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);
80 memset(op_data, 0, sizeof(*op_data));
83 ll_i2gids(op_data->suppgids, i1, i2);
84 op_data->fid1 = *ll_inode2fid(i1);
86 ll_i2gids(op_data->suppgids, i2, i1);
87 op_data->fid1 = *ll_inode2fid(i2);
91 op_data->fid2 = *ll_inode2fid(i2);
94 op_data->namelen = namelen;
95 op_data->create_mode = mode;
96 op_data->mod_time = CURRENT_TIME;
99 void llu_finish_md_op_data(struct md_op_data *op_data)
101 OBD_FREE_PTR(op_data);
104 void obdo_refresh_inode(struct inode *dst,
108 struct intnl_stat *st = llu_i2stat(dst);
109 valid &= src->o_valid;
111 if (valid & (OBD_MD_FLCTIME | OBD_MD_FLMTIME))
112 CDEBUG(D_INODE,"valid "LPX64", cur time %lu/%lu, new %lu/%lu\n",
113 src->o_valid, LTIME_S(st->st_mtime),
114 LTIME_S(st->st_ctime),
115 (long)src->o_mtime, (long)src->o_ctime);
117 if (valid & OBD_MD_FLATIME && src->o_atime > LTIME_S(st->st_atime))
118 LTIME_S(st->st_atime) = src->o_atime;
120 /* mtime is always updated with ctime, but can be set in past.
121 As write and utime(2) may happen within 1 second, and utime's
122 mtime has a priority over write's one, leave mtime from mds
123 for the same ctimes. */
124 if (valid & OBD_MD_FLCTIME && src->o_ctime > LTIME_S(st->st_ctime)) {
125 LTIME_S(st->st_ctime) = src->o_ctime;
126 if (valid & OBD_MD_FLMTIME)
127 LTIME_S(st->st_mtime) = src->o_mtime;
129 if (valid & OBD_MD_FLSIZE && src->o_size > st->st_size)
130 st->st_size = src->o_size;
131 /* optimum IO size */
132 if (valid & OBD_MD_FLBLKSZ)
133 st->st_blksize = src->o_blksize;
134 /* allocation of space */
135 if (valid & OBD_MD_FLBLOCKS && src->o_blocks > st->st_blocks)
136 st->st_blocks = src->o_blocks;
139 int llu_local_open(struct llu_inode_info *lli, struct lookup_intent *it)
141 struct ptlrpc_request *req = it->d.lustre.it_data;
142 struct ll_file_data *fd;
143 struct mdt_body *body;
146 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
147 LASSERT(body != NULL); /* reply already checked out */
148 LASSERT_REPSWABBED(req, DLM_REPLY_REC_OFF); /* and swabbed down */
150 /* already opened? */
151 if (lli->lli_open_count++)
154 LASSERT(!lli->lli_file_data);
156 OBD_ALLOC(fd, sizeof(*fd));
157 /* We can't handle this well without reorganizing ll_file_open and
158 * ll_md_close, so don't even try right now. */
161 memcpy(&fd->fd_mds_och.och_fh, &body->handle, sizeof(body->handle));
162 fd->fd_mds_och.och_magic = OBD_CLIENT_HANDLE_MAGIC;
163 lli->lli_file_data = fd;
165 md_set_open_replay_data(lli->lli_sbi->ll_md_exp,
166 &fd->fd_mds_och, it->d.lustre.it_data);
171 int llu_iop_open(struct pnode *pnode, int flags, mode_t mode)
173 struct inode *inode = pnode->p_base->pb_ino;
174 struct llu_inode_info *lli = llu_i2info(inode);
175 struct intnl_stat *st = llu_i2stat(inode);
176 struct ll_file_data *fd;
177 struct ptlrpc_request *request;
178 struct lookup_intent *it;
179 struct lov_stripe_md *lsm;
183 liblustre_wait_event(0);
185 /* don't do anything for '/' */
186 if (llu_is_root_inode(inode))
189 CDEBUG(D_VFSTRACE, "VFS Op:inode=%llu\n", (long long)st->st_ino);
190 LL_GET_INTENT(inode, it);
192 if (!it->d.lustre.it_disposition) {
196 rc = it_open_error(DISP_OPEN_OPEN, it);
198 GOTO(out_release, rc);
200 rc = llu_local_open(lli, it);
204 if (!S_ISREG(st->st_mode))
205 GOTO(out_release, rc = 0);
207 fd = lli->lli_file_data;
211 if (fd->fd_flags & O_LOV_DELAY_CREATE) {
212 CDEBUG(D_INODE, "object creation was delayed\n");
213 GOTO(out_release, rc);
216 fd->fd_flags &= ~O_LOV_DELAY_CREATE;
218 lli->lli_open_flags = flags & ~(O_CREAT | O_EXCL | O_TRUNC);
221 request = it->d.lustre.it_data;
222 ptlrpc_req_finished(request);
224 it->it_op_release(it);
225 OBD_FREE(it, sizeof(*it));
227 /* libsysio hasn't done anything for O_TRUNC. here we
228 * simply simulate it as open(...); truncate(...); */
229 if (rc == 0 && (flags & O_TRUNC) && S_ISREG(st->st_mode)) {
232 memset(&attr, 0, sizeof(attr));
234 attr.ia_valid |= ATTR_SIZE | ATTR_RAW;
235 rc = llu_setattr_raw(inode, &attr);
237 CERROR("error %d truncate in open()\n", rc);
240 liblustre_wait_event(0);
244 int llu_objects_destroy(struct ptlrpc_request *request, struct inode *dir)
246 struct mdt_body *body;
247 struct lov_mds_md *eadata;
248 struct lov_stripe_md *lsm = NULL;
249 struct obd_trans_info oti = { 0 };
254 /* req is swabbed so this is safe */
255 body = lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
257 if (!(body->valid & OBD_MD_FLEASIZE))
260 if (body->eadatasize == 0) {
261 CERROR("OBD_MD_FLEASIZE set but eadatasize zero\n");
262 GOTO(out, rc = -EPROTO);
265 /* The MDS sent back the EA because we unlinked the last reference
266 * to this file. Use this EA to unlink the objects on the OST.
267 * It's opaque so we don't swab here; we leave it to obd_unpackmd() to
268 * check it is complete and sensible. */
269 eadata = lustre_swab_repbuf(request, REPLY_REC_OFF+1, body->eadatasize,
271 LASSERT(eadata != NULL);
272 if (eadata == NULL) {
273 CERROR("Can't unpack MDS EA data\n");
274 GOTO(out, rc = -EPROTO);
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 lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF + 2,
297 sizeof(struct llog_cookie) *
298 lsm->lsm_stripe_count);
299 if (oti.oti_logcookies == NULL) {
300 oa->o_valid &= ~OBD_MD_FLCOOKIE;
301 body->valid &= ~OBD_MD_FLCOOKIE;
305 rc = obd_destroy(llu_i2obdexp(dir), oa, lsm, &oti, NULL, NULL);
308 CERROR("obd destroy objid 0x"LPX64" error %d\n",
309 lsm->lsm_object_id, rc);
311 obd_free_memmd(llu_i2obdexp(dir), &lsm);
316 int llu_sizeonmds_update(struct inode *inode, struct lustre_handle *fh)
318 struct llu_inode_info *lli = llu_i2info(inode);
319 struct md_op_data op_data;
324 LASSERT(!(lli->lli_flags & LLIF_MDS_SIZE_LOCK));
326 rc = llu_inode_getattr(inode, &oa);
328 CERROR("inode_getattr failed (%d): unable to send a "
329 "Size-on-MDS attribute update for inode %llu/%lu\n",
330 rc, (long long)llu_i2stat(inode)->st_ino,
331 lli->lli_st_generation);
335 md_from_obdo(&op_data, &oa, oa.o_valid);
336 memcpy(&op_data.handle, fh, sizeof(*fh));
337 op_data.flags |= MF_SOM_CHANGE;
339 rc = llu_md_setattr(inode, &op_data);
343 int llu_md_close(struct obd_export *md_exp, struct inode *inode)
345 struct llu_inode_info *lli = llu_i2info(inode);
346 struct ll_file_data *fd = lli->lli_file_data;
347 struct ptlrpc_request *req = NULL;
348 struct obd_client_handle *och = &fd->fd_mds_och;
349 struct intnl_stat *st = llu_i2stat(inode);
350 struct md_op_data op_data = { { 0 } };
354 /* clear group lock, if present */
355 if (fd->fd_flags & LL_FILE_GROUP_LOCKED) {
356 struct lov_stripe_md *lsm = llu_i2info(inode)->lli_smd;
357 fd->fd_flags &= ~(LL_FILE_GROUP_LOCKED|LL_FILE_IGNORE_LOCK);
358 rc = llu_extent_unlock(fd, inode, lsm, LCK_GROUP,
362 op_data.attr.ia_valid = ATTR_MODE | ATTR_ATIME_SET |
363 ATTR_MTIME_SET | ATTR_CTIME_SET;
365 if (!S_ISREG(llu_i2stat(inode)->st_mode)) {
366 op_data.attr.ia_valid |= ATTR_SIZE | ATTR_BLOCKS;
368 /* Inode cannot be dirty. Close the epoch. */
369 op_data.flags |= MF_EPOCH_CLOSE;
370 /* XXX: Send CHANGE flag only if Size-on-MDS inode attributes
371 * are really changed. */
372 op_data.flags |= MF_SOM_CHANGE;
374 /* Pack Size-on-MDS attrinodes if valid. */
375 if ((lli->lli_flags & LLIF_MDS_SIZE_LOCK) ||
376 !llu_local_size(inode))
377 op_data.attr.ia_valid |=
378 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
381 op_data.fid1 = lli->lli_fid;
382 op_data.attr.ia_atime = st->st_atime;
383 op_data.attr.ia_mtime = st->st_mtime;
384 op_data.attr.ia_ctime = st->st_ctime;
385 op_data.attr.ia_size = st->st_size;
386 op_data.attr_blocks = st->st_blocks;
387 op_data.attr.ia_attr_flags = lli->lli_st_flags;
388 op_data.ioepoch = lli->lli_ioepoch;
389 memcpy(&op_data.handle, &och->och_fh, sizeof(op_data.handle));
391 rc = md_close(md_exp, &op_data, och, &req);
393 /* We are the last writer, so the MDS has instructed us to get
394 * the file size and any write cookies, then close again. */
395 rc = llu_sizeonmds_update(inode, &och->och_fh);
397 CERROR("inode %llu mdc Size-on-MDS update failed: "
398 "rc = %d\n", (long long)st->st_ino, rc);
402 CERROR("inode %llu close failed: rc %d\n",
403 (long long)st->st_ino, rc);
405 rc = llu_objects_destroy(req, inode);
407 CERROR("inode %llu ll_objects destroy: rc = %d\n",
408 (long long)st->st_ino, rc);
411 md_clear_open_replay_data(md_exp, och);
412 ptlrpc_req_finished(req);
413 och->och_fh.cookie = DEAD_HANDLE_MAGIC;
414 lli->lli_file_data = NULL;
415 OBD_FREE(fd, sizeof(*fd));
420 int llu_file_release(struct inode *inode)
422 struct ll_file_data *fd;
423 struct llu_sb_info *sbi = llu_i2sbi(inode);
424 struct llu_inode_info *lli = llu_i2info(inode);
428 CDEBUG(D_VFSTRACE, "VFS Op:inode=%llu/%lu\n",
429 (long long)llu_i2stat(inode)->st_ino, lli->lli_st_generation);
431 if (llu_is_root_inode(inode))
434 /* still opened by others? */
435 if (--lli->lli_open_count)
438 fd = lli->lli_file_data;
439 if (!fd) /* no process opened the file after an mcreate */
442 rc2 = llu_md_close(sbi->ll_md_exp, inode);
450 * libsysio require us return 0
452 int llu_iop_close(struct inode *inode)
456 liblustre_wait_event(0);
458 rc = llu_file_release(inode);
460 CERROR("file close error %d\n", rc);
462 /* if open count == 0 && stale_flag is set, should we
463 * remove the inode immediately? */
464 liblustre_wait_event(0);
468 _SYSIO_OFF_T llu_iop_pos(struct inode *ino, _SYSIO_OFF_T off)
472 liblustre_wait_event(0);
474 if (off < 0 || off > ll_file_maxbytes(ino))
480 /* this isn't where truncate starts. roughly:
481 * llu_iop_{open,setattr}->llu_setattr_raw->llu_vmtruncate->llu_truncate
482 * we grab the lock back in setattr_raw to avoid races. */
483 static void llu_truncate(struct inode *inode, obd_flag flags)
485 struct llu_inode_info *lli = llu_i2info(inode);
486 struct intnl_stat *st = llu_i2stat(inode);
487 struct obd_info oinfo = { { { 0 } } };
488 struct obdo oa = { 0 };
491 CDEBUG(D_VFSTRACE, "VFS Op:inode=%llu/%lu(%p) to %llu\n",
492 (long long)st->st_ino, lli->lli_st_generation, inode,
493 (long long)st->st_size);
496 CDEBUG(D_INODE, "truncate on inode %llu with no objects\n",
497 (long long)st->st_ino);
502 oinfo.oi_md = lli->lli_smd;
503 oinfo.oi_policy.l_extent.start = st->st_size;
504 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
506 oa.o_id = lli->lli_smd->lsm_object_id;
507 oa.o_valid = OBD_MD_FLID | OBD_MD_FLFLAGS;
508 oa.o_flags = flags; /* We don't actually want to copy inode flags */
510 obdo_from_inode(&oa, inode,
511 OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLATIME |
512 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
514 obd_adjust_kms(llu_i2obdexp(inode), lli->lli_smd, st->st_size, 1);
516 CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
517 oa.o_id, (long long)st->st_size);
519 /* truncate == punch from new size to absolute end of file */
520 rc = obd_punch_rqset(llu_i2obdexp(inode), &oinfo, NULL);
522 CERROR("obd_truncate fails (%d) ino %llu\n",
523 rc, (long long)st->st_ino);
525 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
526 OBD_MD_FLATIME | OBD_MD_FLMTIME |
533 int llu_vmtruncate(struct inode * inode, loff_t offset, obd_flag flags)
535 llu_i2stat(inode)->st_size = offset;
538 * llu_truncate() is only called from this
539 * point. llu_vmtruncate/llu_truncate split exists to mimic the
540 * structure of Linux VFS truncate code path.
543 llu_truncate(inode, flags);