xref: /linux/fs/xfs/libxfs/xfs_ag_resv.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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_mount.h"
13 #include "xfs_alloc.h"
14 #include "xfs_errortag.h"
15 #include "xfs_error.h"
16 #include "xfs_trace.h"
17 #include "xfs_trans.h"
18 #include "xfs_rmap_btree.h"
19 #include "xfs_btree.h"
20 #include "xfs_refcount_btree.h"
21 #include "xfs_ialloc_btree.h"
22 #include "xfs_ag.h"
23 #include "xfs_ag_resv.h"
24 
25 /*
26  * Per-AG Block Reservations
27  *
28  * For some kinds of allocation group metadata structures, it is advantageous
29  * to reserve a small number of blocks in each AG so that future expansions of
30  * that data structure do not encounter ENOSPC because errors during a btree
31  * split cause the filesystem to go offline.
32  *
33  * Prior to the introduction of reflink, this wasn't an issue because the free
34  * space btrees maintain a reserve of space (the AGFL) to handle any expansion
35  * that may be necessary; and allocations of other metadata (inodes, BMBT,
36  * dir/attr) aren't restricted to a single AG.  However, with reflink it is
37  * possible to allocate all the space in an AG, have subsequent reflink/CoW
38  * activity expand the refcount btree, and discover that there's no space left
39  * to handle that expansion.  Since we can calculate the maximum size of the
40  * refcount btree, we can reserve space for it and avoid ENOSPC.
41  *
42  * Handling per-AG reservations consists of three changes to the allocator's
43  * behavior:  First, because these reservations are always needed, we decrease
44  * the ag_max_usable counter to reflect the size of the AG after the reserved
45  * blocks are taken.  Second, the reservations must be reflected in the
46  * fdblocks count to maintain proper accounting.  Third, each AG must maintain
47  * its own reserved block counter so that we can calculate the amount of space
48  * that must remain free to maintain the reservations.  Fourth, the "remaining
49  * reserved blocks" count must be used when calculating the length of the
50  * longest free extent in an AG and to clamp maxlen in the per-AG allocation
51  * functions.  In other words, we maintain a virtual allocation via in-core
52  * accounting tricks so that we don't have to clean up after a crash. :)
53  *
54  * Reserved blocks can be managed by passing one of the enum xfs_ag_resv_type
55  * values via struct xfs_alloc_arg or directly to the xfs_free_extent
56  * function.  It might seem a little funny to maintain a reservoir of blocks
57  * to feed another reservoir, but the AGFL only holds enough blocks to get
58  * through the next transaction.  The per-AG reservation is to ensure (we
59  * hope) that each AG never runs out of blocks.  Each data structure wanting
60  * to use the reservation system should update ask/used in xfs_ag_resv_init.
61  */
62 
63 /*
64  * Are we critically low on blocks?  For now we'll define that as the number
65  * of blocks we can get our hands on being less than 10% of what we reserved
66  * or less than some arbitrary number (maximum btree height).
67  */
68 bool
69 xfs_ag_resv_critical(
70 	struct xfs_perag		*pag,
71 	enum xfs_ag_resv_type		type)
72 {
73 	xfs_extlen_t			avail;
74 	xfs_extlen_t			orig;
75 
76 	switch (type) {
77 	case XFS_AG_RESV_METADATA:
78 		avail = pag->pagf_freeblks - pag->pag_rmapbt_resv.ar_reserved;
79 		orig = pag->pag_meta_resv.ar_asked;
80 		break;
81 	case XFS_AG_RESV_RMAPBT:
82 		avail = pag->pagf_freeblks + pag->pagf_flcount -
83 			pag->pag_meta_resv.ar_reserved;
84 		orig = pag->pag_rmapbt_resv.ar_asked;
85 		break;
86 	default:
87 		ASSERT(0);
88 		return false;
89 	}
90 
91 	trace_xfs_ag_resv_critical(pag, type, avail);
92 
93 	/* Critically low if less than 10% or max btree height remains. */
94 	return XFS_TEST_ERROR(avail < orig / 10 ||
95 			      avail < pag->pag_mount->m_agbtree_maxlevels,
96 			pag->pag_mount, XFS_ERRTAG_AG_RESV_CRITICAL);
97 }
98 
99 /*
100  * How many blocks are reserved but not used, and therefore must not be
101  * allocated away?
102  */
103 xfs_extlen_t
104 xfs_ag_resv_needed(
105 	struct xfs_perag		*pag,
106 	enum xfs_ag_resv_type		type)
107 {
108 	xfs_extlen_t			len;
109 
110 	len = pag->pag_meta_resv.ar_reserved + pag->pag_rmapbt_resv.ar_reserved;
111 	switch (type) {
112 	case XFS_AG_RESV_METADATA:
113 	case XFS_AG_RESV_RMAPBT:
114 		len -= xfs_perag_resv(pag, type)->ar_reserved;
115 		break;
116 	case XFS_AG_RESV_NONE:
117 		/* empty */
118 		break;
119 	default:
120 		ASSERT(0);
121 	}
122 
123 	trace_xfs_ag_resv_needed(pag, type, len);
124 
125 	return len;
126 }
127 
128 /* Clean out a reservation */
129 static void
130 __xfs_ag_resv_free(
131 	struct xfs_perag		*pag,
132 	enum xfs_ag_resv_type		type)
133 {
134 	struct xfs_ag_resv		*resv;
135 	xfs_extlen_t			oldresv;
136 
137 	trace_xfs_ag_resv_free(pag, type, 0);
138 
139 	resv = xfs_perag_resv(pag, type);
140 	if (pag->pag_agno == 0)
141 		pag->pag_mount->m_ag_max_usable += resv->ar_asked;
142 	/*
143 	 * RMAPBT blocks come from the AGFL and AGFL blocks are always
144 	 * considered "free", so whatever was reserved at mount time must be
145 	 * given back at umount.
146 	 */
147 	if (type == XFS_AG_RESV_RMAPBT)
148 		oldresv = resv->ar_orig_reserved;
149 	else
150 		oldresv = resv->ar_reserved;
151 	xfs_add_fdblocks(pag->pag_mount, oldresv);
152 	resv->ar_reserved = 0;
153 	resv->ar_asked = 0;
154 	resv->ar_orig_reserved = 0;
155 }
156 
157 /* Free a per-AG reservation. */
158 void
159 xfs_ag_resv_free(
160 	struct xfs_perag		*pag)
161 {
162 	__xfs_ag_resv_free(pag, XFS_AG_RESV_RMAPBT);
163 	__xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
164 }
165 
166 static int
167 __xfs_ag_resv_init(
168 	struct xfs_perag		*pag,
169 	enum xfs_ag_resv_type		type,
170 	xfs_extlen_t			ask,
171 	xfs_extlen_t			used)
172 {
173 	struct xfs_mount		*mp = pag->pag_mount;
174 	struct xfs_ag_resv		*resv;
175 	int				error;
176 	xfs_extlen_t			hidden_space;
177 
178 	if (used > ask)
179 		ask = used;
180 
181 	switch (type) {
182 	case XFS_AG_RESV_RMAPBT:
183 		/*
184 		 * Space taken by the rmapbt is not subtracted from fdblocks
185 		 * because the rmapbt lives in the free space.  Here we must
186 		 * subtract the entire reservation from fdblocks so that we
187 		 * always have blocks available for rmapbt expansion.
188 		 */
189 		hidden_space = ask;
190 		break;
191 	case XFS_AG_RESV_METADATA:
192 		/*
193 		 * Space taken by all other metadata btrees are accounted
194 		 * on-disk as used space.  We therefore only hide the space
195 		 * that is reserved but not used by the trees.
196 		 */
197 		hidden_space = ask - used;
198 		break;
199 	default:
200 		ASSERT(0);
201 		return -EINVAL;
202 	}
203 
204 	if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_AG_RESV_FAIL))
205 		error = -ENOSPC;
206 	else
207 		error = xfs_dec_fdblocks(mp, hidden_space, true);
208 	if (error) {
209 		trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
210 				error, _RET_IP_);
211 		xfs_warn(mp,
212 "Per-AG reservation for AG %u failed.  Filesystem may run out of space.",
213 				pag->pag_agno);
214 		return error;
215 	}
216 
217 	/*
218 	 * Reduce the maximum per-AG allocation length by however much we're
219 	 * trying to reserve for an AG.  Since this is a filesystem-wide
220 	 * counter, we only make the adjustment for AG 0.  This assumes that
221 	 * there aren't any AGs hungrier for per-AG reservation than AG 0.
222 	 */
223 	if (pag->pag_agno == 0)
224 		mp->m_ag_max_usable -= ask;
225 
226 	resv = xfs_perag_resv(pag, type);
227 	resv->ar_asked = ask;
228 	resv->ar_orig_reserved = hidden_space;
229 	resv->ar_reserved = ask - used;
230 
231 	trace_xfs_ag_resv_init(pag, type, ask);
232 	return 0;
233 }
234 
235 /* Create a per-AG block reservation. */
236 int
237 xfs_ag_resv_init(
238 	struct xfs_perag		*pag,
239 	struct xfs_trans		*tp)
240 {
241 	struct xfs_mount		*mp = pag->pag_mount;
242 	xfs_extlen_t			ask;
243 	xfs_extlen_t			used;
244 	int				error = 0, error2;
245 	bool				has_resv = false;
246 
247 	/* Create the metadata reservation. */
248 	if (pag->pag_meta_resv.ar_asked == 0) {
249 		ask = used = 0;
250 
251 		error = xfs_refcountbt_calc_reserves(mp, tp, pag, &ask, &used);
252 		if (error)
253 			goto out;
254 
255 		error = xfs_finobt_calc_reserves(pag, tp, &ask, &used);
256 		if (error)
257 			goto out;
258 
259 		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
260 				ask, used);
261 		if (error) {
262 			/*
263 			 * Because we didn't have per-AG reservations when the
264 			 * finobt feature was added we might not be able to
265 			 * reserve all needed blocks.  Warn and fall back to the
266 			 * old and potentially buggy code in that case, but
267 			 * ensure we do have the reservation for the refcountbt.
268 			 */
269 			ask = used = 0;
270 
271 			mp->m_finobt_nores = true;
272 
273 			error = xfs_refcountbt_calc_reserves(mp, tp, pag, &ask,
274 					&used);
275 			if (error)
276 				goto out;
277 
278 			error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
279 					ask, used);
280 			if (error)
281 				goto out;
282 		}
283 		if (ask)
284 			has_resv = true;
285 	}
286 
287 	/* Create the RMAPBT metadata reservation */
288 	if (pag->pag_rmapbt_resv.ar_asked == 0) {
289 		ask = used = 0;
290 
291 		error = xfs_rmapbt_calc_reserves(mp, tp, pag, &ask, &used);
292 		if (error)
293 			goto out;
294 
295 		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_RMAPBT, ask, used);
296 		if (error)
297 			goto out;
298 		if (ask)
299 			has_resv = true;
300 	}
301 
302 out:
303 	/*
304 	 * Initialize the pagf if we have at least one active reservation on the
305 	 * AG. This may have occurred already via reservation calculation, but
306 	 * fall back to an explicit init to ensure the in-core allocbt usage
307 	 * counters are initialized as soon as possible. This is important
308 	 * because filesystems with large perag reservations are susceptible to
309 	 * free space reservation problems that the allocbt counter is used to
310 	 * address.
311 	 */
312 	if (has_resv) {
313 		error2 = xfs_alloc_read_agf(pag, tp, 0, NULL);
314 		if (error2)
315 			return error2;
316 
317 		/*
318 		 * If there isn't enough space in the AG to satisfy the
319 		 * reservation, let the caller know that there wasn't enough
320 		 * space.  Callers are responsible for deciding what to do
321 		 * next, since (in theory) we can stumble along with
322 		 * insufficient reservation if data blocks are being freed to
323 		 * replenish the AG's free space.
324 		 */
325 		if (!error &&
326 		    xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
327 		    xfs_perag_resv(pag, XFS_AG_RESV_RMAPBT)->ar_reserved >
328 		    pag->pagf_freeblks + pag->pagf_flcount)
329 			error = -ENOSPC;
330 	}
331 
332 	return error;
333 }
334 
335 /* Allocate a block from the reservation. */
336 void
337 xfs_ag_resv_alloc_extent(
338 	struct xfs_perag		*pag,
339 	enum xfs_ag_resv_type		type,
340 	struct xfs_alloc_arg		*args)
341 {
342 	struct xfs_ag_resv		*resv;
343 	xfs_extlen_t			len;
344 	uint				field;
345 
346 	trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
347 
348 	switch (type) {
349 	case XFS_AG_RESV_AGFL:
350 		return;
351 	case XFS_AG_RESV_METADATA:
352 	case XFS_AG_RESV_RMAPBT:
353 		resv = xfs_perag_resv(pag, type);
354 		break;
355 	default:
356 		ASSERT(0);
357 		fallthrough;
358 	case XFS_AG_RESV_NONE:
359 		field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
360 				       XFS_TRANS_SB_FDBLOCKS;
361 		xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
362 		return;
363 	}
364 
365 	len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
366 	resv->ar_reserved -= len;
367 	if (type == XFS_AG_RESV_RMAPBT)
368 		return;
369 	/* Allocations of reserved blocks only need on-disk sb updates... */
370 	xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
371 	/* ...but non-reserved blocks need in-core and on-disk updates. */
372 	if (args->len > len)
373 		xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
374 				-((int64_t)args->len - len));
375 }
376 
377 /* Free a block to the reservation. */
378 void
379 xfs_ag_resv_free_extent(
380 	struct xfs_perag		*pag,
381 	enum xfs_ag_resv_type		type,
382 	struct xfs_trans		*tp,
383 	xfs_extlen_t			len)
384 {
385 	xfs_extlen_t			leftover;
386 	struct xfs_ag_resv		*resv;
387 
388 	trace_xfs_ag_resv_free_extent(pag, type, len);
389 
390 	switch (type) {
391 	case XFS_AG_RESV_AGFL:
392 		return;
393 	case XFS_AG_RESV_METADATA:
394 	case XFS_AG_RESV_RMAPBT:
395 		resv = xfs_perag_resv(pag, type);
396 		break;
397 	default:
398 		ASSERT(0);
399 		fallthrough;
400 	case XFS_AG_RESV_NONE:
401 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
402 		fallthrough;
403 	case XFS_AG_RESV_IGNORE:
404 		return;
405 	}
406 
407 	leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
408 	resv->ar_reserved += leftover;
409 	if (type == XFS_AG_RESV_RMAPBT)
410 		return;
411 	/* Freeing into the reserved pool only requires on-disk update... */
412 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
413 	/* ...but freeing beyond that requires in-core and on-disk update. */
414 	if (len > leftover)
415 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
416 }
417