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"
43 * Per-AG Block Reservations
45 * For some kinds of allocation group metadata structures, it is advantageous
46 * to reserve a small number of blocks in each AG so that future expansions of
47 * that data structure do not encounter ENOSPC because errors during a btree
48 * split cause the filesystem to go offline.
50 * Prior to the introduction of reflink, this wasn't an issue because the free
51 * space btrees maintain a reserve of space (the AGFL) to handle any expansion
52 * that may be necessary; and allocations of other metadata (inodes, BMBT,
53 * dir/attr) aren't restricted to a single AG. However, with reflink it is
54 * possible to allocate all the space in an AG, have subsequent reflink/CoW
55 * activity expand the refcount btree, and discover that there's no space left
56 * to handle that expansion. Since we can calculate the maximum size of the
57 * refcount btree, we can reserve space for it and avoid ENOSPC.
59 * Handling per-AG reservations consists of three changes to the allocator's
60 * behavior: First, because these reservations are always needed, we decrease
61 * the ag_max_usable counter to reflect the size of the AG after the reserved
62 * blocks are taken. Second, the reservations must be reflected in the
63 * fdblocks count to maintain proper accounting. Third, each AG must maintain
64 * its own reserved block counter so that we can calculate the amount of space
65 * that must remain free to maintain the reservations. Fourth, the "remaining
66 * reserved blocks" count must be used when calculating the length of the
67 * longest free extent in an AG and to clamp maxlen in the per-AG allocation
68 * functions. In other words, we maintain a virtual allocation via in-core
69 * accounting tricks so that we don't have to clean up after a crash. :)
71 * Reserved blocks can be managed by passing one of the enum xfs_ag_resv_type
72 * values via struct xfs_alloc_arg or directly to the xfs_free_extent
73 * function. It might seem a little funny to maintain a reservoir of blocks
74 * to feed another reservoir, but the AGFL only holds enough blocks to get
75 * through the next transaction. The per-AG reservation is to ensure (we
76 * hope) that each AG never runs out of blocks. Each data structure wanting
77 * to use the reservation system should update ask/used in xfs_ag_resv_init.
81 * Are we critically low on blocks? For now we'll define that as the number
82 * of blocks we can get our hands on being less than 10% of what we reserved
83 * or less than some arbitrary number (maximum btree height).
87 struct xfs_perag *pag,
88 enum xfs_ag_resv_type type)
94 case XFS_AG_RESV_METADATA:
95 avail = pag->pagf_freeblks - pag->pag_agfl_resv.ar_reserved;
96 orig = pag->pag_meta_resv.ar_asked;
98 case XFS_AG_RESV_AGFL:
99 avail = pag->pagf_freeblks + pag->pagf_flcount -
100 pag->pag_meta_resv.ar_reserved;
101 orig = pag->pag_agfl_resv.ar_asked;
108 trace_xfs_ag_resv_critical(pag, type, avail);
110 /* Critically low if less than 10% or max btree height remains. */
111 return avail < orig / 10 || avail < XFS_BTREE_MAXLEVELS;
115 * How many blocks are reserved but not used, and therefore must not be
120 struct xfs_perag *pag,
121 enum xfs_ag_resv_type type)
125 len = pag->pag_meta_resv.ar_reserved + pag->pag_agfl_resv.ar_reserved;
127 case XFS_AG_RESV_METADATA:
128 case XFS_AG_RESV_AGFL:
129 len -= xfs_perag_resv(pag, type)->ar_reserved;
131 case XFS_AG_RESV_NONE:
138 trace_xfs_ag_resv_needed(pag, type, len);
143 /* Clean out a reservation */
146 struct xfs_perag *pag,
147 enum xfs_ag_resv_type type)
149 struct xfs_ag_resv *resv;
150 xfs_extlen_t oldresv;
153 trace_xfs_ag_resv_free(pag, type, 0);
155 resv = xfs_perag_resv(pag, type);
156 pag->pag_mount->m_ag_max_usable += resv->ar_asked;
158 * AGFL blocks are always considered "free", so whatever
159 * was reserved at mount time must be given back at umount.
161 if (type == XFS_AG_RESV_AGFL)
162 oldresv = resv->ar_orig_reserved;
164 oldresv = resv->ar_reserved;
165 error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
166 resv->ar_reserved = 0;
170 trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
175 /* Free a per-AG reservation. */
178 struct xfs_perag *pag)
183 error = __xfs_ag_resv_free(pag, XFS_AG_RESV_AGFL);
184 err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
192 struct xfs_perag *pag,
193 enum xfs_ag_resv_type type,
197 struct xfs_mount *mp = pag->pag_mount;
198 struct xfs_ag_resv *resv;
201 resv = xfs_perag_resv(pag, type);
204 resv->ar_asked = ask;
205 resv->ar_reserved = resv->ar_orig_reserved = ask - used;
206 mp->m_ag_max_usable -= ask;
208 trace_xfs_ag_resv_init(pag, type, ask);
210 error = xfs_mod_fdblocks(mp, -(int64_t)resv->ar_reserved, true);
212 trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
218 /* Create a per-AG block reservation. */
221 struct xfs_perag *pag)
227 /* Create the metadata reservation. */
228 if (pag->pag_meta_resv.ar_asked == 0) {
231 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
237 /* Create the AGFL metadata reservation */
238 if (pag->pag_agfl_resv.ar_asked == 0) {
241 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_AGFL, ask, used);
250 /* Allocate a block from the reservation. */
252 xfs_ag_resv_alloc_extent(
253 struct xfs_perag *pag,
254 enum xfs_ag_resv_type type,
255 struct xfs_alloc_arg *args)
257 struct xfs_ag_resv *resv;
261 trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
264 case XFS_AG_RESV_METADATA:
265 case XFS_AG_RESV_AGFL:
266 resv = xfs_perag_resv(pag, type);
271 case XFS_AG_RESV_NONE:
272 field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
273 XFS_TRANS_SB_FDBLOCKS;
274 xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
278 len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
279 resv->ar_reserved -= len;
280 if (type == XFS_AG_RESV_AGFL)
282 /* Allocations of reserved blocks only need on-disk sb updates... */
283 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
284 /* ...but non-reserved blocks need in-core and on-disk updates. */
286 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
287 -((int64_t)args->len - len));
290 /* Free a block to the reservation. */
292 xfs_ag_resv_free_extent(
293 struct xfs_perag *pag,
294 enum xfs_ag_resv_type type,
295 struct xfs_trans *tp,
298 xfs_extlen_t leftover;
299 struct xfs_ag_resv *resv;
301 trace_xfs_ag_resv_free_extent(pag, type, len);
304 case XFS_AG_RESV_METADATA:
305 case XFS_AG_RESV_AGFL:
306 resv = xfs_perag_resv(pag, type);
311 case XFS_AG_RESV_NONE:
312 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
316 leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
317 resv->ar_reserved += leftover;
318 if (type == XFS_AG_RESV_AGFL)
320 /* Freeing into the reserved pool only requires on-disk update... */
321 xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
322 /* ...but freeing beyond that requires in-core and on-disk update. */
324 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);