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_discard(const struct lu_env *env,
  56                              const struct cl_page_slice *slice,
  57                              struct cl_io *unused)
  58 {
  59         struct cl_page *cp = slice->cpl_page;
  60         struct page *vmpage = cp->cp_vmpage;
  61
  62         if (cp->cp_defer_uptodate && !cp->cp_ra_used && vmpage->mapping != NULL)
  63                 ll_ra_stats_inc(vmpage->mapping->host, RA_STAT_DISCARDED);
  64 }
  65
  66 static void vvp_page_delete(const struct lu_env *env,
  67                             const struct cl_page_slice *slice)
  68 {
  69         struct cl_page *cp = slice->cpl_page;
  70
  71         if (cp->cp_type == CPT_CACHEABLE) {
  72                 struct page *vmpage = cp->cp_vmpage;
  73                 struct inode *inode = vmpage->mapping->host;
  74
  75                 LASSERT(PageLocked(vmpage));
  76                 LASSERT((struct cl_page *)vmpage->private == cp);
  77
  78                 /* Drop the reference count held in vvp_page_init */
  79                 refcount_dec(&cp->cp_ref);
  80
  81                 ClearPagePrivate(vmpage);
  82                 vmpage->private = 0;
  83
  84                 /* clearpageuptodate prevents the page being read by the
  85                  * kernel after it has been deleted from Lustre, which avoids
  86                  * potential stale data reads.  The seqlock allows us to see
  87                  * that a page was potentially deleted and catch the resulting
  88                  * SIGBUS - see ll_filemap_fault() (LU-16160) */
  89                 write_seqlock(&ll_i2info(inode)->lli_page_inv_lock);
  90                 ClearPageUptodate(vmpage);
  91                 write_sequnlock(&ll_i2info(inode)->lli_page_inv_lock);
  92
  93                 /*
  94                  * The reference from vmpage to cl_page is removed,
  95                  * but the reference back is still here. It is removed
  96                  * later in cl_page_free().
  97                  */
  98         }
  99 }
 100
 101 /**
 102  * Handles page transfer errors at VM level.
 103  *
 104  * This takes inode as a separate argument, because inode on which error is to
 105  * be set can be different from \a vmpage inode in case of direct-io.
 106  */
 107 static void vvp_vmpage_error(struct inode *inode, struct page *vmpage,
 108                              int ioret)
 109 {
 110         struct vvp_object *obj = cl_inode2vvp(inode);
 111
 112         if (ioret == 0) {
 113                 ClearPageError(vmpage);
 114                 obj->vob_discard_page_warned = 0;
 115         } else {
 116                 SetPageError(vmpage);
 117                 if (ioret != -ENOSPC &&
 118                     CFS_FAIL_CHECK(OBD_FAIL_LLITE_PANIC_ON_ESTALE))
 119                         LBUG();
 120
 121                 mapping_set_error(inode->i_mapping, ioret);
 122
 123                 if ((ioret == -ESHUTDOWN || ioret == -EINTR ||
 124                      ioret == -EIO) && obj->vob_discard_page_warned == 0) {
 125                         obj->vob_discard_page_warned = 1;
 126                         ll_dirty_page_discard_warn(inode, ioret);
 127                 }
 128         }
 129 }
 130
 131 static void vvp_page_completion_read(const struct lu_env *env,
 132                                      const struct cl_page_slice *slice,
 133                                      int ioret)
 134 {
 135         struct cl_page *cp = slice->cpl_page;
 136         struct page *vmpage = cp->cp_vmpage;
 137         struct inode *inode = vvp_object_inode(cp->cp_obj);
 138
 139         ENTRY;
 140         LASSERT(PageLocked(vmpage));
 141         CL_PAGE_HEADER(D_PAGE, env, cp, "completing READ with %d\n", ioret);
 142
 143         if (cp->cp_defer_uptodate)
 144                 ll_ra_count_put(ll_i2sbi(inode), 1);
 145
 146         if (ioret == 0)  {
 147                 /**
 148                  * cp_defer_uptodate is used for readahead page, and the
 149                  * vmpage Uptodate bit is deferred to set in ll_readpage/
 150                  * ll_io_read_page.
 151                  */
 152                 if (!cp->cp_defer_uptodate)
 153                         SetPageUptodate(vmpage);
 154         } else if (cp->cp_defer_uptodate) {
 155                 cp->cp_defer_uptodate = 0;
 156                 if (ioret == -EAGAIN) {
 157                         /* mirror read failed, it needs to destroy the page
 158                          * because subpage would be from wrong osc when trying
 159                          * to read from a new mirror
 160                          */
 161                         generic_error_remove_page(vmpage->mapping, vmpage);
 162                 }
 163         }
 164
 165         if (cp->cp_sync_io == NULL)
 166                 unlock_page(vmpage);
 167
 168         EXIT;
 169 }
 170
 171 static void vvp_page_completion_write(const struct lu_env *env,
 172                                       const struct cl_page_slice *slice,
 173                                       int ioret)
 174 {
 175         struct cl_page *cp = slice->cpl_page;
 176         struct page *vmpage = cp->cp_vmpage;
 177
 178         ENTRY;
 179         CL_PAGE_HEADER(D_PAGE, env, cp, "completing WRITE with %d\n", ioret);
 180
 181         if (cp->cp_sync_io != NULL) {
 182                 LASSERT(PageLocked(vmpage));
 183                 LASSERT(!PageWriteback(vmpage));
 184         } else {
 185                 LASSERT(PageWriteback(vmpage));
 186                 /*
 187                  * Only mark the page error only when it's an async write
 188                  * because applications won't wait for IO to finish.
 189                  */
 190                 vvp_vmpage_error(vvp_object_inode(cp->cp_obj), vmpage, ioret);
 191
 192                 end_page_writeback(vmpage);
 193         }
 194         EXIT;
 195 }
 196
 197 static const struct cl_page_operations vvp_page_ops = {
 198         .cpo_delete        = vvp_page_delete,
 199         .cpo_discard       = vvp_page_discard,
 200         .io = {
 201                 [CRT_READ] = {
 202                         .cpo_completion = vvp_page_completion_read,
 203                 },
 204                 [CRT_WRITE] = {
 205                         .cpo_completion = vvp_page_completion_write,
 206                 },
 207         },
 208 };
 209
 210 static const struct cl_page_operations vvp_transient_page_ops = {
 211 };
 212
 213 int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
 214                 struct cl_page *page, pgoff_t index)
 215 {
 216         struct cl_page_slice *cpl = cl_object_page_slice(obj, page);
 217         struct page *vmpage = page->cp_vmpage;
 218
 219         CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 220
 221         if (page->cp_type == CPT_TRANSIENT) {
 222                 /* DIO pages are referenced by userspace, we don't need to take
 223                  * a reference on them. (contrast with get_page() call above)
 224                  */
 225                 cl_page_slice_add(page, cpl, obj,
 226                                   &vvp_transient_page_ops);
 227         } else {
 228                 get_page(vmpage);
 229                 /* in cache, decref in cl_page_delete() */
 230                 refcount_inc(&page->cp_ref);
 231                 SetPagePrivate(vmpage);
 232                 vmpage->private = (unsigned long)page;
 233                 cl_page_slice_add(page, cpl, obj, &vvp_page_ops);
 234         }
 235
 236         return 0;
 237 }