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