xref: /linux/fs/xfs/libxfs/xfs_ag_resv.c (revision 5c35a02c545a7bbe77f3a1ae337d9e29beed079b)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Copyright (C) 2016 Oracle.  All Rights Reserved.
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_sb.h"
13 #include "xfs_mount.h"
14 #include "xfs_defer.h"
15 #include "xfs_alloc.h"
16 #include "xfs_errortag.h"
17 #include "xfs_error.h"
18 #include "xfs_trace.h"
19 #include "xfs_cksum.h"
20 #include "xfs_trans.h"
21 #include "xfs_bit.h"
22 #include "xfs_bmap.h"
23 #include "xfs_bmap_btree.h"
24 #include "xfs_ag_resv.h"
25 #include "xfs_trans_space.h"
26 #include "xfs_rmap_btree.h"
27 #include "xfs_btree.h"
28 #include "xfs_refcount_btree.h"
29 #include "xfs_ialloc_btree.h"
30 
31 /*
32  * Per-AG Block Reservations
33  *
34  * For some kinds of allocation group metadata structures, it is advantageous
35  * to reserve a small number of blocks in each AG so that future expansions of
36  * that data structure do not encounter ENOSPC because errors during a btree
37  * split cause the filesystem to go offline.
38  *
39  * Prior to the introduction of reflink, this wasn't an issue because the free
40  * space btrees maintain a reserve of space (the AGFL) to handle any expansion
41  * that may be necessary; and allocations of other metadata (inodes, BMBT,
42  * dir/attr) aren't restricted to a single AG.  However, with reflink it is
43  * possible to allocate all the space in an AG, have subsequent reflink/CoW
44  * activity expand the refcount btree, and discover that there's no space left
45  * to handle that expansion.  Since we can calculate the maximum size of the
46  * refcount btree, we can reserve space for it and avoid ENOSPC.
47  *
48  * Handling per-AG reservations consists of three changes to the allocator's
49  * behavior:  First, because these reservations are always needed, we decrease
50  * the ag_max_usable counter to reflect the size of the AG after the reserved
51  * blocks are taken.  Second, the reservations must be reflected in the
52  * fdblocks count to maintain proper accounting.  Third, each AG must maintain
53  * its own reserved block counter so that we can calculate the amount of space
54  * that must remain free to maintain the reservations.  Fourth, the "remaining
55  * reserved blocks" count must be used when calculating the length of the
56  * longest free extent in an AG and to clamp maxlen in the per-AG allocation
57  * functions.  In other words, we maintain a virtual allocation via in-core
58  * accounting tricks so that we don't have to clean up after a crash. :)
59  *
60  * Reserved blocks can be managed by passing one of the enum xfs_ag_resv_type
61  * values via struct xfs_alloc_arg or directly to the xfs_free_extent
62  * function.  It might seem a little funny to maintain a reservoir of blocks
63  * to feed another reservoir, but the AGFL only holds enough blocks to get
64  * through the next transaction.  The per-AG reservation is to ensure (we
65  * hope) that each AG never runs out of blocks.  Each data structure wanting
66  * to use the reservation system should update ask/used in xfs_ag_resv_init.
67  */
68 
69 /*
70  * Are we critically low on blocks?  For now we'll define that as the number
71  * of blocks we can get our hands on being less than 10% of what we reserved
72  * or less than some arbitrary number (maximum btree height).
73  */
74 bool
75 xfs_ag_resv_critical(
76 	struct xfs_perag		*pag,
77 	enum xfs_ag_resv_type		type)
78 {
79 	xfs_extlen_t			avail;
80 	xfs_extlen_t			orig;
81 
82 	switch (type) {
83 	case XFS_AG_RESV_METADATA:
84 		avail = pag->pagf_freeblks - pag->pag_rmapbt_resv.ar_reserved;
85 		orig = pag->pag_meta_resv.ar_asked;
86 		break;
87 	case XFS_AG_RESV_RMAPBT:
88 		avail = pag->pagf_freeblks + pag->pagf_flcount -
89 			pag->pag_meta_resv.ar_reserved;
90 		orig = pag->pag_rmapbt_resv.ar_asked;
91 		break;
92 	default:
93 		ASSERT(0);
94 		return false;
95 	}
96 
97 	trace_xfs_ag_resv_critical(pag, type, avail);
98 
99 	/* Critically low if less than 10% or max btree height remains. */
100 	return XFS_TEST_ERROR(avail < orig / 10 || avail < XFS_BTREE_MAXLEVELS,
101 			pag->pag_mount, XFS_ERRTAG_AG_RESV_CRITICAL);
102 }
103 
104 /*
105  * How many blocks are reserved but not used, and therefore must not be
106  * allocated away?
107  */
108 xfs_extlen_t
109 xfs_ag_resv_needed(
110 	struct xfs_perag		*pag,
111 	enum xfs_ag_resv_type		type)
112 {
113 	xfs_extlen_t			len;
114 
115 	len = pag->pag_meta_resv.ar_reserved + pag->pag_rmapbt_resv.ar_reserved;
116 	switch (type) {
117 	case XFS_AG_RESV_METADATA:
118 	case XFS_AG_RESV_RMAPBT:
119 		len -= xfs_perag_resv(pag, type)->ar_reserved;
120 		break;
121 	case XFS_AG_RESV_NONE:
122 		/* empty */
123 		break;
124 	default:
125 		ASSERT(0);
126 	}
127 
128 	trace_xfs_ag_resv_needed(pag, type, len);
129 
130 	return len;
131 }
132 
133 /* Clean out a reservation */
134 static int
135 __xfs_ag_resv_free(
136 	struct xfs_perag		*pag,
137 	enum xfs_ag_resv_type		type)
138 {
139 	struct xfs_ag_resv		*resv;
140 	xfs_extlen_t			oldresv;
141 	int				error;
142 
143 	trace_xfs_ag_resv_free(pag, type, 0);
144 
145 	resv = xfs_perag_resv(pag, type);
146 	if (pag->pag_agno == 0)
147 		pag->pag_mount->m_ag_max_usable += resv->ar_asked;
148 	/*
149 	 * RMAPBT blocks come from the AGFL and AGFL blocks are always
150 	 * considered "free", so whatever was reserved at mount time must be
151 	 * given back at umount.
152 	 */
153 	if (type == XFS_AG_RESV_RMAPBT)
154 		oldresv = resv->ar_orig_reserved;
155 	else
156 		oldresv = resv->ar_reserved;
157 	error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
158 	resv->ar_reserved = 0;
159 	resv->ar_asked = 0;
160 	resv->ar_orig_reserved = 0;
161 
162 	if (error)
163 		trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
164 				error, _RET_IP_);
165 	return error;
166 }
167 
168 /* Free a per-AG reservation. */
169 int
170 xfs_ag_resv_free(
171 	struct xfs_perag		*pag)
172 {
173 	int				error;
174 	int				err2;
175 
176 	error = __xfs_ag_resv_free(pag, XFS_AG_RESV_RMAPBT);
177 	err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
178 	if (err2 && !error)
179 		error = err2;
180 	return error;
181 }
182 
183 static int
184 __xfs_ag_resv_init(
185 	struct xfs_perag		*pag,
186 	enum xfs_ag_resv_type		type,
187 	xfs_extlen_t			ask,
188 	xfs_extlen_t			used)
189 {
190 	struct xfs_mount		*mp = pag->pag_mount;
191 	struct xfs_ag_resv		*resv;
192 	int				error;
193 	xfs_extlen_t			hidden_space;
194 
195 	if (used > ask)
196 		ask = used;
197 
198 	switch (type) {
199 	case XFS_AG_RESV_RMAPBT:
200 		/*
201 		 * Space taken by the rmapbt is not subtracted from fdblocks
202 		 * because the rmapbt lives in the free space.  Here we must
203 		 * subtract the entire reservation from fdblocks so that we
204 		 * always have blocks available for rmapbt expansion.
205 		 */
206 		hidden_space = ask;
207 		break;
208 	case XFS_AG_RESV_METADATA:
209 		/*
210 		 * Space taken by all other metadata btrees are accounted
211 		 * on-disk as used space.  We therefore only hide the space
212 		 * that is reserved but not used by the trees.
213 		 */
214 		hidden_space = ask - used;
215 		break;
216 	default:
217 		ASSERT(0);
218 		return -EINVAL;
219 	}
220 	error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space, true);
221 	if (error) {
222 		trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
223 				error, _RET_IP_);
224 		xfs_warn(mp,
225 "Per-AG reservation for AG %u failed.  Filesystem may run out of space.",
226 				pag->pag_agno);
227 		return error;
228 	}
229 
230 	/*
231 	 * Reduce the maximum per-AG allocation length by however much we're
232 	 * trying to reserve for an AG.  Since this is a filesystem-wide
233 	 * counter, we only make the adjustment for AG 0.  This assumes that
234 	 * there aren't any AGs hungrier for per-AG reservation than AG 0.
235 	 */
236 	if (pag->pag_agno == 0)
237 		mp->m_ag_max_usable -= ask;
238 
239 	resv = xfs_perag_resv(pag, type);
240 	resv->ar_asked = ask;
241 	resv->ar_orig_reserved = hidden_space;
242 	resv->ar_reserved = ask - used;
243 
244 	trace_xfs_ag_resv_init(pag, type, ask);
245 	return 0;
246 }
247 
248 /* Create a per-AG block reservation. */
249 int
250 xfs_ag_resv_init(
251 	struct xfs_perag		*pag)
252 {
253 	struct xfs_mount		*mp = pag->pag_mount;
254 	xfs_agnumber_t			agno = pag->pag_agno;
255 	xfs_extlen_t			ask;
256 	xfs_extlen_t			used;
257 	int				error = 0;
258 
259 	/* Create the metadata reservation. */
260 	if (pag->pag_meta_resv.ar_asked == 0) {
261 		ask = used = 0;
262 
263 		error = xfs_refcountbt_calc_reserves(mp, agno, &ask, &used);
264 		if (error)
265 			goto out;
266 
267 		error = xfs_finobt_calc_reserves(mp, agno, &ask, &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 			/*
275 			 * Because we didn't have per-AG reservations when the
276 			 * finobt feature was added we might not be able to
277 			 * reserve all needed blocks.  Warn and fall back to the
278 			 * old and potentially buggy code in that case, but
279 			 * ensure we do have the reservation for the refcountbt.
280 			 */
281 			ask = used = 0;
282 
283 			mp->m_inotbt_nores = true;
284 
285 			error = xfs_refcountbt_calc_reserves(mp, agno, &ask,
286 					&used);
287 			if (error)
288 				goto out;
289 
290 			error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
291 					ask, used);
292 			if (error)
293 				goto out;
294 		}
295 	}
296 
297 	/* Create the RMAPBT metadata reservation */
298 	if (pag->pag_rmapbt_resv.ar_asked == 0) {
299 		ask = used = 0;
300 
301 		error = xfs_rmapbt_calc_reserves(mp, agno, &ask, &used);
302 		if (error)
303 			goto out;
304 
305 		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_RMAPBT, ask, used);
306 		if (error)
307 			goto out;
308 	}
309 
310 #ifdef DEBUG
311 	/* need to read in the AGF for the ASSERT below to work */
312 	error = xfs_alloc_pagf_init(pag->pag_mount, NULL, pag->pag_agno, 0);
313 	if (error)
314 		return error;
315 
316 	ASSERT(xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
317 	       xfs_perag_resv(pag, XFS_AG_RESV_RMAPBT)->ar_reserved <=
318 	       pag->pagf_freeblks + pag->pagf_flcount);
319 #endif
320 out:
321 	return error;
322 }
323 
324 /* Allocate a block from the reservation. */
325 void
326 xfs_ag_resv_alloc_extent(
327 	struct xfs_perag		*pag,
328 	enum xfs_ag_resv_type		type,
329 	struct xfs_alloc_arg		*args)
330 {
331 	struct xfs_ag_resv		*resv;
332 	xfs_extlen_t			len;
333 	uint				field;
334 
335 	trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
336 
337 	switch (type) {
338 	case XFS_AG_RESV_AGFL:
339 		return;
340 	case XFS_AG_RESV_METADATA:
341 	case XFS_AG_RESV_RMAPBT:
342 		resv = xfs_perag_resv(pag, type);
343 		break;
344 	default:
345 		ASSERT(0);
346 		/* fall through */
347 	case XFS_AG_RESV_NONE:
348 		field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
349 				       XFS_TRANS_SB_FDBLOCKS;
350 		xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
351 		return;
352 	}
353 
354 	len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
355 	resv->ar_reserved -= len;
356 	if (type == XFS_AG_RESV_RMAPBT)
357 		return;
358 	/* Allocations of reserved blocks only need on-disk sb updates... */
359 	xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
360 	/* ...but non-reserved blocks need in-core and on-disk updates. */
361 	if (args->len > len)
362 		xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
363 				-((int64_t)args->len - len));
364 }
365 
366 /* Free a block to the reservation. */
367 void
368 xfs_ag_resv_free_extent(
369 	struct xfs_perag		*pag,
370 	enum xfs_ag_resv_type		type,
371 	struct xfs_trans		*tp,
372 	xfs_extlen_t			len)
373 {
374 	xfs_extlen_t			leftover;
375 	struct xfs_ag_resv		*resv;
376 
377 	trace_xfs_ag_resv_free_extent(pag, type, len);
378 
379 	switch (type) {
380 	case XFS_AG_RESV_AGFL:
381 		return;
382 	case XFS_AG_RESV_METADATA:
383 	case XFS_AG_RESV_RMAPBT:
384 		resv = xfs_perag_resv(pag, type);
385 		break;
386 	default:
387 		ASSERT(0);
388 		/* fall through */
389 	case XFS_AG_RESV_NONE:
390 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
391 		return;
392 	}
393 
394 	leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
395 	resv->ar_reserved += leftover;
396 	if (type == XFS_AG_RESV_RMAPBT)
397 		return;
398 	/* Freeing into the reserved pool only requires on-disk update... */
399 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
400 	/* ...but freeing beyond that requires in-core and on-disk update. */
401 	if (len > leftover)
402 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
403 }
404