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
xfs_ag_resv_critical(struct xfs_perag * pag,enum xfs_ag_resv_type type)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
xfs_ag_resv_needed(struct xfs_perag * pag,enum xfs_ag_resv_type type)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
__xfs_ag_resv_free(struct xfs_perag * pag,enum xfs_ag_resv_type type)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
xfs_ag_resv_free(struct xfs_perag * pag)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
__xfs_ag_resv_init(struct xfs_perag * pag,enum xfs_ag_resv_type type,xfs_extlen_t ask,xfs_extlen_t used)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
xfs_ag_resv_init(struct xfs_perag * pag,struct xfs_trans * tp)250 xfs_ag_resv_init(
251 	struct xfs_perag		*pag,
252 	struct xfs_trans		*tp)
253 {
254 	struct xfs_mount		*mp = pag->pag_mount;
255 	xfs_agnumber_t			agno = pag->pag_agno;
256 	xfs_extlen_t			ask;
257 	xfs_extlen_t			used;
258 	int				error = 0;
259 
260 	/* Create the metadata reservation. */
261 	if (pag->pag_meta_resv.ar_asked == 0) {
262 		ask = used = 0;
263 
264 		error = xfs_refcountbt_calc_reserves(mp, tp, agno, &ask, &used);
265 		if (error)
266 			goto out;
267 
268 		error = xfs_finobt_calc_reserves(mp, tp, agno, &ask, &used);
269 		if (error)
270 			goto out;
271 
272 		error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
273 				ask, used);
274 		if (error) {
275 			/*
276 			 * Because we didn't have per-AG reservations when the
277 			 * finobt feature was added we might not be able to
278 			 * reserve all needed blocks.  Warn and fall back to the
279 			 * old and potentially buggy code in that case, but
280 			 * ensure we do have the reservation for the refcountbt.
281 			 */
282 			ask = used = 0;
283 
284 			mp->m_finobt_nores = true;
285 
286 			error = xfs_refcountbt_calc_reserves(mp, tp, agno, &ask,
287 					&used);
288 			if (error)
289 				goto out;
290 
291 			error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
292 					ask, used);
293 			if (error)
294 				goto out;
295 		}
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, agno, &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 	}
310 
311 #ifdef DEBUG
312 	/* need to read in the AGF for the ASSERT below to work */
313 	error = xfs_alloc_pagf_init(pag->pag_mount, tp, pag->pag_agno, 0);
314 	if (error)
315 		return error;
316 
317 	ASSERT(xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
318 	       xfs_perag_resv(pag, XFS_AG_RESV_RMAPBT)->ar_reserved <=
319 	       pag->pagf_freeblks + pag->pagf_flcount);
320 #endif
321 out:
322 	return error;
323 }
324 
325 /* Allocate a block from the reservation. */
326 void
xfs_ag_resv_alloc_extent(struct xfs_perag * pag,enum xfs_ag_resv_type type,struct xfs_alloc_arg * args)327 xfs_ag_resv_alloc_extent(
328 	struct xfs_perag		*pag,
329 	enum xfs_ag_resv_type		type,
330 	struct xfs_alloc_arg		*args)
331 {
332 	struct xfs_ag_resv		*resv;
333 	xfs_extlen_t			len;
334 	uint				field;
335 
336 	trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
337 
338 	switch (type) {
339 	case XFS_AG_RESV_AGFL:
340 		return;
341 	case XFS_AG_RESV_METADATA:
342 	case XFS_AG_RESV_RMAPBT:
343 		resv = xfs_perag_resv(pag, type);
344 		break;
345 	default:
346 		ASSERT(0);
347 		/* fall through */
348 	case XFS_AG_RESV_NONE:
349 		field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
350 				       XFS_TRANS_SB_FDBLOCKS;
351 		xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
352 		return;
353 	}
354 
355 	len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
356 	resv->ar_reserved -= len;
357 	if (type == XFS_AG_RESV_RMAPBT)
358 		return;
359 	/* Allocations of reserved blocks only need on-disk sb updates... */
360 	xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
361 	/* ...but non-reserved blocks need in-core and on-disk updates. */
362 	if (args->len > len)
363 		xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
364 				-((int64_t)args->len - len));
365 }
366 
367 /* Free a block to the reservation. */
368 void
xfs_ag_resv_free_extent(struct xfs_perag * pag,enum xfs_ag_resv_type type,struct xfs_trans * tp,xfs_extlen_t len)369 xfs_ag_resv_free_extent(
370 	struct xfs_perag		*pag,
371 	enum xfs_ag_resv_type		type,
372 	struct xfs_trans		*tp,
373 	xfs_extlen_t			len)
374 {
375 	xfs_extlen_t			leftover;
376 	struct xfs_ag_resv		*resv;
377 
378 	trace_xfs_ag_resv_free_extent(pag, type, len);
379 
380 	switch (type) {
381 	case XFS_AG_RESV_AGFL:
382 		return;
383 	case XFS_AG_RESV_METADATA:
384 	case XFS_AG_RESV_RMAPBT:
385 		resv = xfs_perag_resv(pag, type);
386 		break;
387 	default:
388 		ASSERT(0);
389 		/* fall through */
390 	case XFS_AG_RESV_NONE:
391 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
392 		return;
393 	}
394 
395 	leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
396 	resv->ar_reserved += leftover;
397 	if (type == XFS_AG_RESV_RMAPBT)
398 		return;
399 	/* Freeing into the reserved pool only requires on-disk update... */
400 	xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
401 	/* ...but freeing beyond that requires in-core and on-disk update. */
402 	if (len > leftover)
403 		xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
404 }
405