1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Lustre Light file operations
6 * Copyright (c) 2002-2004 Cluster File Systems, Inc.
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #define DEBUG_SUBSYSTEM S_LLITE
30 #include <sys/types.h>
32 #include <sys/queue.h>
48 #include "llite_lib.h"
50 /* Pack the required supplementary groups into the supplied groups array.
51 * If we don't need to use the groups from the target inode(s) then we
52 * instead pack one or more groups from the user's supplementary group
53 * array in case it might be useful. Not needed if doing an MDS-side upcall. */
54 void ll_i2gids(__u32 *suppgids, struct inode *i1, struct inode *i2)
57 LASSERT(suppgids != NULL);
59 if (in_group_p(i1->i_stbuf.st_gid))
60 suppgids[0] = i1->i_stbuf.st_gid;
65 if (in_group_p(i2->i_stbuf.st_gid))
66 suppgids[1] = i2->i_stbuf.st_gid;
74 void llu_prep_md_op_data(struct md_op_data *op_data, struct inode *i1,
75 struct inode *i2, const char *name, int namelen,
78 LASSERT(i1 != NULL || i2 != NULL);
82 ll_i2gids(op_data->op_suppgids, i1, i2);
83 op_data->op_fid1 = *ll_inode2fid(i1);
85 ll_i2gids(op_data->op_suppgids, i2, i1);
86 op_data->op_fid1 = *ll_inode2fid(i2);
90 op_data->op_fid2 = *ll_inode2fid(i2);
92 fid_zero(&op_data->op_fid2);
94 op_data->op_opc = opc;
95 op_data->op_name = name;
96 op_data->op_mode = mode;
97 op_data->op_namelen = namelen;
98 op_data->op_mod_time = CURRENT_TIME;
99 op_data->op_data = NULL;
102 void llu_finish_md_op_data(struct md_op_data *op_data)
104 OBD_FREE_PTR(op_data);
107 void obdo_refresh_inode(struct inode *dst,
111 struct intnl_stat *st = llu_i2stat(dst);
112 valid &= src->o_valid;
114 if (valid & (OBD_MD_FLCTIME | OBD_MD_FLMTIME))
115 CDEBUG(D_INODE,"valid "LPX64", cur time "CFS_TIME_T"/"CFS_TIME_T
117 src->o_valid, LTIME_S(st->st_mtime),
118 LTIME_S(st->st_ctime),
119 (long)src->o_mtime, (long)src->o_ctime);
121 if (valid & OBD_MD_FLATIME && src->o_atime > LTIME_S(st->st_atime))
122 LTIME_S(st->st_atime) = src->o_atime;
124 /* mtime is always updated with ctime, but can be set in past.
125 As write and utime(2) may happen within 1 second, and utime's
126 mtime has a priority over write's one, leave mtime from mds
127 for the same ctimes. */
128 if (valid & OBD_MD_FLCTIME && src->o_ctime > LTIME_S(st->st_ctime)) {
129 LTIME_S(st->st_ctime) = src->o_ctime;
130 if (valid & OBD_MD_FLMTIME)
131 LTIME_S(st->st_mtime) = src->o_mtime;
133 if (valid & OBD_MD_FLSIZE && src->o_size > st->st_size)
134 st->st_size = src->o_size;
135 /* optimum IO size */
136 if (valid & OBD_MD_FLBLKSZ)
137 st->st_blksize = src->o_blksize;
138 /* allocation of space */
139 if (valid & OBD_MD_FLBLOCKS && src->o_blocks > st->st_blocks)
140 st->st_blocks = src->o_blocks;
143 int llu_local_open(struct llu_inode_info *lli, struct lookup_intent *it)
145 struct ptlrpc_request *req = it->d.lustre.it_data;
146 struct ll_file_data *fd;
147 struct mdt_body *body;
150 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
151 LASSERT(body != NULL);
153 /* already opened? */
154 if (lli->lli_open_count++)
157 LASSERT(!lli->lli_file_data);
159 OBD_ALLOC(fd, sizeof(*fd));
160 /* We can't handle this well without reorganizing ll_file_open and
161 * ll_md_close, so don't even try right now. */
164 memcpy(&fd->fd_mds_och.och_fh, &body->handle, sizeof(body->handle));
165 fd->fd_mds_och.och_magic = OBD_CLIENT_HANDLE_MAGIC;
166 fd->fd_mds_och.och_fid = lli->lli_fid;
167 lli->lli_file_data = fd;
169 md_set_open_replay_data(lli->lli_sbi->ll_md_exp,
170 &fd->fd_mds_och, it->d.lustre.it_data);
175 int llu_iop_open(struct pnode *pnode, int flags, mode_t mode)
177 struct inode *inode = pnode->p_base->pb_ino;
178 struct llu_inode_info *lli = llu_i2info(inode);
179 struct intnl_stat *st = llu_i2stat(inode);
180 struct ll_file_data *fd;
181 struct ptlrpc_request *request;
182 struct lookup_intent *it;
183 struct lov_stripe_md *lsm;
187 liblustre_wait_event(0);
189 /* don't do anything for '/' */
190 if (llu_is_root_inode(inode))
193 CDEBUG(D_VFSTRACE, "VFS Op:inode=%llu\n", (long long)st->st_ino);
194 LL_GET_INTENT(inode, it);
196 if (!it->d.lustre.it_disposition) {
200 rc = it_open_error(DISP_OPEN_OPEN, it);
202 GOTO(out_release, rc);
204 rc = llu_local_open(lli, it);
208 if (!S_ISREG(st->st_mode))
209 GOTO(out_release, rc = 0);
211 fd = lli->lli_file_data;
215 if (fd->fd_flags & O_LOV_DELAY_CREATE) {
216 CDEBUG(D_INODE, "object creation was delayed\n");
217 GOTO(out_release, rc);
220 fd->fd_flags &= ~O_LOV_DELAY_CREATE;
222 lli->lli_open_flags = flags & ~(O_CREAT | O_EXCL | O_TRUNC);
225 request = it->d.lustre.it_data;
226 ptlrpc_req_finished(request);
228 it->it_op_release(it);
229 OBD_FREE(it, sizeof(*it));
231 /* libsysio hasn't done anything for O_TRUNC. here we
232 * simply simulate it as open(...); truncate(...); */
233 if (rc == 0 && (flags & O_TRUNC) && S_ISREG(st->st_mode)) {
236 memset(&attr, 0, sizeof(attr));
238 attr.ia_valid |= ATTR_SIZE | ATTR_RAW;
239 rc = llu_setattr_raw(inode, &attr);
241 CERROR("error %d truncate in open()\n", rc);
244 liblustre_wait_event(0);
248 int llu_objects_destroy(struct ptlrpc_request *req, struct inode *dir)
250 struct mdt_body *body;
251 struct lov_mds_md *eadata;
252 struct lov_stripe_md *lsm = NULL;
253 struct obd_trans_info oti = { 0 };
258 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
260 if (!(body->valid & OBD_MD_FLEASIZE))
263 if (body->eadatasize == 0) {
264 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
265 GOTO(out, rc = -EPROTO);
268 /* The MDS sent back the EA because we unlinked the last reference
269 * to this file. Use this EA to unlink the objects on the OST.
270 * It's opaque so we don't swab here; we leave it to obd_unpackmd() to
271 * check it is complete and sensible. */
272 eadata = req_capsule_server_sized_get(&req->rq_pill, &RMF_MDT_MD,
275 LASSERT(eadata != NULL);
277 rc = obd_unpackmd(llu_i2obdexp(dir), &lsm, eadata,body->eadatasize);
279 CERROR("obd_unpackmd: %d\n", rc);
282 LASSERT(rc >= sizeof(*lsm));
286 GOTO(out_free_memmd, rc = -ENOMEM);
288 oa->o_id = lsm->lsm_object_id;
289 oa->o_gr = lsm->lsm_object_gr;
290 oa->o_mode = body->mode & S_IFMT;
291 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLGROUP;
293 if (body->valid & OBD_MD_FLCOOKIE) {
294 oa->o_valid |= OBD_MD_FLCOOKIE;
296 req_capsule_server_sized_get(&req->rq_pill,
298 sizeof(struct llog_cookie) *
299 lsm->lsm_stripe_count);
300 if (oti.oti_logcookies == NULL) {
301 oa->o_valid &= ~OBD_MD_FLCOOKIE;
302 body->valid &= ~OBD_MD_FLCOOKIE;
306 rc = obd_destroy(llu_i2obdexp(dir), oa, lsm, &oti, NULL);
309 CERROR("obd destroy objid 0x"LPX64" error %d\n",
310 lsm->lsm_object_id, rc);
312 obd_free_memmd(llu_i2obdexp(dir), &lsm);
317 int llu_sizeonmds_update(struct inode *inode, struct md_open_data *mod,
318 struct lustre_handle *fh, __u64 ioepoch)
320 struct llu_inode_info *lli = llu_i2info(inode);
321 struct llu_sb_info *sbi = llu_i2sbi(inode);
322 struct md_op_data op_data = {{ 0 }};
327 LASSERT(!(lli->lli_flags & LLIF_MDS_SIZE_LOCK));
328 LASSERT(sbi->ll_lco.lco_flags & OBD_CONNECT_SOM);
330 rc = llu_inode_getattr(inode, &oa);
333 CDEBUG(D_INODE, "objid "LPX64" is already destroyed\n",
334 lli->lli_smd->lsm_object_id);
336 CERROR("inode_getattr failed (%d): unable to send a "
337 "Size-on-MDS attribute update for inode %llu/%lu\n",
338 rc, (long long)llu_i2stat(inode)->st_ino,
339 lli->lli_st_generation);
343 md_from_obdo(&op_data, &oa, oa.o_valid);
344 memcpy(&op_data.op_handle, fh, sizeof(*fh));
345 op_data.op_ioepoch = ioepoch;
346 op_data.op_flags |= MF_SOM_CHANGE;
348 rc = llu_md_setattr(inode, &op_data, &mod);
352 int llu_md_close(struct obd_export *md_exp, struct inode *inode)
354 struct llu_inode_info *lli = llu_i2info(inode);
355 struct ll_file_data *fd = lli->lli_file_data;
356 struct ptlrpc_request *req = NULL;
357 struct obd_client_handle *och = &fd->fd_mds_och;
358 struct intnl_stat *st = llu_i2stat(inode);
359 struct md_op_data op_data = { { 0 } };
363 /* clear group lock, if present */
364 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
365 struct lov_stripe_md *lsm = llu_i2info(inode)->lli_smd;
366 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
367 rc = llu_extent_unlock(fd, inode, lsm, LCK_GROUP,
371 op_data.op_attr.ia_valid = ATTR_MODE | ATTR_ATIME_SET |
372 ATTR_MTIME_SET | ATTR_CTIME_SET;
374 if (fd->fd_flags & FMODE_WRITE) {
375 struct llu_sb_info *sbi = llu_i2sbi(inode);
376 if (!(sbi->ll_lco.lco_flags & OBD_CONNECT_SOM) ||
377 !S_ISREG(llu_i2stat(inode)->st_mode)) {
378 op_data.op_attr.ia_valid |= ATTR_SIZE | ATTR_BLOCKS;
380 /* Inode cannot be dirty. Close the epoch. */
381 op_data.op_flags |= MF_EPOCH_CLOSE;
382 /* XXX: Send CHANGE flag only if Size-on-MDS inode attributes
383 * are really changed. */
384 op_data.op_flags |= MF_SOM_CHANGE;
386 /* Pack Size-on-MDS attributes if we are in IO epoch and
387 * attributes are valid. */
388 LASSERT(!(lli->lli_flags & LLIF_MDS_SIZE_LOCK));
389 if (!llu_local_size(inode))
390 op_data.op_attr.ia_valid |=
391 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
394 op_data.op_fid1 = lli->lli_fid;
395 op_data.op_attr.ia_atime = st->st_atime;
396 op_data.op_attr.ia_mtime = st->st_mtime;
397 op_data.op_attr.ia_ctime = st->st_ctime;
398 op_data.op_attr.ia_size = st->st_size;
399 op_data.op_attr_blocks = st->st_blocks;
400 op_data.op_attr.ia_attr_flags = lli->lli_st_flags;
401 op_data.op_ioepoch = lli->lli_ioepoch;
402 memcpy(&op_data.op_handle, &och->och_fh, sizeof(op_data.op_handle));
404 rc = md_close(md_exp, &op_data, och->och_mod, &req);
409 /* We are the last writer, so the MDS has instructed us to get
410 * the file size and any write cookies, then close again. */
411 LASSERT(fd->fd_flags & FMODE_WRITE);
412 rc = llu_sizeonmds_update(inode, och->och_mod, &och->och_fh,
415 CERROR("inode %llu mdc Size-on-MDS update failed: "
416 "rc = %d\n", (long long)st->st_ino, rc);
420 CERROR("inode %llu close failed: rc %d\n",
421 (long long)st->st_ino, rc);
423 rc = llu_objects_destroy(req, inode);
425 CERROR("inode %llu ll_objects destroy: rc = %d\n",
426 (long long)st->st_ino, rc);
430 ptlrpc_close_replay_seq(req);
431 md_clear_open_replay_data(md_exp, och);
432 ptlrpc_req_finished(req);
433 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
434 lli->lli_file_data = NULL;
435 OBD_FREE(fd, sizeof(*fd));
440 int llu_file_release(struct inode *inode)
442 struct ll_file_data *fd;
443 struct llu_sb_info *sbi = llu_i2sbi(inode);
444 struct llu_inode_info *lli = llu_i2info(inode);
448 CDEBUG(D_VFSTRACE, "VFS Op:inode=%llu/%lu\n",
449 (long long)llu_i2stat(inode)->st_ino, lli->lli_st_generation);
451 if (llu_is_root_inode(inode))
454 /* still opened by others? */
455 if (--lli->lli_open_count)
458 fd = lli->lli_file_data;
459 if (!fd) /* no process opened the file after an mcreate */
462 rc2 = llu_md_close(sbi->ll_md_exp, inode);
470 * libsysio require us return 0
472 int llu_iop_close(struct inode *inode)
476 liblustre_wait_event(0);
478 rc = llu_file_release(inode);
480 CERROR("file close error %d\n", rc);
482 /* if open count == 0 && stale_flag is set, should we
483 * remove the inode immediately? */
484 liblustre_wait_idle();
488 _SYSIO_OFF_T llu_iop_pos(struct inode *ino, _SYSIO_OFF_T off)
492 liblustre_wait_event(0);
494 if (off < 0 || off > ll_file_maxbytes(ino))
500 /* this isn't where truncate starts. roughly:
501 * llu_iop_{open,setattr}->llu_setattr_raw->llu_vmtruncate->llu_truncate
502 * we grab the lock back in setattr_raw to avoid races. */
503 static void llu_truncate(struct inode *inode, obd_flag flags)
505 struct llu_inode_info *lli = llu_i2info(inode);
506 struct intnl_stat *st = llu_i2stat(inode);
507 struct obd_info oinfo = { { { 0 } } };
508 struct obdo oa = { 0 };
511 CDEBUG(D_VFSTRACE, "VFS Op:inode=%llu/%lu(%p) to %llu\n",
512 (long long)st->st_ino, lli->lli_st_generation, inode,
513 (long long)st->st_size);
516 CDEBUG(D_INODE, "truncate on inode %llu with no objects\n",
517 (long long)st->st_ino);
522 oinfo.oi_md = lli->lli_smd;
523 oinfo.oi_policy.l_extent.start = st->st_size;
524 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
526 oa.o_id = lli->lli_smd->lsm_object_id;
527 oa.o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS;
528 oa.o_flags = flags; /* We don't actually want to copy inode flags */
530 obdo_from_inode(&oa, inode,
531 OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLATIME |
532 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
534 obd_adjust_kms(llu_i2obdexp(inode), lli->lli_smd, st->st_size, 1);
536 CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
537 oa.o_id, (long long)st->st_size);
539 /* truncate == punch from new size to absolute end of file */
540 rc = obd_punch_rqset(llu_i2obdexp(inode), &oinfo, NULL);
542 CERROR("obd_truncate fails (%d) ino %llu\n",
543 rc, (long long)st->st_ino);
545 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
546 OBD_MD_FLATIME | OBD_MD_FLMTIME |
553 int llu_vmtruncate(struct inode * inode, loff_t offset, obd_flag flags)
555 llu_i2stat(inode)->st_size = offset;
558 * llu_truncate() is only called from this
559 * point. llu_vmtruncate/llu_truncate split exists to mimic the
560 * structure of Linux VFS truncate code path.
563 llu_truncate(inode, flags);