fs/xfs/libxfs/xfs_ag_resv.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Copyright (C) 2016 Oracle.  All Rights Reserved.
   4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
   5  */
   6 #include "xfs.h"
   7 #include "xfs_fs.h"
   8 #include "xfs_shared.h"
   9 #include "xfs_format.h"
  10 #include "xfs_log_format.h"
  11 #include "xfs_trans_resv.h"
  12 #include "xfs_mount.h"
  13 #include "xfs_alloc.h"
  14 #include "xfs_errortag.h"
  15 #include "xfs_error.h"
  16 #include "xfs_trace.h"
  17 #include "xfs_trans.h"
  18 #include "xfs_rmap_btree.h"
  19 #include "xfs_btree.h"
  20 #include "xfs_refcount_btree.h"
  21 #include "xfs_ialloc_btree.h"
  22 #include "xfs_sb.h"
  23 #include "xfs_ag_resv.h"
  24
  25 /*
  26  * Per-AG Block Reservations
  27  *
  28  * For some kinds of allocation group metadata structures, it is advantageous
  29  * to reserve a small number of blocks in each AG so that future expansions of
  30  * that data structure do not encounter ENOSPC because errors during a btree
  31  * split cause the filesystem to go offline.
  32  *
  33  * Prior to the introduction of reflink, this wasn't an issue because the free
  34  * space btrees maintain a reserve of space (the AGFL) to handle any expansion
  35  * that may be necessary; and allocations of other metadata (inodes, BMBT,
  36  * dir/attr) aren't restricted to a single AG.  However, with reflink it is
  37  * possible to allocate all the space in an AG, have subsequent reflink/CoW
  38  * activity expand the refcount btree, and discover that there's no space left
  39  * to handle that expansion.  Since we can calculate the maximum size of the
  40  * refcount btree, we can reserve space for it and avoid ENOSPC.
  41  *
  42  * Handling per-AG reservations consists of three changes to the allocator's
  43  * behavior:  First, because these reservations are always needed, we decrease
  44  * the ag_max_usable counter to reflect the size of the AG after the reserved
  45  * blocks are taken.  Second, the reservations must be reflected in the
  46  * fdblocks count to maintain proper accounting.  Third, each AG must maintain
  47  * its own reserved block counter so that we can calculate the amount of space
  48  * that must remain free to maintain the reservations.  Fourth, the "remaining
  49  * reserved blocks" count must be used when calculating the length of the
  50  * longest free extent in an AG and to clamp maxlen in the per-AG allocation
  51  * functions.  In other words, we maintain a virtual allocation via in-core
  52  * accounting tricks so that we don't have to clean up after a crash. :)
  53  *
  54  * Reserved blocks can be managed by passing one of the enum xfs_ag_resv_type
  55  * values via struct xfs_alloc_arg or directly to the xfs_free_extent
  56  * function.  It might seem a little funny to maintain a reservoir of blocks
  57  * to feed another reservoir, but the AGFL only holds enough blocks to get
  58  * through the next transaction.  The per-AG reservation is to ensure (we
  59  * hope) that each AG never runs out of blocks.  Each data structure wanting
  60  * to use the reservation system should update ask/used in xfs_ag_resv_init.
  61  */
  62
  63 /*
  64  * Are we critically low on blocks?  For now we'll define that as the number
  65  * of blocks we can get our hands on being less than 10% of what we reserved
  66  * or less than some arbitrary number (maximum btree height).
  67  */
  68 bool
  69 xfs_ag_resv_critical(
  70         struct xfs_perag                *pag,
  71         enum xfs_ag_resv_type           type)
  72 {
  73         xfs_extlen_t                    avail;
  74         xfs_extlen_t                    orig;
  75
  76         switch (type) {
  77         case XFS_AG_RESV_METADATA:
  78                 avail = pag->pagf_freeblks - pag->pag_rmapbt_resv.ar_reserved;
  79                 orig = pag->pag_meta_resv.ar_asked;
  80                 break;
  81         case XFS_AG_RESV_RMAPBT:
  82                 avail = pag->pagf_freeblks + pag->pagf_flcount -
  83                         pag->pag_meta_resv.ar_reserved;
  84                 orig = pag->pag_rmapbt_resv.ar_asked;
  85                 break;
  86         default:
  87                 ASSERT(0);
  88                 return false;
  89         }
  90
  91         trace_xfs_ag_resv_critical(pag, type, avail);
  92
  93         /* Critically low if less than 10% or max btree height remains. */
  94         return XFS_TEST_ERROR(avail < orig / 10 || avail < XFS_BTREE_MAXLEVELS,
  95                         pag->pag_mount, XFS_ERRTAG_AG_RESV_CRITICAL);
  96 }
  97
  98 /*
  99  * How many blocks are reserved but not used, and therefore must not be
 100  * allocated away?
 101  */
 102 xfs_extlen_t
 103 xfs_ag_resv_needed(
 104         struct xfs_perag                *pag,
 105         enum xfs_ag_resv_type           type)
 106 {
 107         xfs_extlen_t                    len;
 108
 109         len = pag->pag_meta_resv.ar_reserved + pag->pag_rmapbt_resv.ar_reserved;
 110         switch (type) {
 111         case XFS_AG_RESV_METADATA:
 112         case XFS_AG_RESV_RMAPBT:
 113                 len -= xfs_perag_resv(pag, type)->ar_reserved;
 114                 break;
 115         case XFS_AG_RESV_NONE:
 116                 /* empty */
 117                 break;
 118         default:
 119                 ASSERT(0);
 120         }
 121
 122         trace_xfs_ag_resv_needed(pag, type, len);
 123
 124         return len;
 125 }
 126
 127 /* Clean out a reservation */
 128 static int
 129 __xfs_ag_resv_free(
 130         struct xfs_perag                *pag,
 131         enum xfs_ag_resv_type           type)
 132 {
 133         struct xfs_ag_resv              *resv;
 134         xfs_extlen_t                    oldresv;
 135         int                             error;
 136
 137         trace_xfs_ag_resv_free(pag, type, 0);
 138
 139         resv = xfs_perag_resv(pag, type);
 140         if (pag->pag_agno == 0)
 141                 pag->pag_mount->m_ag_max_usable += resv->ar_asked;
 142         /*
 143          * RMAPBT blocks come from the AGFL and AGFL blocks are always
 144          * considered "free", so whatever was reserved at mount time must be
 145          * given back at umount.
 146          */
 147         if (type == XFS_AG_RESV_RMAPBT)
 148                 oldresv = resv->ar_orig_reserved;
 149         else
 150                 oldresv = resv->ar_reserved;
 151         error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
 152         resv->ar_reserved = 0;
 153         resv->ar_asked = 0;
 154         resv->ar_orig_reserved = 0;
 155
 156         if (error)
 157                 trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
 158                                 error, _RET_IP_);
 159         return error;
 160 }
 161
 162 /* Free a per-AG reservation. */
 163 int
 164 xfs_ag_resv_free(
 165         struct xfs_perag                *pag)
 166 {
 167         int                             error;
 168         int                             err2;
 169
 170         error = __xfs_ag_resv_free(pag, XFS_AG_RESV_RMAPBT);
 171         err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
 172         if (err2 && !error)
 173                 error = err2;
 174         return error;
 175 }
 176
 177 static int
 178 __xfs_ag_resv_init(
 179         struct xfs_perag                *pag,
 180         enum xfs_ag_resv_type           type,
 181         xfs_extlen_t                    ask,
 182         xfs_extlen_t                    used)
 183 {
 184         struct xfs_mount                *mp = pag->pag_mount;
 185         struct xfs_ag_resv              *resv;
 186         int                             error;
 187         xfs_extlen_t                    hidden_space;
 188
 189         if (used > ask)
 190                 ask = used;
 191
 192         switch (type) {
 193         case XFS_AG_RESV_RMAPBT:
 194                 /*
 195                  * Space taken by the rmapbt is not subtracted from fdblocks
 196                  * because the rmapbt lives in the free space.  Here we must
 197                  * subtract the entire reservation from fdblocks so that we
 198                  * always have blocks available for rmapbt expansion.
 199                  */
 200                 hidden_space = ask;
 201                 break;
 202         case XFS_AG_RESV_METADATA:
 203                 /*
 204                  * Space taken by all other metadata btrees are accounted
 205                  * on-disk as used space.  We therefore only hide the space
 206                  * that is reserved but not used by the trees.
 207                  */
 208                 hidden_space = ask - used;
 209                 break;
 210         default:
 211                 ASSERT(0);
 212                 return -EINVAL;
 213         }
 214         error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space, true);
 215         if (error) {
 216                 trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
 217                                 error, _RET_IP_);
 218                 xfs_warn(mp,
 219 "Per-AG reservation for AG %u failed.  Filesystem may run out of space.",
 220                                 pag->pag_agno);
 221                 return error;
 222         }
 223
 224         /*
 225          * Reduce the maximum per-AG allocation length by however much we're
 226          * trying to reserve for an AG.  Since this is a filesystem-wide
 227          * counter, we only make the adjustment for AG 0.  This assumes that
 228          * there aren't any AGs hungrier for per-AG reservation than AG 0.
 229          */
 230         if (pag->pag_agno == 0)
 231                 mp->m_ag_max_usable -= ask;
 232
 233         resv = xfs_perag_resv(pag, type);
 234         resv->ar_asked = ask;
 235         resv->ar_orig_reserved = hidden_space;
 236         resv->ar_reserved = ask - used;
 237
 238         trace_xfs_ag_resv_init(pag, type, ask);
 239         return 0;
 240 }
 241
 242 /* Create a per-AG block reservation. */
 243 int
 244 xfs_ag_resv_init(
 245         struct xfs_perag                *pag,
 246         struct xfs_trans                *tp)
 247 {
 248         struct xfs_mount                *mp = pag->pag_mount;
 249         xfs_agnumber_t                  agno = pag->pag_agno;
 250         xfs_extlen_t                    ask;
 251         xfs_extlen_t                    used;
 252         int                             error = 0;
 253
 254         /* Create the metadata reservation. */
 255         if (pag->pag_meta_resv.ar_asked == 0) {
 256                 ask = used = 0;
 257
 258                 error = xfs_refcountbt_calc_reserves(mp, tp, agno, &ask, &used);
 259                 if (error)
 260                         goto out;
 261
 262                 error = xfs_finobt_calc_reserves(mp, tp, agno, &ask, &used);
 263                 if (error)
 264                         goto out;
 265
 266                 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
 267                                 ask, used);
 268                 if (error) {
 269                         /*
 270                          * Because we didn't have per-AG reservations when the
 271                          * finobt feature was added we might not be able to
 272                          * reserve all needed blocks.  Warn and fall back to the
 273                          * old and potentially buggy code in that case, but
 274                          * ensure we do have the reservation for the refcountbt.
 275                          */
 276                         ask = used = 0;
 277
 278                         mp->m_finobt_nores = true;
 279
 280                         error = xfs_refcountbt_calc_reserves(mp, tp, agno, &ask,
 281                                         &used);
 282                         if (error)
 283                                 goto out;
 284
 285                         error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
 286                                         ask, used);
 287                         if (error)
 288                                 goto out;
 289                 }
 290         }
 291
 292         /* Create the RMAPBT metadata reservation */
 293         if (pag->pag_rmapbt_resv.ar_asked == 0) {
 294                 ask = used = 0;
 295
 296                 error = xfs_rmapbt_calc_reserves(mp, tp, agno, &ask, &used);
 297                 if (error)
 298                         goto out;
 299
 300                 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_RMAPBT, ask, used);
 301                 if (error)
 302                         goto out;
 303         }
 304
 305 #ifdef DEBUG
 306         /* need to read in the AGF for the ASSERT below to work */
 307         error = xfs_alloc_pagf_init(pag->pag_mount, tp, pag->pag_agno, 0);
 308         if (error)
 309                 return error;
 310
 311         ASSERT(xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
 312                xfs_perag_resv(pag, XFS_AG_RESV_RMAPBT)->ar_reserved <=
 313                pag->pagf_freeblks + pag->pagf_flcount);
 314 #endif
 315 out:
 316         return error;
 317 }
 318
 319 /* Allocate a block from the reservation. */
 320 void
 321 xfs_ag_resv_alloc_extent(
 322         struct xfs_perag                *pag,
 323         enum xfs_ag_resv_type           type,
 324         struct xfs_alloc_arg            *args)
 325 {
 326         struct xfs_ag_resv              *resv;
 327         xfs_extlen_t                    len;
 328         uint                            field;
 329
 330         trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
 331
 332         switch (type) {
 333         case XFS_AG_RESV_AGFL:
 334                 return;
 335         case XFS_AG_RESV_METADATA:
 336         case XFS_AG_RESV_RMAPBT:
 337                 resv = xfs_perag_resv(pag, type);
 338                 break;
 339         default:
 340                 ASSERT(0);
 341                 /* fall through */
 342         case XFS_AG_RESV_NONE:
 343                 field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
 344                                        XFS_TRANS_SB_FDBLOCKS;
 345                 xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
 346                 return;
 347         }
 348
 349         len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
 350         resv->ar_reserved -= len;
 351         if (type == XFS_AG_RESV_RMAPBT)
 352                 return;
 353         /* Allocations of reserved blocks only need on-disk sb updates... */
 354         xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
 355         /* ...but non-reserved blocks need in-core and on-disk updates. */
 356         if (args->len > len)
 357                 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
 358                                 -((int64_t)args->len - len));
 359 }
 360
 361 /* Free a block to the reservation. */
 362 void
 363 xfs_ag_resv_free_extent(
 364         struct xfs_perag                *pag,
 365         enum xfs_ag_resv_type           type,
 366         struct xfs_trans                *tp,
 367         xfs_extlen_t                    len)
 368 {
 369         xfs_extlen_t                    leftover;
 370         struct xfs_ag_resv              *resv;
 371
 372         trace_xfs_ag_resv_free_extent(pag, type, len);
 373
 374         switch (type) {
 375         case XFS_AG_RESV_AGFL:
 376                 return;
 377         case XFS_AG_RESV_METADATA:
 378         case XFS_AG_RESV_RMAPBT:
 379                 resv = xfs_perag_resv(pag, type);
 380                 break;
 381         default:
 382                 ASSERT(0);
 383                 /* fall through */
 384         case XFS_AG_RESV_NONE:
 385                 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
 386                 return;
 387         }
 388
 389         leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
 390         resv->ar_reserved += leftover;
 391         if (type == XFS_AG_RESV_RMAPBT)
 392                 return;
 393         /* Freeing into the reserved pool only requires on-disk update... */
 394         xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
 395         /* ...but freeing beyond that requires in-core and on-disk update. */
 396         if (len > leftover)
 397                 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
 398 }