xref: /openbmc/linux/fs/xfs/libxfs/xfs_ag_resv.c (revision a977d045)
1 /*
2  * Copyright (C) 2016 Oracle.  All Rights Reserved.
3  *
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
5  *
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.
10  *
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.
15  *
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.
19  */
20 #include "xfs.h"
21 #include "xfs_fs.h"
22 #include "xfs_shared.h"
23 #include "xfs_format.h"
24 #include "xfs_log_format.h"
25 #include "xfs_trans_resv.h"
26 #include "xfs_sb.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"
34 #include "xfs_bit.h"
35 #include "xfs_bmap.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"
43 
44 /*
45  * Per-AG Block Reservations
46  *
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.
51  *
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.
60  *
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. :)
72  *
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.
80  */
81 
82 /*
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).
86  */
87 bool
88 xfs_ag_resv_critical(
89 	struct xfs_perag		*pag,
90 	enum xfs_ag_resv_type		type)
91 {
92 	xfs_extlen_t			avail;
93 	xfs_extlen_t			orig;
94 
95 	switch (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;
99 		break;
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;
104 		break;
105 	default:
106 		ASSERT(0);
107 		return false;
108 	}
109 
110 	trace_xfs_ag_resv_critical(pag, type, avail);
111 
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);
115 }
116 
117 /*
118  * How many blocks are reserved but not used, and therefore must not be
119  * allocated away?
120  */
121 xfs_extlen_t
122 xfs_ag_resv_needed(
123 	struct xfs_perag		*pag,
124 	enum xfs_ag_resv_type		type)
125 {
126 	xfs_extlen_t			len;
127 
128 	len = pag->pag_meta_resv.ar_reserved + pag->pag_agfl_resv.ar_reserved;
129 	switch (type) {
130 	case XFS_AG_RESV_METADATA:
131 	case XFS_AG_RESV_AGFL:
132 		len -= xfs_perag_resv(pag, type)->ar_reserved;
133 		break;
134 	case XFS_AG_RESV_NONE:
135 		/* empty */
136 		break;
137 	default:
138 		ASSERT(0);
139 	}
140 
141 	trace_xfs_ag_resv_needed(pag, type, len);
142 
143 	return len;
144 }
145 
146 /* Clean out a reservation */
147 static int
148 __xfs_ag_resv_free(
149 	struct xfs_perag		*pag,
150 	enum xfs_ag_resv_type		type)
151 {
152 	struct xfs_ag_resv		*resv;
153 	xfs_extlen_t			oldresv;
154 	int				error;
155 
156 	trace_xfs_ag_resv_free(pag, type, 0);
157 
158 	resv = xfs_perag_resv(pag, type);
159 	pag->pag_mount->m_ag_max_usable += resv->ar_asked;
160 	/*
161 	 * AGFL blocks are always considered "free", so whatever
162 	 * was reserved at mount time must be given back at umount.
163 	 */
164 	if (type == XFS_AG_RESV_AGFL)
165 		oldresv = resv->ar_orig_reserved;
166 	else
167 		oldresv = resv->ar_reserved;
168 	error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
169 	resv->ar_reserved = 0;
170 	resv->ar_asked = 0;
171 
172 	if (error)
173 		trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
174 				error, _RET_IP_);
175 	return error;
176 }
177 
178 /* Free a per-AG reservation. */
179 int
180 xfs_ag_resv_free(
181 	struct xfs_perag		*pag)
182 {
183 	int				error;
184 	int				err2;
185 
186 	error = __xfs_ag_resv_free(pag, XFS_AG_RESV_AGFL);
187 	err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
188 	if (err2 && !error)
189 		error = err2;
190 	return error;
191 }
192 
193 static int
194 __xfs_ag_resv_init(
195 	struct xfs_perag		*pag,
196 	enum xfs_ag_resv_type		type,
197 	xfs_extlen_t			ask,
198 	xfs_extlen_t			used)
199 {
200 	struct xfs_mount		*mp = pag->pag_mount;
201 	struct xfs_ag_resv		*resv;
202 	int				error;
203 	xfs_extlen_t			reserved;
204 
205 	if (used > ask)
206 		ask = used;
207 	reserved = ask - used;
208 
209 	error = xfs_mod_fdblocks(mp, -(int64_t)reserved, true);
210 	if (error) {
211 		trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
212 				error, _RET_IP_);
213 		xfs_warn(mp,
214 "Per-AG reservation for AG %u failed.  Filesystem may run out of space.",
215 				pag->pag_agno);
216 		return error;
217 	}
218 
219 	mp->m_ag_max_usable -= ask;
220 
221 	resv = xfs_perag_resv(pag, type);
222 	resv->ar_asked = ask;
223 	resv->ar_reserved = resv->ar_orig_reserved = reserved;
224 
225 	trace_xfs_ag_resv_init(pag, type, ask);
226 	return 0;
227 }
228 
229 /* Create a per-AG block reservation. */
230 int
231 xfs_ag_resv_init(
232 	struct xfs_perag		*pag)
233 {
234 	struct xfs_mount		*mp = pag->pag_mount;
235 	xfs_agnumber_t			agno = pag->pag_agno;
236 	xfs_extlen_t			ask;
237 	xfs_extlen_t			used;
238 	int				error = 0;
239 
240 	/* Create the metadata reservation. */
241 	if (pag->pag_meta_resv.ar_asked == 0) {
242 		ask = used = 0;
243 
244 		error = xfs_refcountbt_calc_reserves(mp, agno, &ask, &used);
245 		if (error)
246 			goto out;
247 
248 		error = xfs_finobt_calc_reserves(mp, agno, &ask, &used);
249 		if (error)
250 			goto out;
251 
252 		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
253 				ask, used);
254 		if (error) {
255 			/*
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.
261 			 */
262 			ask = used = 0;
263 
264 			mp->m_inotbt_nores = true;
265 
266 			error = xfs_refcountbt_calc_reserves(mp, agno, &ask,
267 					&used);
268 			if (error)
269 				goto out;
270 
271 			error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
272 					ask, used);
273 			if (error)
274 				goto out;
275 		}
276 	}
277 
278 	/* Create the AGFL metadata reservation */
279 	if (pag->pag_agfl_resv.ar_asked == 0) {
280 		ask = used = 0;
281 
282 		error = xfs_rmapbt_calc_reserves(mp, agno, &ask, &used);
283 		if (error)
284 			goto out;
285 
286 		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_AGFL, ask, used);
287 		if (error)
288 			goto out;
289 	}
290 
291 #ifdef DEBUG
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);
294 	if (error)
295 		return error;
296 
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);
300 #endif
301 out:
302 	return error;
303 }
304 
305 /* Allocate a block from the reservation. */
306 void
307 xfs_ag_resv_alloc_extent(
308 	struct xfs_perag		*pag,
309 	enum xfs_ag_resv_type		type,
310 	struct xfs_alloc_arg		*args)
311 {
312 	struct xfs_ag_resv		*resv;
313 	xfs_extlen_t			len;
314 	uint				field;
315 
316 	trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
317 
318 	switch (type) {
319 	case XFS_AG_RESV_METADATA:
320 	case XFS_AG_RESV_AGFL:
321 		resv = xfs_perag_resv(pag, type);
322 		break;
323 	default:
324 		ASSERT(0);
325 		/* fall through */
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);
330 		return;
331 	}
332 
333 	len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
334 	resv->ar_reserved -= len;
335 	if (type == XFS_AG_RESV_AGFL)
336 		return;
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. */
340 	if (args->len > len)
341 		xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
342 				-((int64_t)args->len - len));
343 }
344 
345 /* Free a block to the reservation. */
346 void
347 xfs_ag_resv_free_extent(
348 	struct xfs_perag		*pag,
349 	enum xfs_ag_resv_type		type,
350 	struct xfs_trans		*tp,
351 	xfs_extlen_t			len)
352 {
353 	xfs_extlen_t			leftover;
354 	struct xfs_ag_resv		*resv;
355 
356 	trace_xfs_ag_resv_free_extent(pag, type, len);
357 
358 	switch (type) {
359 	case XFS_AG_RESV_METADATA:
360 	case XFS_AG_RESV_AGFL:
361 		resv = xfs_perag_resv(pag, type);
362 		break;
363 	default:
364 		ASSERT(0);
365 		/* fall through */
366 	case XFS_AG_RESV_NONE:
367 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
368 		return;
369 	}
370 
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)
374 		return;
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. */
378 	if (len > leftover)
379 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
380 }
381