fs/xfs/libxfs/xfs_ag_resv.c

   1 /*
   2  * Copyright (C) 2016 Oracle.  All Rights Reserved.
   3  *
   4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version 2
   9  * of the License, or (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it would be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write the Free Software Foundation,
  18  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301, USA.
  19  */
  20 #include "xfs.h"
  21 #include "xfs_fs.h"
  22 #include "xfs_shared.h"
  23 #include "xfs_format.h"
  24 #include "xfs_log_format.h"
  25 #include "xfs_trans_resv.h"
  26 #include "xfs_sb.h"
  27 #include "xfs_mount.h"
  28 #include "xfs_defer.h"
  29 #include "xfs_alloc.h"
  30 #include "xfs_error.h"
  31 #include "xfs_trace.h"
  32 #include "xfs_cksum.h"
  33 #include "xfs_trans.h"
  34 #include "xfs_bit.h"
  35 #include "xfs_bmap.h"
  36 #include "xfs_bmap_btree.h"
  37 #include "xfs_ag_resv.h"
  38 #include "xfs_trans_space.h"
  39 #include "xfs_rmap_btree.h"
  40 #include "xfs_btree.h"
  41 #include "xfs_refcount_btree.h"
  42 #include "xfs_ialloc_btree.h"
  43
  44 /*
  45  * Per-AG Block Reservations
  46  *
  47  * For some kinds of allocation group metadata structures, it is advantageous
  48  * to reserve a small number of blocks in each AG so that future expansions of
  49  * that data structure do not encounter ENOSPC because errors during a btree
  50  * split cause the filesystem to go offline.
  51  *
  52  * Prior to the introduction of reflink, this wasn't an issue because the free
  53  * space btrees maintain a reserve of space (the AGFL) to handle any expansion
  54  * that may be necessary; and allocations of other metadata (inodes, BMBT,
  55  * dir/attr) aren't restricted to a single AG.  However, with reflink it is
  56  * possible to allocate all the space in an AG, have subsequent reflink/CoW
  57  * activity expand the refcount btree, and discover that there's no space left
  58  * to handle that expansion.  Since we can calculate the maximum size of the
  59  * refcount btree, we can reserve space for it and avoid ENOSPC.
  60  *
  61  * Handling per-AG reservations consists of three changes to the allocator's
  62  * behavior:  First, because these reservations are always needed, we decrease
  63  * the ag_max_usable counter to reflect the size of the AG after the reserved
  64  * blocks are taken.  Second, the reservations must be reflected in the
  65  * fdblocks count to maintain proper accounting.  Third, each AG must maintain
  66  * its own reserved block counter so that we can calculate the amount of space
  67  * that must remain free to maintain the reservations.  Fourth, the "remaining
  68  * reserved blocks" count must be used when calculating the length of the
  69  * longest free extent in an AG and to clamp maxlen in the per-AG allocation
  70  * functions.  In other words, we maintain a virtual allocation via in-core
  71  * accounting tricks so that we don't have to clean up after a crash. :)
  72  *
  73  * Reserved blocks can be managed by passing one of the enum xfs_ag_resv_type
  74  * values via struct xfs_alloc_arg or directly to the xfs_free_extent
  75  * function.  It might seem a little funny to maintain a reservoir of blocks
  76  * to feed another reservoir, but the AGFL only holds enough blocks to get
  77  * through the next transaction.  The per-AG reservation is to ensure (we
  78  * hope) that each AG never runs out of blocks.  Each data structure wanting
  79  * to use the reservation system should update ask/used in xfs_ag_resv_init.
  80  */
  81
  82 /*
  83  * Are we critically low on blocks?  For now we'll define that as the number
  84  * of blocks we can get our hands on being less than 10% of what we reserved
  85  * or less than some arbitrary number (maximum btree height).
  86  */
  87 bool
  88 xfs_ag_resv_critical(
  89         struct xfs_perag                *pag,
  90         enum xfs_ag_resv_type           type)
  91 {
  92         xfs_extlen_t                    avail;
  93         xfs_extlen_t                    orig;
  94
  95         switch (type) {
  96         case XFS_AG_RESV_METADATA:
  97                 avail = pag->pagf_freeblks - pag->pag_agfl_resv.ar_reserved;
  98                 orig = pag->pag_meta_resv.ar_asked;
  99                 break;
 100         case XFS_AG_RESV_AGFL:
 101                 avail = pag->pagf_freeblks + pag->pagf_flcount -
 102                         pag->pag_meta_resv.ar_reserved;
 103                 orig = pag->pag_agfl_resv.ar_asked;
 104                 break;
 105         default:
 106                 ASSERT(0);
 107                 return false;
 108         }
 109
 110         trace_xfs_ag_resv_critical(pag, type, avail);
 111
 112         /* Critically low if less than 10% or max btree height remains. */
 113         return XFS_TEST_ERROR(avail < orig / 10 || avail < XFS_BTREE_MAXLEVELS,
 114                         pag->pag_mount, XFS_ERRTAG_AG_RESV_CRITICAL,
 115                         XFS_RANDOM_AG_RESV_CRITICAL);
 116 }
 117
 118 /*
 119  * How many blocks are reserved but not used, and therefore must not be
 120  * allocated away?
 121  */
 122 xfs_extlen_t
 123 xfs_ag_resv_needed(
 124         struct xfs_perag                *pag,
 125         enum xfs_ag_resv_type           type)
 126 {
 127         xfs_extlen_t                    len;
 128
 129         len = pag->pag_meta_resv.ar_reserved + pag->pag_agfl_resv.ar_reserved;
 130         switch (type) {
 131         case XFS_AG_RESV_METADATA:
 132         case XFS_AG_RESV_AGFL:
 133                 len -= xfs_perag_resv(pag, type)->ar_reserved;
 134                 break;
 135         case XFS_AG_RESV_NONE:
 136                 /* empty */
 137                 break;
 138         default:
 139                 ASSERT(0);
 140         }
 141
 142         trace_xfs_ag_resv_needed(pag, type, len);
 143
 144         return len;
 145 }
 146
 147 /* Clean out a reservation */
 148 static int
 149 __xfs_ag_resv_free(
 150         struct xfs_perag                *pag,
 151         enum xfs_ag_resv_type           type)
 152 {
 153         struct xfs_ag_resv              *resv;
 154         xfs_extlen_t                    oldresv;
 155         int                             error;
 156
 157         trace_xfs_ag_resv_free(pag, type, 0);
 158
 159         resv = xfs_perag_resv(pag, type);
 160         if (pag->pag_agno == 0)
 161                 pag->pag_mount->m_ag_max_usable += resv->ar_asked;
 162         /*
 163          * AGFL blocks are always considered "free", so whatever
 164          * was reserved at mount time must be given back at umount.
 165          */
 166         if (type == XFS_AG_RESV_AGFL)
 167                 oldresv = resv->ar_orig_reserved;
 168         else
 169                 oldresv = resv->ar_reserved;
 170         error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
 171         resv->ar_reserved = 0;
 172         resv->ar_asked = 0;
 173
 174         if (error)
 175                 trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
 176                                 error, _RET_IP_);
 177         return error;
 178 }
 179
 180 /* Free a per-AG reservation. */
 181 int
 182 xfs_ag_resv_free(
 183         struct xfs_perag                *pag)
 184 {
 185         int                             error;
 186         int                             err2;
 187
 188         error = __xfs_ag_resv_free(pag, XFS_AG_RESV_AGFL);
 189         err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
 190         if (err2 && !error)
 191                 error = err2;
 192         return error;
 193 }
 194
 195 static int
 196 __xfs_ag_resv_init(
 197         struct xfs_perag                *pag,
 198         enum xfs_ag_resv_type           type,
 199         xfs_extlen_t                    ask,
 200         xfs_extlen_t                    used)
 201 {
 202         struct xfs_mount                *mp = pag->pag_mount;
 203         struct xfs_ag_resv              *resv;
 204         int                             error;
 205         xfs_extlen_t                    reserved;
 206
 207         if (used > ask)
 208                 ask = used;
 209         reserved = ask - used;
 210
 211         error = xfs_mod_fdblocks(mp, -(int64_t)reserved, true);
 212         if (error) {
 213                 trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
 214                                 error, _RET_IP_);
 215                 xfs_warn(mp,
 216 "Per-AG reservation for AG %u failed.  Filesystem may run out of space.",
 217                                 pag->pag_agno);
 218                 return error;
 219         }
 220
 221         /*
 222          * Reduce the maximum per-AG allocation length by however much we're
 223          * trying to reserve for an AG.  Since this is a filesystem-wide
 224          * counter, we only make the adjustment for AG 0.  This assumes that
 225          * there aren't any AGs hungrier for per-AG reservation than AG 0.
 226          */
 227         if (pag->pag_agno == 0)
 228                 mp->m_ag_max_usable -= ask;
 229
 230         resv = xfs_perag_resv(pag, type);
 231         resv->ar_asked = ask;
 232         resv->ar_reserved = resv->ar_orig_reserved = reserved;
 233
 234         trace_xfs_ag_resv_init(pag, type, ask);
 235         return 0;
 236 }
 237
 238 /* Create a per-AG block reservation. */
 239 int
 240 xfs_ag_resv_init(
 241         struct xfs_perag                *pag)
 242 {
 243         struct xfs_mount                *mp = pag->pag_mount;
 244         xfs_agnumber_t                  agno = pag->pag_agno;
 245         xfs_extlen_t                    ask;
 246         xfs_extlen_t                    used;
 247         int                             error = 0;
 248
 249         /* Create the metadata reservation. */
 250         if (pag->pag_meta_resv.ar_asked == 0) {
 251                 ask = used = 0;
 252
 253                 error = xfs_refcountbt_calc_reserves(mp, agno, &ask, &used);
 254                 if (error)
 255                         goto out;
 256
 257                 error = xfs_finobt_calc_reserves(mp, agno, &ask, &used);
 258                 if (error)
 259                         goto out;
 260
 261                 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
 262                                 ask, used);
 263                 if (error) {
 264                         /*
 265                          * Because we didn't have per-AG reservations when the
 266                          * finobt feature was added we might not be able to
 267                          * reserve all needed blocks.  Warn and fall back to the
 268                          * old and potentially buggy code in that case, but
 269                          * ensure we do have the reservation for the refcountbt.
 270                          */
 271                         ask = used = 0;
 272
 273                         mp->m_inotbt_nores = true;
 274
 275                         error = xfs_refcountbt_calc_reserves(mp, agno, &ask,
 276                                         &used);
 277                         if (error)
 278                                 goto out;
 279
 280                         error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
 281                                         ask, used);
 282                         if (error)
 283                                 goto out;
 284                 }
 285         }
 286
 287         /* Create the AGFL metadata reservation */
 288         if (pag->pag_agfl_resv.ar_asked == 0) {
 289                 ask = used = 0;
 290
 291                 error = xfs_rmapbt_calc_reserves(mp, agno, &ask, &used);
 292                 if (error)
 293                         goto out;
 294
 295                 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_AGFL, ask, used);
 296                 if (error)
 297                         goto out;
 298         }
 299
 300 #ifdef DEBUG
 301         /* need to read in the AGF for the ASSERT below to work */
 302         error = xfs_alloc_pagf_init(pag->pag_mount, NULL, pag->pag_agno, 0);
 303         if (error)
 304                 return error;
 305
 306         ASSERT(xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
 307                xfs_perag_resv(pag, XFS_AG_RESV_AGFL)->ar_reserved <=
 308                pag->pagf_freeblks + pag->pagf_flcount);
 309 #endif
 310 out:
 311         return error;
 312 }
 313
 314 /* Allocate a block from the reservation. */
 315 void
 316 xfs_ag_resv_alloc_extent(
 317         struct xfs_perag                *pag,
 318         enum xfs_ag_resv_type           type,
 319         struct xfs_alloc_arg            *args)
 320 {
 321         struct xfs_ag_resv              *resv;
 322         xfs_extlen_t                    len;
 323         uint                            field;
 324
 325         trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
 326
 327         switch (type) {
 328         case XFS_AG_RESV_METADATA:
 329         case XFS_AG_RESV_AGFL:
 330                 resv = xfs_perag_resv(pag, type);
 331                 break;
 332         default:
 333                 ASSERT(0);
 334                 /* fall through */
 335         case XFS_AG_RESV_NONE:
 336                 field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
 337                                        XFS_TRANS_SB_FDBLOCKS;
 338                 xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
 339                 return;
 340         }
 341
 342         len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
 343         resv->ar_reserved -= len;
 344         if (type == XFS_AG_RESV_AGFL)
 345                 return;
 346         /* Allocations of reserved blocks only need on-disk sb updates... */
 347         xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
 348         /* ...but non-reserved blocks need in-core and on-disk updates. */
 349         if (args->len > len)
 350                 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
 351                                 -((int64_t)args->len - len));
 352 }
 353
 354 /* Free a block to the reservation. */
 355 void
 356 xfs_ag_resv_free_extent(
 357         struct xfs_perag                *pag,
 358         enum xfs_ag_resv_type           type,
 359         struct xfs_trans                *tp,
 360         xfs_extlen_t                    len)
 361 {
 362         xfs_extlen_t                    leftover;
 363         struct xfs_ag_resv              *resv;
 364
 365         trace_xfs_ag_resv_free_extent(pag, type, len);
 366
 367         switch (type) {
 368         case XFS_AG_RESV_METADATA:
 369         case XFS_AG_RESV_AGFL:
 370                 resv = xfs_perag_resv(pag, type);
 371                 break;
 372         default:
 373                 ASSERT(0);
 374                 /* fall through */
 375         case XFS_AG_RESV_NONE:
 376                 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
 377                 return;
 378         }
 379
 380         leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
 381         resv->ar_reserved += leftover;
 382         if (type == XFS_AG_RESV_AGFL)
 383                 return;
 384         /* Freeing into the reserved pool only requires on-disk update... */
 385         xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
 386         /* ...but freeing beyond that requires in-core and on-disk update. */
 387         if (len > leftover)
 388                 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
 389 }