2 * Copyright (C) 2016 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
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.
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.
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.
22 #include "xfs_shared.h"
23 #include "xfs_format.h"
24 #include "xfs_log_format.h"
25 #include "xfs_trans_resv.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"
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"
45 * Per-AG Block Reservations
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.
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.
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. :)
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.
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).
89 struct xfs_perag *pag,
90 enum xfs_ag_resv_type 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;
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;
110 trace_xfs_ag_resv_critical(pag, type, avail);
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);
118 * How many blocks are reserved but not used, and therefore must not be
123 struct xfs_perag *pag,
124 enum xfs_ag_resv_type type)
128 len = pag->pag_meta_resv.ar_reserved + pag->pag_agfl_resv.ar_reserved;
130 case XFS_AG_RESV_METADATA:
131 case XFS_AG_RESV_AGFL:
132 len -= xfs_perag_resv(pag, type)->ar_reserved;
134 case XFS_AG_RESV_NONE:
141 trace_xfs_ag_resv_needed(pag, type, len);
146 /* Clean out a reservation */
149 struct xfs_perag *pag,
150 enum xfs_ag_resv_type type)
152 struct xfs_ag_resv *resv;
153 xfs_extlen_t oldresv;
156 trace_xfs_ag_resv_free(pag, type, 0);
158 resv = xfs_perag_resv(pag, type);
159 pag->pag_mount->m_ag_max_usable += resv->ar_asked;
161 * AGFL blocks are always considered "free", so whatever
162 * was reserved at mount time must be given back at umount.
164 if (type == XFS_AG_RESV_AGFL)
165 oldresv = resv->ar_orig_reserved;
167 oldresv = resv->ar_reserved;
168 error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
169 resv->ar_reserved = 0;
173 trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
178 /* Free a per-AG reservation. */
181 struct xfs_perag *pag)
186 error = __xfs_ag_resv_free(pag, XFS_AG_RESV_AGFL);
187 err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
195 struct xfs_perag *pag,
196 enum xfs_ag_resv_type type,
200 struct xfs_mount *mp = pag->pag_mount;
201 struct xfs_ag_resv *resv;
203 xfs_extlen_t reserved;
207 reserved = ask - used;
209 error = xfs_mod_fdblocks(mp, -(int64_t)reserved, true);
211 trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
214 "Per-AG reservation for AG %u failed. Filesystem may run out of space.",
219 mp->m_ag_max_usable -= ask;
221 resv = xfs_perag_resv(pag, type);
222 resv->ar_asked = ask;
223 resv->ar_reserved = resv->ar_orig_reserved = reserved;
225 trace_xfs_ag_resv_init(pag, type, ask);
229 /* Create a per-AG block reservation. */
232 struct xfs_perag *pag)
234 struct xfs_mount *mp = pag->pag_mount;
235 xfs_agnumber_t agno = pag->pag_agno;
240 /* Create the metadata reservation. */
241 if (pag->pag_meta_resv.ar_asked == 0) {
244 error = xfs_refcountbt_calc_reserves(mp, agno, &ask, &used);
248 error = xfs_finobt_calc_reserves(mp, agno, &ask, &used);
252 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
256 * Because we didn't have per-AG reservations when the
257 * finobt feature was added we might not be able to
258 * reserve all needed blocks. Warn and fall back to the
259 * old and potentially buggy code in that case, but
260 * ensure we do have the reservation for the refcountbt.
264 mp->m_inotbt_nores = true;
266 error = xfs_refcountbt_calc_reserves(mp, agno, &ask,
271 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
278 /* Create the AGFL metadata reservation */
279 if (pag->pag_agfl_resv.ar_asked == 0) {
282 error = xfs_rmapbt_calc_reserves(mp, agno, &ask, &used);
286 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_AGFL, ask, used);
292 /* need to read in the AGF for the ASSERT below to work */
293 error = xfs_alloc_pagf_init(pag->pag_mount, NULL, pag->pag_agno, 0);
297 ASSERT(xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
298 xfs_perag_resv(pag, XFS_AG_RESV_AGFL)->ar_reserved <=
299 pag->pagf_freeblks + pag->pagf_flcount);
305 /* Allocate a block from the reservation. */
307 xfs_ag_resv_alloc_extent(
308 struct xfs_perag *pag,
309 enum xfs_ag_resv_type type,
310 struct xfs_alloc_arg *args)
312 struct xfs_ag_resv *resv;
316 trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
319 case XFS_AG_RESV_METADATA:
320 case XFS_AG_RESV_AGFL:
321 resv = xfs_perag_resv(pag, type);
326 case XFS_AG_RESV_NONE:
327 field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
328 XFS_TRANS_SB_FDBLOCKS;
329 xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
333 len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
334 resv->ar_reserved -= len;
335 if (type == XFS_AG_RESV_AGFL)
337 /* Allocations of reserved blocks only need on-disk sb updates... */
338 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
339 /* ...but non-reserved blocks need in-core and on-disk updates. */
341 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
342 -((int64_t)args->len - len));
345 /* Free a block to the reservation. */
347 xfs_ag_resv_free_extent(
348 struct xfs_perag *pag,
349 enum xfs_ag_resv_type type,
350 struct xfs_trans *tp,
353 xfs_extlen_t leftover;
354 struct xfs_ag_resv *resv;
356 trace_xfs_ag_resv_free_extent(pag, type, len);
359 case XFS_AG_RESV_METADATA:
360 case XFS_AG_RESV_AGFL:
361 resv = xfs_perag_resv(pag, type);
366 case XFS_AG_RESV_NONE:
367 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
371 leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
372 resv->ar_reserved += leftover;
373 if (type == XFS_AG_RESV_AGFL)
375 /* Freeing into the reserved pool only requires on-disk update... */
376 xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
377 /* ...but freeing beyond that requires in-core and on-disk update. */
379 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);