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
23 /*
24 * Per-AG Block Reservations
25 *
26 * For some kinds of allocation group metadata structures, it is advantageous
27 * to reserve a small number of blocks in each AG so that future expansions of
28 * that data structure do not encounter ENOSPC because errors during a btree
29 * split cause the filesystem to go offline.
30 *
31 * Prior to the introduction of reflink, this wasn't an issue because the free
32 * space btrees maintain a reserve of space (the AGFL) to handle any expansion
33 * that may be necessary; and allocations of other metadata (inodes, BMBT,
34 * dir/attr) aren't restricted to a single AG. However, with reflink it is
35 * possible to allocate all the space in an AG, have subsequent reflink/CoW
36 * activity expand the refcount btree, and discover that there's no space left
37 * to handle that expansion. Since we can calculate the maximum size of the
38 * refcount btree, we can reserve space for it and avoid ENOSPC.
39 *
40 * Handling per-AG reservations consists of three changes to the allocator's
41 * behavior: First, because these reservations are always needed, we decrease
42 * the ag_max_usable counter to reflect the size of the AG after the reserved
43 * blocks are taken. Second, the reservations must be reflected in the
44 * fdblocks count to maintain proper accounting. Third, each AG must maintain
45 * its own reserved block counter so that we can calculate the amount of space
46 * that must remain free to maintain the reservations. Fourth, the "remaining
47 * reserved blocks" count must be used when calculating the length of the
48 * longest free extent in an AG and to clamp maxlen in the per-AG allocation
49 * functions. In other words, we maintain a virtual allocation via in-core
50 * accounting tricks so that we don't have to clean up after a crash. :)
51 *
52 * Reserved blocks can be managed by passing one of the enum xfs_ag_resv_type
53 * values via struct xfs_alloc_arg or directly to the xfs_free_extent
54 * function. It might seem a little funny to maintain a reservoir of blocks
55 * to feed another reservoir, but the AGFL only holds enough blocks to get
56 * through the next transaction. The per-AG reservation is to ensure (we
57 * hope) that each AG never runs out of blocks. Each data structure wanting
58 * to use the reservation system should update ask/used in xfs_ag_resv_init.
59 */
60
61 /*
62 * Are we critically low on blocks? For now we'll define that as the number
63 * of blocks we can get our hands on being less than 10% of what we reserved
64 * or less than some arbitrary number (maximum btree height).
65 */
66 bool
67 xfs_ag_resv_critical(
68 struct xfs_perag *pag,
69 enum xfs_ag_resv_type type)
70 {
71 xfs_extlen_t avail;
72 xfs_extlen_t orig;
73
74 switch (type) {
75 case XFS_AG_RESV_METADATA:
76 avail = pag->pagf_freeblks - pag->pag_rmapbt_resv.ar_reserved;
77 orig = pag->pag_meta_resv.ar_asked;
78 break;
79 case XFS_AG_RESV_RMAPBT:
80 avail = pag->pagf_freeblks + pag->pagf_flcount -
81 pag->pag_meta_resv.ar_reserved;
82 orig = pag->pag_rmapbt_resv.ar_asked;
83 break;
84 default:
85 ASSERT(0);
86 return false;
87 }
88
89 trace_xfs_ag_resv_critical(pag, type, avail);
90
91 /* Critically low if less than 10% or max btree height remains. */
92 return XFS_TEST_ERROR(avail < orig / 10 || avail < XFS_BTREE_MAXLEVELS,
93 pag->pag_mount, XFS_ERRTAG_AG_RESV_CRITICAL);
94 }
95
96 /*
97 * How many blocks are reserved but not used, and therefore must not be
98 * allocated away?
99 */
100 xfs_extlen_t
101 xfs_ag_resv_needed(
102 struct xfs_perag *pag,
103 enum xfs_ag_resv_type type)
104 {
105 xfs_extlen_t len;
106
107 len = pag->pag_meta_resv.ar_reserved + pag->pag_rmapbt_resv.ar_reserved;
108 switch (type) {
109 case XFS_AG_RESV_METADATA:
110 case XFS_AG_RESV_RMAPBT:
111 len -= xfs_perag_resv(pag, type)->ar_reserved;
112 break;
113 case XFS_AG_RESV_NONE:
114 /* empty */
115 break;
116 default:
117 ASSERT(0);
118 }
119
120 trace_xfs_ag_resv_needed(pag, type, len);
121
122 return len;
123 }
124
125 /* Clean out a reservation */
126 static int
127 __xfs_ag_resv_free(
128 struct xfs_perag *pag,
129 enum xfs_ag_resv_type type)
130 {
131 struct xfs_ag_resv *resv;
132 xfs_extlen_t oldresv;
133 int error;
134
135 trace_xfs_ag_resv_free(pag, type, 0);
136
137 resv = xfs_perag_resv(pag, type);
138 if (pag->pag_agno == 0)
139 pag->pag_mount->m_ag_max_usable += resv->ar_asked;
140 /*
141 * RMAPBT blocks come from the AGFL and AGFL blocks are always
142 * considered "free", so whatever was reserved at mount time must be
143 * given back at umount.
144 */
145 if (type == XFS_AG_RESV_RMAPBT)
146 oldresv = resv->ar_orig_reserved;
147 else
148 oldresv = resv->ar_reserved;
149 error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
150 resv->ar_reserved = 0;
151 resv->ar_asked = 0;
152 resv->ar_orig_reserved = 0;
153
154 if (error)
155 trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno,
156 error, _RET_IP_);
157 return error;
158 }
159
160 /* Free a per-AG reservation. */
161 int
162 xfs_ag_resv_free(
163 struct xfs_perag *pag)
164 {
165 int error;
166 int err2;
167
168 error = __xfs_ag_resv_free(pag, XFS_AG_RESV_RMAPBT);
169 err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA);
170 if (err2 && !error)
171 error = err2;
172 return error;
173 }
174
175 static int
176 __xfs_ag_resv_init(
177 struct xfs_perag *pag,
178 enum xfs_ag_resv_type type,
179 xfs_extlen_t ask,
180 xfs_extlen_t used)
181 {
182 struct xfs_mount *mp = pag->pag_mount;
183 struct xfs_ag_resv *resv;
184 int error;
185 xfs_extlen_t hidden_space;
186
187 if (used > ask)
188 ask = used;
189
190 switch (type) {
191 case XFS_AG_RESV_RMAPBT:
192 /*
193 * Space taken by the rmapbt is not subtracted from fdblocks
194 * because the rmapbt lives in the free space. Here we must
195 * subtract the entire reservation from fdblocks so that we
196 * always have blocks available for rmapbt expansion.
197 */
198 hidden_space = ask;
199 break;
200 case XFS_AG_RESV_METADATA:
201 /*
202 * Space taken by all other metadata btrees are accounted
203 * on-disk as used space. We therefore only hide the space
204 * that is reserved but not used by the trees.
205 */
206 hidden_space = ask - used;
207 break;
208 default:
209 ASSERT(0);
210 return -EINVAL;
211 }
212 error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space, true);
213 if (error) {
214 trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
215 error, _RET_IP_);
216 xfs_warn(mp,
217 "Per-AG reservation for AG %u failed. Filesystem may run out of space.",
218 pag->pag_agno);
219 return error;
220 }
221
222 /*
223 * Reduce the maximum per-AG allocation length by however much we're
224 * trying to reserve for an AG. Since this is a filesystem-wide
225 * counter, we only make the adjustment for AG 0. This assumes that
226 * there aren't any AGs hungrier for per-AG reservation than AG 0.
227 */
228 if (pag->pag_agno == 0)
229 mp->m_ag_max_usable -= ask;
230
231 resv = xfs_perag_resv(pag, type);
232 resv->ar_asked = ask;
233 resv->ar_orig_reserved = hidden_space;
234 resv->ar_reserved = ask - used;
235
236 trace_xfs_ag_resv_init(pag, type, ask);
237 return 0;
238 }
239
240 /* Create a per-AG block reservation. */
241 int
242 xfs_ag_resv_init(
243 struct xfs_perag *pag,
244 struct xfs_trans *tp)
245 {
246 struct xfs_mount *mp = pag->pag_mount;
247 xfs_agnumber_t agno = pag->pag_agno;
248 xfs_extlen_t ask;
249 xfs_extlen_t used;
250 int error = 0;
251
252 /* Create the metadata reservation. */
253 if (pag->pag_meta_resv.ar_asked == 0) {
254 ask = used = 0;
255
256 error = xfs_refcountbt_calc_reserves(mp, tp, agno, &ask, &used);
257 if (error)
258 goto out;
259
260 error = xfs_finobt_calc_reserves(mp, tp, agno, &ask, &used);
261 if (error)
262 goto out;
263
264 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
265 ask, used);
266 if (error) {
267 /*
268 * Because we didn't have per-AG reservations when the
269 * finobt feature was added we might not be able to
270 * reserve all needed blocks. Warn and fall back to the
271 * old and potentially buggy code in that case, but
272 * ensure we do have the reservation for the refcountbt.
273 */
274 ask = used = 0;
275
276 mp->m_finobt_nores = true;
277
278 error = xfs_refcountbt_calc_reserves(mp, tp, agno, &ask,
279 &used);
280 if (error)
281 goto out;
282
283 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA,
284 ask, used);
285 if (error)
286 goto out;
287 }
288 }
289
290 /* Create the RMAPBT metadata reservation */
291 if (pag->pag_rmapbt_resv.ar_asked == 0) {
292 ask = used = 0;
293
294 error = xfs_rmapbt_calc_reserves(mp, tp, agno, &ask, &used);
295 if (error)
296 goto out;
297
298 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_RMAPBT, ask, used);
299 if (error)
300 goto out;
301 }
302
303 #ifdef DEBUG
304 /* need to read in the AGF for the ASSERT below to work */
305 error = xfs_alloc_pagf_init(pag->pag_mount, tp, pag->pag_agno, 0);
306 if (error)
307 return error;
308
309 ASSERT(xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved +
310 xfs_perag_resv(pag, XFS_AG_RESV_RMAPBT)->ar_reserved <=
311 pag->pagf_freeblks + pag->pagf_flcount);
312 #endif
313 out:
314 return error;
315 }
316
317 /* Allocate a block from the reservation. */
318 void
319 xfs_ag_resv_alloc_extent(
320 struct xfs_perag *pag,
321 enum xfs_ag_resv_type type,
322 struct xfs_alloc_arg *args)
323 {
324 struct xfs_ag_resv *resv;
325 xfs_extlen_t len;
326 uint field;
327
328 trace_xfs_ag_resv_alloc_extent(pag, type, args->len);
329
330 switch (type) {
331 case XFS_AG_RESV_AGFL:
332 return;
333 case XFS_AG_RESV_METADATA:
334 case XFS_AG_RESV_RMAPBT:
335 resv = xfs_perag_resv(pag, type);
336 break;
337 default:
338 ASSERT(0);
339 /* fall through */
340 case XFS_AG_RESV_NONE:
341 field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS :
342 XFS_TRANS_SB_FDBLOCKS;
343 xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len);
344 return;
345 }
346
347 len = min_t(xfs_extlen_t, args->len, resv->ar_reserved);
348 resv->ar_reserved -= len;
349 if (type == XFS_AG_RESV_RMAPBT)
350 return;
351 /* Allocations of reserved blocks only need on-disk sb updates... */
352 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len);
353 /* ...but non-reserved blocks need in-core and on-disk updates. */
354 if (args->len > len)
355 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS,
356 -((int64_t)args->len - len));
357 }
358
359 /* Free a block to the reservation. */
360 void
361 xfs_ag_resv_free_extent(
362 struct xfs_perag *pag,
363 enum xfs_ag_resv_type type,
364 struct xfs_trans *tp,
365 xfs_extlen_t len)
366 {
367 xfs_extlen_t leftover;
368 struct xfs_ag_resv *resv;
369
370 trace_xfs_ag_resv_free_extent(pag, type, len);
371
372 switch (type) {
373 case XFS_AG_RESV_AGFL:
374 return;
375 case XFS_AG_RESV_METADATA:
376 case XFS_AG_RESV_RMAPBT:
377 resv = xfs_perag_resv(pag, type);
378 break;
379 default:
380 ASSERT(0);
381 /* fall through */
382 case XFS_AG_RESV_NONE:
383 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len);
384 return;
385 }
386
387 leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved);
388 resv->ar_reserved += leftover;
389 if (type == XFS_AG_RESV_RMAPBT)
390 return;
391 /* Freeing into the reserved pool only requires on-disk update... */
392 xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len);
393 /* ...but freeing beyond that requires in-core and on-disk update. */
394 if (len > leftover)
395 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover);
396 }