpage);
  73         }
  74 }
  75
  76 static void vvp_page_assume(const struct lu_env *env,
  77                             const struct cl_page_slice *slice,
  78                             struct cl_io *unused)
  79 {
  80         struct page *vmpage = cl2vm_page(slice);
  81
  82         LASSERT(vmpage != NULL);
  83         LASSERT(PageLocked(vmpage));
  84         wait_on_page_writeback(vmpage);
  85 }
  86
  87 static void vvp_page_unassume(const struct lu_env *env,
  88                               const struct cl_page_slice *slice,
  89                               struct cl_io *unused)
  90 {
  91         struct page *vmpage = cl2vm_page(slice);
  92
  93         LASSERT(vmpage != NULL);
  94         LASSERT(PageLocked(vmpage));
  95 }
  96
  97 static void vvp_page_discard(const struct lu_env *env,
  98                              const struct cl_page_slice *slice,
  99                              struct cl_io *unused)
 100 {
 101         struct cl_page *cp = slice->cpl_page;
 102         struct page *vmpage = cp->cp_vmpage;
 103
 104         if (cp->cp_defer_uptodate && !cp->cp_ra_used && vmpage->mapping != NULL)
 105                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
 106 }
 107
 108 static void vvp_page_delete(const struct lu_env *env,
 109                             const struct cl_page_slice *slice)
 110 {
 111         struct page      *vmpage = cl2vm_page(slice);
 112         struct cl_page   *page   = slice->cpl_page;
 113         int refc;
 114
 115         LASSERT(PageLocked(vmpage));
 116         LASSERT((struct cl_page *)vmpage->private == page);
 117
 118
 119         /* Drop the reference count held in vvp_page_init */
 120         refc = atomic_dec_return(&page->cp_ref);
 121         LASSERTF(refc >= 1, "page = %p, refc = %d\n", page, refc);
 122
 123         ClearPagePrivate(vmpage);
 124         vmpage->private = 0;
 125         /*
 126          * Reference from vmpage to cl_page is removed, but the reference back
 127          * is still here. It is removed later in vvp_page_fini().
 128          */
 129 }
 130
 131 static int vvp_page_prep_read(const struct lu_env *env,
 132                               const struct cl_page_slice *slice,
 133                               struct cl_io *unused)
 134 {
 135         ENTRY;
 136         /* Skip the page already marked as PG_uptodate. */
 137         RETURN(PageUptodate(cl2vm_page(slice)) ? -EALREADY : 0);
 138 }
 139
 140 static int vvp_page_prep_write(const struct lu_env *env,
 141                                const struct cl_page_slice *slice,
 142                                struct cl_io *unused)
 143 {
 144         struct page *vmpage = cl2vm_page(slice);
 145         struct cl_page *pg = slice->cpl_page;
 146
 147         LASSERT(PageLocked(vmpage));
 148         LASSERT(!PageDirty(vmpage));
 149
 150         /* ll_writepage path is not a sync write, so need to set page writeback
 151          * flag
 152          */
 153         if (pg->cp_sync_io == NULL)
 154                 set_page_writeback(vmpage);
 155
 156         return 0;
 157 }
 158
 159 /**
 160  * Handles page transfer errors at VM level.
 161  *
 162  * This takes inode as a separate argument, because inode on which error is to
 163  * be set can be different from \a vmpage inode in case of direct-io.
 164  */
 165 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage,
 166                              int ioret)
 167 {
 168         struct vvp_object *obj = cl_inode2vvp(inode);
 169
 170         if (ioret == 0) {
 171                 ClearPageError(vmpage);
 172                 obj->vob_discard_page_warned = 0;
 173         } else {
 174                 SetPageError(vmpage);
 175                 if (ioret == -ENOSPC)
 176                         set_bit(AS_ENOSPC, &inode->i_mapping->flags);
 177                 else
 178                         set_bit(AS_EIO, &inode->i_mapping->flags);
 179
 180                 if ((ioret == -ESHUTDOWN || ioret == -EINTR ||
 181                      ioret == -EIO) && obj->vob_discard_page_warned == 0) {
 182                         obj->vob_discard_page_warned = 1;
 183                         ll_dirty_page_discard_warn(inode, ioret);
 184                 }
 185         }
 186 }
 187
 188 static void vvp_page_completion_read(const struct lu_env *env,
 189                                      const struct cl_page_slice *slice,
 190                                      int ioret)
 191 {
 192         struct cl_page *cp = slice->cpl_page;
 193         struct page *vmpage = cp->cp_vmpage;
 194         struct inode *inode = vvp_object_inode(cp->cp_obj);
 195
 196         ENTRY;
 197         LASSERT(PageLocked(vmpage));
 198         CL_PAGE_HEADER(D_PAGE, env, cp, "completing READ with %d\n", ioret);
 199
 200         if (cp->cp_defer_uptodate)
 201                 ll_ra_count_put(ll_i2sbi(inode), 1);
 202
 203         if (ioret == 0)  {
 204                 if (!cp->cp_defer_uptodate)
 205                         SetPageUptodate(vmpage);
 206         } else if (cp->cp_defer_uptodate) {
 207                 cp->cp_defer_uptodate = 0;
 208                 if (ioret == -EAGAIN) {
 209                         /* mirror read failed, it needs to destroy the page
 210                          * because subpage would be from wrong osc when trying
 211                          * to read from a new mirror
 212                          */
 213                         generic_error_remove_page(vmpage->mapping, vmpage);
 214                 }
 215         }
 216
 217         if (cp->cp_sync_io == NULL)
 218                 unlock_page(vmpage);
 219
 220         EXIT;
 221 }
 222
 223 static void vvp_page_completion_write(const struct lu_env *env,
 224                                       const struct cl_page_slice *slice,
 225                                       int ioret)
 226 {
 227         struct vvp_page *vpg    = cl2vvp_page(slice);
 228         struct cl_page  *pg     = slice->cpl_page;
 229         struct page     *vmpage = vpg->vpg_page;
 230
 231         ENTRY;
 232         CL_PAGE_HEADER(D_PAGE, env, pg, "completing WRITE with %d\n", ioret);
 233
 234         if (pg->cp_sync_io != NULL) {
 235                 LASSERT(PageLocked(vmpage));
 236                 LASSERT(!PageWriteback(vmpage));
 237         } else {
 238                 LASSERT(PageWriteback(vmpage));
 239                 /*
 240                  * Only mark the page error only when it's an async write
 241                  * because applications won't wait for IO to finish.
 242                  */
 243                 vvp_vmpage_error(vvp_object_inode(pg->cp_obj), vmpage, ioret);
 244
 245                 end_page_writeback(vmpage);
 246         }
 247         EXIT;
 248 }
 249
 250 /**
 251  * Implements cl_page_operations::cpo_make_ready() method.
 252  *
 253  * This is called to yank a page from the transfer cache and to send it out as
 254  * a part of transfer. This function try-locks the page. If try-lock failed,
 255  * page is owned by some concurrent IO, and should be skipped (this is bad,
 256  * but hopefully rare situation, as it usually results in transfer being
 257  * shorter than possible).
 258  *
 259  * \retval 0      success, page can be placed into transfer
 260  *
 261  * \retval -EAGAIN page is either used by concurrent IO has been
 262  * truncated. Skip it.
 263  */
 264 static int vvp_page_make_ready(const struct lu_env *env,
 265                                const struct cl_page_slice *slice)
 266 {
 267         struct page *vmpage = cl2vm_page(slice);
 268         struct cl_page *pg = slice->cpl_page;
 269         int result = 0;
 270
 271         lock_page(vmpage);
 272         if (clear_page_dirty_for_io(vmpage)) {
 273                 LASSERT(pg->cp_state == CPS_CACHED);
 274                 /* This actually clears the dirty bit in the radix
 275                  * tree.
 276                  */
 277                 set_page_writeback(vmpage);
 278                 CL_PAGE_HEADER(D_PAGE, env, pg, "readied\n");
 279         } else if (pg->cp_state == CPS_PAGEOUT) {
 280                 /* is it possible for osc_flush_async_page() to already
 281                  * make it ready?
 282                  */
 283                 result = -EALREADY;
 284         } else {
 285                 CL_PAGE_DEBUG(D_ERROR, env, pg, "Unexpecting page state %d.\n",
 286                               pg->cp_state);
 287                 LBUG();
 288         }
 289         unlock_page(vmpage);
 290         RETURN(result);
 291 }
 292
 293 static int vvp_page_print(const struct lu_env *env,
 294                           const struct cl_page_slice *slice,
 295                           void *cookie, lu_printer_t printer)
 296 {
 297         struct vvp_page *vpg    = cl2vvp_page(slice);
 298         struct page     *vmpage = vpg->vpg_page;
 299
 300         (*printer)(env, cookie,
 301                    LUSTRE_VVP_NAME"-page@%p vm@%p ", vpg, vmpage);
 302
 303         if (vmpage != NULL) {
 304                 (*printer)(env, cookie, "%lx %d:%d %lx %lu %slru",
 305                            (long)vmpage->flags, page_count(vmpage),
 306                            page_mapcount(vmpage), vmpage->private,
 307                            page_index(vmpage),
 308                            list_empty(&vmpage->lru) ? "not-" : "");
 309         }
 310
 311         (*printer)(env, cookie, "\n");
 312
 313         return 0;
 314 }
 315
 316 static int vvp_page_fail(const struct lu_env *env,
 317                          const struct cl_page_slice *slice)
 318 {
 319         /*
 320          * Cached read?
 321          */
 322         LBUG();
 323
 324         return 0;
 325 }
 326
 327 static const struct cl_page_operations vvp_page_ops = {
 328         .cpo_assume        = vvp_page_assume,
 329         .cpo_unassume      = vvp_page_unassume,
 330         .cpo_discard       = vvp_page_discard,
 331         .cpo_delete        = vvp_page_delete,
 332         .cpo_fini          = vvp_page_fini,
 333         .cpo_print         = vvp_page_print,
 334         .io = {
 335                 [CRT_READ] = {
 336                         .cpo_prep       = vvp_page_prep_read,
 337                         .cpo_completion = vvp_page_completion_read,
 338                         .cpo_make_ready = vvp_page_fail,
 339                 },
 340                 [CRT_WRITE] = {
 341                         .cpo_prep       = vvp_page_prep_write,
 342                         .cpo_completion = vvp_page_completion_write,
 343                         .cpo_make_ready = vvp_page_make_ready,
 344                 },
 345         },
 346 };
 347
 348 static const struct cl_page_operations vvp_transient_page_ops = {
 349         .cpo_print              = vvp_page_print,
 350 };
 351
 352 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 353                 struct cl_page *page, pgoff_t index)
 354 {
 355         struct vvp_page *vpg = cl_object_page_slice(obj, page);
 356         struct page     *vmpage = page->cp_vmpage;
 357
 358         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 359
 360         vpg->vpg_page = vmpage;
 361
 362         if (page->cp_type == CPT_TRANSIENT) {
 363                 /* DIO pages are referenced by userspace, we don't need to take
 364                  * a reference on them. (contrast with get_page() call above)
 365                  */
 366                 cl_page_slice_add(page, &vpg->vpg_cl, obj,
 367                                   &vvp_transient_page_ops);
 368         } else {
 369                 get_page(vmpage);
 370                 /* in cache, decref in vvp_page_delete */
 371                 atomic_inc(&page->cp_ref);
 372                 SetPagePrivate(vmpage);
 373                 vmpage->private = (unsigned long)page;
 374                 cl_page_slice_add(page, &vpg->vpg_cl, obj,
 375                                 &vvp_page_ops);
 376         }
 377
 378         return 0;
 379 }