1 /* 2 * Copyright (C) 2016 Oracle. All Rights Reserved. 3 * 4 * Author: Darrick J. Wong <darrick.wong@oracle.com> 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public License 8 * as published by the Free Software Foundation; either version 2 9 * of the License, or (at your option) any later version. 10 * 11 * This program is distributed in the hope that it would be useful, 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 * GNU General Public License for more details. 15 * 16 * You should have received a copy of the GNU General Public License 17 * along with this program; if not, write the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. 19 */ 20 #include "xfs.h" 21 #include "xfs_fs.h" 22 #include "xfs_shared.h" 23 #include "xfs_format.h" 24 #include "xfs_log_format.h" 25 #include "xfs_trans_resv.h" 26 #include "xfs_sb.h" 27 #include "xfs_mount.h" 28 #include "xfs_defer.h" 29 #include "xfs_alloc.h" 30 #include "xfs_error.h" 31 #include "xfs_trace.h" 32 #include "xfs_cksum.h" 33 #include "xfs_trans.h" 34 #include "xfs_bit.h" 35 #include "xfs_bmap.h" 36 #include "xfs_bmap_btree.h" 37 #include "xfs_ag_resv.h" 38 #include "xfs_trans_space.h" 39 #include "xfs_rmap_btree.h" 40 #include "xfs_btree.h" 41 #include "xfs_refcount_btree.h" 42 #include "xfs_ialloc_btree.h" 43 44 /* 45 * Per-AG Block Reservations 46 * 47 * For some kinds of allocation group metadata structures, it is advantageous 48 * to reserve a small number of blocks in each AG so that future expansions of 49 * that data structure do not encounter ENOSPC because errors during a btree 50 * split cause the filesystem to go offline. 51 * 52 * Prior to the introduction of reflink, this wasn't an issue because the free 53 * space btrees maintain a reserve of space (the AGFL) to handle any expansion 54 * that may be necessary; and allocations of other metadata (inodes, BMBT, 55 * dir/attr) aren't restricted to a single AG. However, with reflink it is 56 * possible to allocate all the space in an AG, have subsequent reflink/CoW 57 * activity expand the refcount btree, and discover that there's no space left 58 * to handle that expansion. Since we can calculate the maximum size of the 59 * refcount btree, we can reserve space for it and avoid ENOSPC. 60 * 61 * Handling per-AG reservations consists of three changes to the allocator's 62 * behavior: First, because these reservations are always needed, we decrease 63 * the ag_max_usable counter to reflect the size of the AG after the reserved 64 * blocks are taken. Second, the reservations must be reflected in the 65 * fdblocks count to maintain proper accounting. Third, each AG must maintain 66 * its own reserved block counter so that we can calculate the amount of space 67 * that must remain free to maintain the reservations. Fourth, the "remaining 68 * reserved blocks" count must be used when calculating the length of the 69 * longest free extent in an AG and to clamp maxlen in the per-AG allocation 70 * functions. In other words, we maintain a virtual allocation via in-core 71 * accounting tricks so that we don't have to clean up after a crash. :) 72 * 73 * Reserved blocks can be managed by passing one of the enum xfs_ag_resv_type 74 * values via struct xfs_alloc_arg or directly to the xfs_free_extent 75 * function. It might seem a little funny to maintain a reservoir of blocks 76 * to feed another reservoir, but the AGFL only holds enough blocks to get 77 * through the next transaction. The per-AG reservation is to ensure (we 78 * hope) that each AG never runs out of blocks. Each data structure wanting 79 * to use the reservation system should update ask/used in xfs_ag_resv_init. 80 */ 81 82 /* 83 * Are we critically low on blocks? For now we'll define that as the number 84 * of blocks we can get our hands on being less than 10% of what we reserved 85 * or less than some arbitrary number (maximum btree height). 86 */ 87 bool 88 xfs_ag_resv_critical( 89 struct xfs_perag *pag, 90 enum xfs_ag_resv_type type) 91 { 92 xfs_extlen_t avail; 93 xfs_extlen_t orig; 94 95 switch (type) { 96 case XFS_AG_RESV_METADATA: 97 avail = pag->pagf_freeblks - pag->pag_agfl_resv.ar_reserved; 98 orig = pag->pag_meta_resv.ar_asked; 99 break; 100 case XFS_AG_RESV_AGFL: 101 avail = pag->pagf_freeblks + pag->pagf_flcount - 102 pag->pag_meta_resv.ar_reserved; 103 orig = pag->pag_agfl_resv.ar_asked; 104 break; 105 default: 106 ASSERT(0); 107 return false; 108 } 109 110 trace_xfs_ag_resv_critical(pag, type, avail); 111 112 /* Critically low if less than 10% or max btree height remains. */ 113 return XFS_TEST_ERROR(avail < orig / 10 || avail < XFS_BTREE_MAXLEVELS, 114 pag->pag_mount, XFS_ERRTAG_AG_RESV_CRITICAL); 115 } 116 117 /* 118 * How many blocks are reserved but not used, and therefore must not be 119 * allocated away? 120 */ 121 xfs_extlen_t 122 xfs_ag_resv_needed( 123 struct xfs_perag *pag, 124 enum xfs_ag_resv_type type) 125 { 126 xfs_extlen_t len; 127 128 len = pag->pag_meta_resv.ar_reserved + pag->pag_agfl_resv.ar_reserved; 129 switch (type) { 130 case XFS_AG_RESV_METADATA: 131 case XFS_AG_RESV_AGFL: 132 len -= xfs_perag_resv(pag, type)->ar_reserved; 133 break; 134 case XFS_AG_RESV_NONE: 135 /* empty */ 136 break; 137 default: 138 ASSERT(0); 139 } 140 141 trace_xfs_ag_resv_needed(pag, type, len); 142 143 return len; 144 } 145 146 /* Clean out a reservation */ 147 static int 148 __xfs_ag_resv_free( 149 struct xfs_perag *pag, 150 enum xfs_ag_resv_type type) 151 { 152 struct xfs_ag_resv *resv; 153 xfs_extlen_t oldresv; 154 int error; 155 156 trace_xfs_ag_resv_free(pag, type, 0); 157 158 resv = xfs_perag_resv(pag, type); 159 if (pag->pag_agno == 0) 160 pag->pag_mount->m_ag_max_usable += resv->ar_asked; 161 /* 162 * AGFL blocks are always considered "free", so whatever 163 * was reserved at mount time must be given back at umount. 164 */ 165 if (type == XFS_AG_RESV_AGFL) 166 oldresv = resv->ar_orig_reserved; 167 else 168 oldresv = resv->ar_reserved; 169 error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true); 170 resv->ar_reserved = 0; 171 resv->ar_asked = 0; 172 173 if (error) 174 trace_xfs_ag_resv_free_error(pag->pag_mount, pag->pag_agno, 175 error, _RET_IP_); 176 return error; 177 } 178 179 /* Free a per-AG reservation. */ 180 int 181 xfs_ag_resv_free( 182 struct xfs_perag *pag) 183 { 184 int error; 185 int err2; 186 187 error = __xfs_ag_resv_free(pag, XFS_AG_RESV_AGFL); 188 err2 = __xfs_ag_resv_free(pag, XFS_AG_RESV_METADATA); 189 if (err2 && !error) 190 error = err2; 191 return error; 192 } 193 194 static int 195 __xfs_ag_resv_init( 196 struct xfs_perag *pag, 197 enum xfs_ag_resv_type type, 198 xfs_extlen_t ask, 199 xfs_extlen_t used) 200 { 201 struct xfs_mount *mp = pag->pag_mount; 202 struct xfs_ag_resv *resv; 203 int error; 204 xfs_extlen_t reserved; 205 206 if (used > ask) 207 ask = used; 208 reserved = ask - used; 209 210 error = xfs_mod_fdblocks(mp, -(int64_t)reserved, true); 211 if (error) { 212 trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno, 213 error, _RET_IP_); 214 xfs_warn(mp, 215 "Per-AG reservation for AG %u failed. Filesystem may run out of space.", 216 pag->pag_agno); 217 return error; 218 } 219 220 /* 221 * Reduce the maximum per-AG allocation length by however much we're 222 * trying to reserve for an AG. Since this is a filesystem-wide 223 * counter, we only make the adjustment for AG 0. This assumes that 224 * there aren't any AGs hungrier for per-AG reservation than AG 0. 225 */ 226 if (pag->pag_agno == 0) 227 mp->m_ag_max_usable -= ask; 228 229 resv = xfs_perag_resv(pag, type); 230 resv->ar_asked = ask; 231 resv->ar_reserved = resv->ar_orig_reserved = reserved; 232 233 trace_xfs_ag_resv_init(pag, type, ask); 234 return 0; 235 } 236 237 /* Create a per-AG block reservation. */ 238 int 239 xfs_ag_resv_init( 240 struct xfs_perag *pag) 241 { 242 struct xfs_mount *mp = pag->pag_mount; 243 xfs_agnumber_t agno = pag->pag_agno; 244 xfs_extlen_t ask; 245 xfs_extlen_t used; 246 int error = 0; 247 248 /* Create the metadata reservation. */ 249 if (pag->pag_meta_resv.ar_asked == 0) { 250 ask = used = 0; 251 252 error = xfs_refcountbt_calc_reserves(mp, agno, &ask, &used); 253 if (error) 254 goto out; 255 256 error = xfs_finobt_calc_reserves(mp, agno, &ask, &used); 257 if (error) 258 goto out; 259 260 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA, 261 ask, used); 262 if (error) { 263 /* 264 * Because we didn't have per-AG reservations when the 265 * finobt feature was added we might not be able to 266 * reserve all needed blocks. Warn and fall back to the 267 * old and potentially buggy code in that case, but 268 * ensure we do have the reservation for the refcountbt. 269 */ 270 ask = used = 0; 271 272 mp->m_inotbt_nores = true; 273 274 error = xfs_refcountbt_calc_reserves(mp, agno, &ask, 275 &used); 276 if (error) 277 goto out; 278 279 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_METADATA, 280 ask, used); 281 if (error) 282 goto out; 283 } 284 } 285 286 /* Create the AGFL metadata reservation */ 287 if (pag->pag_agfl_resv.ar_asked == 0) { 288 ask = used = 0; 289 290 error = xfs_rmapbt_calc_reserves(mp, agno, &ask, &used); 291 if (error) 292 goto out; 293 294 error = __xfs_ag_resv_init(pag, XFS_AG_RESV_AGFL, ask, used); 295 if (error) 296 goto out; 297 } 298 299 #ifdef DEBUG 300 /* need to read in the AGF for the ASSERT below to work */ 301 error = xfs_alloc_pagf_init(pag->pag_mount, NULL, pag->pag_agno, 0); 302 if (error) 303 return error; 304 305 ASSERT(xfs_perag_resv(pag, XFS_AG_RESV_METADATA)->ar_reserved + 306 xfs_perag_resv(pag, XFS_AG_RESV_AGFL)->ar_reserved <= 307 pag->pagf_freeblks + pag->pagf_flcount); 308 #endif 309 out: 310 return error; 311 } 312 313 /* Allocate a block from the reservation. */ 314 void 315 xfs_ag_resv_alloc_extent( 316 struct xfs_perag *pag, 317 enum xfs_ag_resv_type type, 318 struct xfs_alloc_arg *args) 319 { 320 struct xfs_ag_resv *resv; 321 xfs_extlen_t len; 322 uint field; 323 324 trace_xfs_ag_resv_alloc_extent(pag, type, args->len); 325 326 switch (type) { 327 case XFS_AG_RESV_METADATA: 328 case XFS_AG_RESV_AGFL: 329 resv = xfs_perag_resv(pag, type); 330 break; 331 default: 332 ASSERT(0); 333 /* fall through */ 334 case XFS_AG_RESV_NONE: 335 field = args->wasdel ? XFS_TRANS_SB_RES_FDBLOCKS : 336 XFS_TRANS_SB_FDBLOCKS; 337 xfs_trans_mod_sb(args->tp, field, -(int64_t)args->len); 338 return; 339 } 340 341 len = min_t(xfs_extlen_t, args->len, resv->ar_reserved); 342 resv->ar_reserved -= len; 343 if (type == XFS_AG_RESV_AGFL) 344 return; 345 /* Allocations of reserved blocks only need on-disk sb updates... */ 346 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_RES_FDBLOCKS, -(int64_t)len); 347 /* ...but non-reserved blocks need in-core and on-disk updates. */ 348 if (args->len > len) 349 xfs_trans_mod_sb(args->tp, XFS_TRANS_SB_FDBLOCKS, 350 -((int64_t)args->len - len)); 351 } 352 353 /* Free a block to the reservation. */ 354 void 355 xfs_ag_resv_free_extent( 356 struct xfs_perag *pag, 357 enum xfs_ag_resv_type type, 358 struct xfs_trans *tp, 359 xfs_extlen_t len) 360 { 361 xfs_extlen_t leftover; 362 struct xfs_ag_resv *resv; 363 364 trace_xfs_ag_resv_free_extent(pag, type, len); 365 366 switch (type) { 367 case XFS_AG_RESV_METADATA: 368 case XFS_AG_RESV_AGFL: 369 resv = xfs_perag_resv(pag, type); 370 break; 371 default: 372 ASSERT(0); 373 /* fall through */ 374 case XFS_AG_RESV_NONE: 375 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, (int64_t)len); 376 return; 377 } 378 379 leftover = min_t(xfs_extlen_t, len, resv->ar_asked - resv->ar_reserved); 380 resv->ar_reserved += leftover; 381 if (type == XFS_AG_RESV_AGFL) 382 return; 383 /* Freeing into the reserved pool only requires on-disk update... */ 384 xfs_trans_mod_sb(tp, XFS_TRANS_SB_RES_FDBLOCKS, len); 385 /* ...but freeing beyond that requires in-core and on-disk update. */ 386 if (len > leftover) 387 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, len - leftover); 388 } 389