xref: /linux/fs/xfs/libxfs/xfs_ag_resv.c (revision 8137a49e1567726eb10fcf55ad141ac19804ca6b)
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 || avail < XFS_BTREE_MAXLEVELS,
95 			pag->pag_mount, XFS_ERRTAG_AG_RESV_CRITICAL);
96 }
97 
98 /*
99  * How many blocks are reserved but not used, and therefore must not be
100  * allocated away?
101  */
102 xfs_extlen_t
103 xfs_ag_resv_needed(
104 	struct xfs_perag		*pag,
105 	enum xfs_ag_resv_type		type)
106 {
107 	xfs_extlen_t			len;
108 
109 	len = pag->pag_meta_resv.ar_reserved + pag->pag_rmapbt_resv.ar_reserved;
110 	switch (type) {
111 	case XFS_AG_RESV_METADATA:
112 	case XFS_AG_RESV_RMAPBT:
113 		len -= xfs_perag_resv(pag, type)->ar_reserved;
114 		break;
115 	case XFS_AG_RESV_NONE:
116 		/* empty */
117 		break;
118 	default:
119 		ASSERT(0);
120 	}
121 
122 	trace_xfs_ag_resv_needed(pag, type, len);
123 
124 	return len;
125 }
126 
127 /* Clean out a reservation */
128 static int
129 __xfs_ag_resv_free(
130 	struct xfs_perag		*pag,
131 	enum xfs_ag_resv_type		type)
132 {
133 	struct xfs_ag_resv		*resv;
134 	xfs_extlen_t			oldresv;
135 	int				error;
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 	error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
152 	resv->ar_reserved = 0;
153 	resv->ar_asked = 0;
154 	resv->ar_orig_reserved = 0;
155 
156 	if (error)
157 		trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
158 				error, _RET_IP_);
159 	return error;
160 }
161 
162 /* Free a per-AG reservation. */
163 int
164 xfs_ag_resv_free(
165 	struct xfs_perag		*pag)
166 {
167 	int				error;
168 	int				err2;
169 
170 	error = __xfs_ag_resv_free(pag, XFS_AG_RESV_RMAPBT);
171 	err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
172 	if (err2 && !error)
173 		error = err2;
174 	return error;
175 }
176 
177 static int
178 __xfs_ag_resv_init(
179 	struct xfs_perag		*pag,
180 	enum xfs_ag_resv_type		type,
181 	xfs_extlen_t			ask,
182 	xfs_extlen_t			used)
183 {
184 	struct xfs_mount		*mp = pag->pag_mount;
185 	struct xfs_ag_resv		*resv;
186 	int				error;
187 	xfs_extlen_t			hidden_space;
188 
189 	if (used > ask)
190 		ask = used;
191 
192 	switch (type) {
193 	case XFS_AG_RESV_RMAPBT:
194 		/*
195 		 * Space taken by the rmapbt is not subtracted from fdblocks
196 		 * because the rmapbt lives in the free space.  Here we must
197 		 * subtract the entire reservation from fdblocks so that we
198 		 * always have blocks available for rmapbt expansion.
199 		 */
200 		hidden_space = ask;
201 		break;
202 	case XFS_AG_RESV_METADATA:
203 		/*
204 		 * Space taken by all other metadata btrees are accounted
205 		 * on-disk as used space.  We therefore only hide the space
206 		 * that is reserved but not used by the trees.
207 		 */
208 		hidden_space = ask - used;
209 		break;
210 	default:
211 		ASSERT(0);
212 		return -EINVAL;
213 	}
214 
215 	if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_AG_RESV_FAIL))
216 		error = -ENOSPC;
217 	else
218 		error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space, true);
219 	if (error) {
220 		trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
221 				error, _RET_IP_);
222 		xfs_warn(mp,
223 "Per-AG reservation for AG %u failed.  Filesystem may run out of space.",
224 				pag->pag_agno);
225 		return error;
226 	}
227 
228 	/*
229 	 * Reduce the maximum per-AG allocation length by however much we're
230 	 * trying to reserve for an AG.  Since this is a filesystem-wide
231 	 * counter, we only make the adjustment for AG 0.  This assumes that
232 	 * there aren't any AGs hungrier for per-AG reservation than AG 0.
233 	 */
234 	if (pag->pag_agno == 0)
235 		mp->m_ag_max_usable -= ask;
236 
237 	resv = xfs_perag_resv(pag, type);
238 	resv->ar_asked = ask;
239 	resv->ar_orig_reserved = hidden_space;
240 	resv->ar_reserved = ask - used;
241 
242 	trace_xfs_ag_resv_init(pag, type, ask);
243 	return 0;
244 }
245 
246 /* Create a per-AG block reservation. */
247 int
248 xfs_ag_resv_init(
249 	struct xfs_perag		*pag,
250 	struct xfs_trans		*tp)
251 {
252 	struct xfs_mount		*mp = pag->pag_mount;
253 	xfs_extlen_t			ask;
254 	xfs_extlen_t			used;
255 	int				error = 0, error2;
256 	bool				has_resv = false;
257 
258 	/* Create the metadata reservation. */
259 	if (pag->pag_meta_resv.ar_asked == 0) {
260 		ask = used = 0;
261 
262 		error = xfs_refcountbt_calc_reserves(mp, tp, pag, &ask, &used);
263 		if (error)
264 			goto out;
265 
266 		error = xfs_finobt_calc_reserves(mp, tp, pag, &ask, &used);
267 		if (error)
268 			goto out;
269 
270 		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
271 				ask, used);
272 		if (error) {
273 			/*
274 			 * Because we didn't have per-AG reservations when the
275 			 * finobt feature was added we might not be able to
276 			 * reserve all needed blocks.  Warn and fall back to the
277 			 * old and potentially buggy code in that case, but
278 			 * ensure we do have the reservation for the refcountbt.
279 			 */
280 			ask = used = 0;
281 
282 			mp->m_finobt_nores = true;
283 
284 			error = xfs_refcountbt_calc_reserves(mp, tp, pag, &ask,
285 					&used);
286 			if (error)
287 				goto out;
288 
289 			error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
290 					ask, used);
291 			if (error)
292 				goto out;
293 		}
294 		if (ask)
295 			has_resv = true;
296 	}
297 
298 	/* Create the RMAPBT metadata reservation */
299 	if (pag->pag_rmapbt_resv.ar_asked == 0) {
300 		ask = used = 0;
301 
302 		error = xfs_rmapbt_calc_reserves(mp, tp, pag, &ask, &used);
303 		if (error)
304 			goto out;
305 
306 		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_RMAPBT, ask, used);
307 		if (error)
308 			goto out;
309 		if (ask)
310 			has_resv = true;
311 	}
312 
313 out:
314 	/*
315 	 * Initialize the pagf if we have at least one active reservation on the
316 	 * AG. This may have occurred already via reservation calculation, but
317 	 * fall back to an explicit init to ensure the in-core allocbt usage
318 	 * counters are initialized as soon as possible. This is important
319 	 * because filesystems with large perag reservations are susceptible to
320 	 * free space reservation problems that the allocbt counter is used to
321 	 * address.
322 	 */
323 	if (has_resv) {
324 		error2 = xfs_alloc_pagf_init(mp, tp, pag->pag_agno, 0);
325 		if (error2)
326 			return error2;
327 
328 		/*
329 		 * If there isn't enough space in the AG to satisfy the
330 		 * reservation, let the caller know that there wasn't enough
331 		 * space.  Callers are responsible for deciding what to do
332 		 * next, since (in theory) we can stumble along with
333 		 * insufficient reservation if data blocks are being freed to
334 		 * replenish the AG's free space.
335 		 */
336 		if (!error &&
337 		    xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
338 		    xfs_perag_resv(pag, XFS_AG_RESV_RMAPBT)->ar_reserved >
339 		    pag->pagf_freeblks + pag->pagf_flcount)
340 			error = -ENOSPC;
341 	}
342 
343 	return error;
344 }
345 
346 /* Allocate a block from the reservation. */
347 void
348 xfs_ag_resv_alloc_extent(
349 	struct xfs_perag		*pag,
350 	enum xfs_ag_resv_type		type,
351 	struct xfs_alloc_arg		*args)
352 {
353 	struct xfs_ag_resv		*resv;
354 	xfs_extlen_t			len;
355 	uint				field;
356 
357 	trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
358 
359 	switch (type) {
360 	case XFS_AG_RESV_AGFL:
361 		return;
362 	case XFS_AG_RESV_METADATA:
363 	case XFS_AG_RESV_RMAPBT:
364 		resv = xfs_perag_resv(pag, type);
365 		break;
366 	default:
367 		ASSERT(0);
368 		fallthrough;
369 	case XFS_AG_RESV_NONE:
370 		field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
371 				       XFS_TRANS_SB_FDBLOCKS;
372 		xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
373 		return;
374 	}
375 
376 	len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
377 	resv->ar_reserved -= len;
378 	if (type == XFS_AG_RESV_RMAPBT)
379 		return;
380 	/* Allocations of reserved blocks only need on-disk sb updates... */
381 	xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
382 	/* ...but non-reserved blocks need in-core and on-disk updates. */
383 	if (args->len > len)
384 		xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
385 				-((int64_t)args->len - len));
386 }
387 
388 /* Free a block to the reservation. */
389 void
390 xfs_ag_resv_free_extent(
391 	struct xfs_perag		*pag,
392 	enum xfs_ag_resv_type		type,
393 	struct xfs_trans		*tp,
394 	xfs_extlen_t			len)
395 {
396 	xfs_extlen_t			leftover;
397 	struct xfs_ag_resv		*resv;
398 
399 	trace_xfs_ag_resv_free_extent(pag, type, len);
400 
401 	switch (type) {
402 	case XFS_AG_RESV_AGFL:
403 		return;
404 	case XFS_AG_RESV_METADATA:
405 	case XFS_AG_RESV_RMAPBT:
406 		resv = xfs_perag_resv(pag, type);
407 		break;
408 	default:
409 		ASSERT(0);
410 		fallthrough;
411 	case XFS_AG_RESV_NONE:
412 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
413 		return;
414 	}
415 
416 	leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
417 	resv->ar_reserved += leftover;
418 	if (type == XFS_AG_RESV_RMAPBT)
419 		return;
420 	/* Freeing into the reserved pool only requires on-disk update... */
421 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
422 	/* ...but freeing beyond that requires in-core and on-disk update. */
423 	if (len > leftover)
424 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
425 }
426