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.
20 #include "libxfs_priv.h"
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_trace.h"
31 #include "xfs_cksum.h"
32 #include "xfs_trans.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"
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
);
117 * How many blocks are reserved but not used, and therefore must not be
122 struct xfs_perag
*pag
,
123 enum xfs_ag_resv_type type
)
127 len
= pag
->pag_meta_resv
.ar_reserved
+ pag
->pag_agfl_resv
.ar_reserved
;
129 case XFS_AG_RESV_METADATA
:
130 case XFS_AG_RESV_AGFL
:
131 len
-= xfs_perag_resv(pag
, type
)->ar_reserved
;
133 case XFS_AG_RESV_NONE
:
140 trace_xfs_ag_resv_needed(pag
, type
, len
);
145 /* Clean out a reservation */
148 struct xfs_perag
*pag
,
149 enum xfs_ag_resv_type type
)
151 struct xfs_ag_resv
*resv
;
152 xfs_extlen_t oldresv
;
155 trace_xfs_ag_resv_free(pag
, type
, 0);
157 resv
= xfs_perag_resv(pag
, type
);
158 pag
->pag_mount
->m_ag_max_usable
+= resv
->ar_asked
;
160 * AGFL blocks are always considered "free", so whatever
161 * was reserved at mount time must be given back at umount.
163 if (type
== XFS_AG_RESV_AGFL
)
164 oldresv
= resv
->ar_orig_reserved
;
166 oldresv
= resv
->ar_reserved
;
167 error
= xfs_mod_fdblocks(pag
->pag_mount
, oldresv
, true);
168 resv
->ar_reserved
= 0;
172 trace_xfs_ag_resv_free_error(pag
->pag_mount
, pag
->pag_agno
,
177 /* Free a per-AG reservation. */
180 struct xfs_perag
*pag
)
185 error
= __xfs_ag_resv_free(pag
, XFS_AG_RESV_AGFL
);
186 err2
= __xfs_ag_resv_free(pag
, XFS_AG_RESV_METADATA
);
194 struct xfs_perag
*pag
,
195 enum xfs_ag_resv_type type
,
199 struct xfs_mount
*mp
= pag
->pag_mount
;
200 struct xfs_ag_resv
*resv
;
202 xfs_extlen_t reserved
;
206 reserved
= ask
- used
;
208 error
= xfs_mod_fdblocks(mp
, -(int64_t)reserved
, true);
210 trace_xfs_ag_resv_init_error(pag
->pag_mount
, pag
->pag_agno
,
213 "Per-AG reservation for AG %u failed. Filesystem may run out of space.",
218 mp
->m_ag_max_usable
-= ask
;
220 resv
= xfs_perag_resv(pag
, type
);
221 resv
->ar_asked
= ask
;
222 resv
->ar_reserved
= resv
->ar_orig_reserved
= reserved
;
224 trace_xfs_ag_resv_init(pag
, type
, ask
);
228 /* Create a per-AG block reservation. */
231 struct xfs_perag
*pag
)
233 struct xfs_mount
*mp
= pag
->pag_mount
;
234 xfs_agnumber_t agno
= pag
->pag_agno
;
239 /* Create the metadata reservation. */
240 if (pag
->pag_meta_resv
.ar_asked
== 0) {
243 error
= xfs_refcountbt_calc_reserves(mp
, agno
, &ask
, &used
);
247 error
= xfs_finobt_calc_reserves(mp
, agno
, &ask
, &used
);
251 error
= __xfs_ag_resv_init(pag
, XFS_AG_RESV_METADATA
,
255 * Because we didn't have per-AG reservations when the
256 * finobt feature was added we might not be able to
257 * reserve all needed blocks. Warn and fall back to the
258 * old and potentially buggy code in that case, but
259 * ensure we do have the reservation for the refcountbt.
263 mp
->m_inotbt_nores
= true;
265 error
= xfs_refcountbt_calc_reserves(mp
, agno
, &ask
,
270 error
= __xfs_ag_resv_init(pag
, XFS_AG_RESV_METADATA
,
277 /* Create the AGFL metadata reservation */
278 if (pag
->pag_agfl_resv
.ar_asked
== 0) {
281 error
= xfs_rmapbt_calc_reserves(mp
, agno
, &ask
, &used
);
285 error
= __xfs_ag_resv_init(pag
, XFS_AG_RESV_AGFL
, ask
, used
);
291 /* need to read in the AGF for the ASSERT below to work */
292 error
= xfs_alloc_pagf_init(pag
->pag_mount
, NULL
, pag
->pag_agno
, 0);
296 ASSERT(xfs_perag_resv(pag
, XFS_AG_RESV_METADATA
)->ar_reserved
+
297 xfs_perag_resv(pag
, XFS_AG_RESV_AGFL
)->ar_reserved
<=
298 pag
->pagf_freeblks
+ pag
->pagf_flcount
);
304 /* Allocate a block from the reservation. */
306 xfs_ag_resv_alloc_extent(
307 struct xfs_perag
*pag
,
308 enum xfs_ag_resv_type type
,
309 struct xfs_alloc_arg
*args
)
311 struct xfs_ag_resv
*resv
;
315 trace_xfs_ag_resv_alloc_extent(pag
, type
, args
->len
);
318 case XFS_AG_RESV_METADATA
:
319 case XFS_AG_RESV_AGFL
:
320 resv
= xfs_perag_resv(pag
, type
);
325 case XFS_AG_RESV_NONE
:
326 field
= args
->wasdel
? XFS_TRANS_SB_RES_FDBLOCKS
:
327 XFS_TRANS_SB_FDBLOCKS
;
328 xfs_trans_mod_sb(args
->tp
, field
, -(int64_t)args
->len
);
332 len
= min_t(xfs_extlen_t
, args
->len
, resv
->ar_reserved
);
333 resv
->ar_reserved
-= len
;
334 if (type
== XFS_AG_RESV_AGFL
)
336 /* Allocations of reserved blocks only need on-disk sb updates... */
337 xfs_trans_mod_sb(args
->tp
, XFS_TRANS_SB_RES_FDBLOCKS
, -(int64_t)len
);
338 /* ...but non-reserved blocks need in-core and on-disk updates. */
340 xfs_trans_mod_sb(args
->tp
, XFS_TRANS_SB_FDBLOCKS
,
341 -((int64_t)args
->len
- len
));
344 /* Free a block to the reservation. */
346 xfs_ag_resv_free_extent(
347 struct xfs_perag
*pag
,
348 enum xfs_ag_resv_type type
,
349 struct xfs_trans
*tp
,
352 xfs_extlen_t leftover
;
353 struct xfs_ag_resv
*resv
;
355 trace_xfs_ag_resv_free_extent(pag
, type
, len
);
358 case XFS_AG_RESV_METADATA
:
359 case XFS_AG_RESV_AGFL
:
360 resv
= xfs_perag_resv(pag
, type
);
365 case XFS_AG_RESV_NONE
:
366 xfs_trans_mod_sb(tp
, XFS_TRANS_SB_FDBLOCKS
, (int64_t)len
);
370 leftover
= min_t(xfs_extlen_t
, len
, resv
->ar_asked
- resv
->ar_reserved
);
371 resv
->ar_reserved
+= leftover
;
372 if (type
== XFS_AG_RESV_AGFL
)
374 /* Freeing into the reserved pool only requires on-disk update... */
375 xfs_trans_mod_sb(tp
, XFS_TRANS_SB_RES_FDBLOCKS
, len
);
376 /* ...but freeing beyond that requires in-core and on-disk update. */
378 xfs_trans_mod_sb(tp
, XFS_TRANS_SB_FDBLOCKS
, len
- leftover
);