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"
44 * Per-AG Block Reservations
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.
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.
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. :)
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.
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).
88 struct xfs_perag *pag,
89 enum xfs_ag_resv_type 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;
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;
109 trace_xfs_ag_resv_critical(pag, type, avail);
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 XFS_RANDOM_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;
204 resv = xfs_perag_resv(pag, type);
207 resv->ar_asked = ask;
208 resv->ar_reserved = resv->ar_orig_reserved = ask - used;
209 mp->m_ag_max_usable -= ask;
211 trace_xfs_ag_resv_init(pag, type, ask);
213 error = xfs_mod_fdblocks(mp, -(int64_t)resv->ar_reserved, true);
215 trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
221 /* Create a per-AG block reservation. */
224 struct xfs_perag *pag)
230 /* Create the metadata reservation. */
231 if (pag->pag_meta_resv.ar_asked == 0) {
234 error = xfs_refcountbt_calc_reserves(pag->pag_mount,
235 pag->pag_agno, &ask, &used);
239 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
245 /* Create the AGFL metadata reservation */
246 if (pag->pag_agfl_resv.ar_asked == 0) {
249 error = xfs_rmapbt_calc_reserves(pag->pag_mount, pag->pag_agno,
254 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_AGFL, ask, used);
263 /* Allocate a block from the reservation. */
265 xfs_ag_resv_alloc_extent(
266 struct xfs_perag *pag,
267 enum xfs_ag_resv_type type,
268 struct xfs_alloc_arg *args)
270 struct xfs_ag_resv *resv;
274 trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
277 case XFS_AG_RESV_METADATA:
278 case XFS_AG_RESV_AGFL:
279 resv = xfs_perag_resv(pag, type);
284 case XFS_AG_RESV_NONE:
285 field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
286 XFS_TRANS_SB_FDBLOCKS;
287 xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
291 len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
292 resv->ar_reserved -= len;
293 if (type == XFS_AG_RESV_AGFL)
295 /* Allocations of reserved blocks only need on-disk sb updates... */
296 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
297 /* ...but non-reserved blocks need in-core and on-disk updates. */
299 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
300 -((int64_t)args->len - len));
303 /* Free a block to the reservation. */
305 xfs_ag_resv_free_extent(
306 struct xfs_perag *pag,
307 enum xfs_ag_resv_type type,
308 struct xfs_trans *tp,
311 xfs_extlen_t leftover;
312 struct xfs_ag_resv *resv;
314 trace_xfs_ag_resv_free_extent(pag, type, len);
317 case XFS_AG_RESV_METADATA:
318 case XFS_AG_RESV_AGFL:
319 resv = xfs_perag_resv(pag, type);
324 case XFS_AG_RESV_NONE:
325 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
329 leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
330 resv->ar_reserved += leftover;
331 if (type == XFS_AG_RESV_AGFL)
333 /* Freeing into the reserved pool only requires on-disk update... */
334 xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
335 /* ...but freeing beyond that requires in-core and on-disk update. */
337 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);