1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Directory code for lustre client.
42 #include <linux/pagemap.h>
44 #include <linux/version.h>
45 #include <linux/smp_lock.h>
46 #include <asm/uaccess.h>
47 #include <linux/buffer_head.h> // for wait_on_buffer
49 #define DEBUG_SUBSYSTEM S_LLITE
51 #include <obd_support.h>
52 #include <obd_class.h>
53 #include <lustre_lib.h>
54 #include <lustre/lustre_idl.h>
55 #include <lustre_lite.h>
56 #include <lustre_dlm.h>
57 #include <lustre_fid.h>
58 #include "llite_internal.h"
60 #ifndef HAVE_PAGE_CHECKED
61 #ifdef HAVE_PG_FS_MISC
62 #define PageChecked(page) test_bit(PG_fs_misc, &(page)->flags)
63 #define SetPageChecked(page) set_bit(PG_fs_misc, &(page)->flags)
65 #error PageChecked or PageFsMisc not defined in kernel
70 * (new) readdir implementation overview.
72 * Original lustre readdir implementation cached exact copy of raw directory
73 * pages on the client. These pages were indexed in client page cache by
74 * logical offset in the directory file. This design, while very simple and
75 * intuitive had some inherent problems:
77 * . it implies that byte offset to the directory entry serves as a
78 * telldir(3)/seekdir(3) cookie, but that offset is not stable: in
79 * ext3/htree directory entries may move due to splits, and more
82 * . it is incompatible with the design of split directories for cmd3,
83 * that assumes that names are distributed across nodes based on their
84 * hash, and so readdir should be done in hash order.
86 * New readdir implementation does readdir in hash order, and uses hash of a
87 * file name as a telldir/seekdir cookie. This led to number of complications:
89 * . hash is not unique, so it cannot be used to index cached directory
90 * pages on the client (note, that it requires a whole pageful of hash
91 * collided entries to cause two pages to have identical hashes);
93 * . hash is not unique, so it cannot, strictly speaking, be used as an
94 * entry cookie. ext3/htree has the same problem and lustre implementation
95 * mimics their solution: seekdir(hash) positions directory at the first
96 * entry with the given hash.
102 * Client caches directory pages using hash of the first entry as an index. As
103 * noted above hash is not unique, so this solution doesn't work as is:
104 * special processing is needed for "page hash chains" (i.e., sequences of
105 * pages filled with entries all having the same hash value).
107 * First, such chains have to be detected. To this end, server returns to the
108 * client the hash of the first entry on the page next to one returned. When
109 * client detects that this hash is the same as hash of the first entry on the
110 * returned page, page hash collision has to be handled. Pages in the
111 * hash chain, except first one, are termed "overflow pages".
113 * Solution to index uniqueness problem is to not cache overflow
114 * pages. Instead, when page hash collision is detected, all overflow pages
115 * from emerging chain are immediately requested from the server and placed in
116 * a special data structure (struct ll_dir_chain). This data structure is used
117 * by ll_readdir() to process entries from overflow pages. When readdir
118 * invocation finishes, overflow pages are discarded. If page hash collision
119 * chain weren't completely processed, next call to readdir will again detect
120 * page hash collision, again read overflow pages in, process next portion of
121 * entries and again discard the pages. This is not as wasteful as it looks,
122 * because, given reasonable hash, page hash collisions are extremely rare.
124 * 1. directory positioning
126 * When seekdir(hash) is called, original
137 * identification of and access to overflow pages
147 /* returns the page unlocked, but with a reference */
148 static int ll_dir_readpage(struct file *file, struct page *page)
150 struct inode *inode = page->mapping->host;
151 struct ptlrpc_request *request;
152 struct mdt_body *body;
159 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
161 hash = fd->fd_dir.lfd_next;
163 struct ll_inode_info *lli = ll_i2info(inode);
165 cfs_spin_lock(&lli->lli_sa_lock);
167 LASSERT(lli->lli_sai->sai_pid == cfs_curproc_pid());
169 LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
170 hash = lli->lli_sa_pos;
171 cfs_spin_unlock(&lli->lli_sa_lock);
173 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
174 inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
176 oc = ll_mdscapa_get(inode);
177 rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
178 oc, hash, page, &request);
181 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
182 /* Checked by mdc_readpage() */
183 LASSERT(body != NULL);
185 if (body->valid & OBD_MD_FLSIZE)
186 cl_isize_write(inode, body->size);
187 SetPageUptodate(page);
189 ptlrpc_req_finished(request);
196 struct address_space_operations ll_dir_aops = {
197 .readpage = ll_dir_readpage,
200 static void ll_check_page(struct inode *dir, struct page *page)
202 /* XXX: check page format later */
203 SetPageChecked(page);
206 static void ll_release_page(struct page *page, __u64 hash,
207 __u64 start, __u64 end)
211 if (likely(page->mapping != NULL)) {
212 ll_truncate_complete_page(page);
216 CWARN("NULL mapping page %p, truncated by others: "
217 "hash("LPX64") | start("LPX64") | end("LPX64")\n",
218 page, hash, start, end);
220 page_cache_release(page);
224 * Find, kmap and return page that contains given hash.
226 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
227 __u64 *start, __u64 *end)
229 struct address_space *mapping = dir->i_mapping;
231 * Complement of hash is used as an index so that
232 * radix_tree_gang_lookup() can be used to find a page with starting
233 * hash _smaller_ than one we are looking for.
235 unsigned long offset = hash_x_index(*hash);
239 TREE_READ_LOCK_IRQ(mapping);
240 found = radix_tree_gang_lookup(&mapping->page_tree,
241 (void **)&page, offset, 1);
243 struct lu_dirpage *dp;
245 page_cache_get(page);
246 TREE_READ_UNLOCK_IRQ(mapping);
248 * In contrast to find_lock_page() we are sure that directory
249 * page cannot be truncated (while DLM lock is held) and,
250 * hence, can avoid restart.
252 * In fact, page cannot be locked here at all, because
253 * ll_dir_readpage() does synchronous io.
256 if (PageUptodate(page)) {
258 #if BITS_PER_LONG == 32
259 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
260 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
263 *start = le64_to_cpu(dp->ldp_hash_start);
264 *end = le64_to_cpu(dp->ldp_hash_end);
266 LASSERTF(*start <= *hash, "start = "LPX64",end = "
267 LPX64",hash = "LPX64"\n", *start, *end, *hash);
268 if (*hash > *end || (*end != *start && *hash == *end)) {
269 ll_release_page(page, *hash, *start, *end);
273 page_cache_release(page);
274 page = ERR_PTR(-EIO);
278 TREE_READ_UNLOCK_IRQ(mapping);
284 struct page *ll_get_dir_page(struct file *filp, struct inode *dir, __u64 hash,
285 int exact, struct ll_dir_chain *chain)
287 ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
288 struct address_space *mapping = dir->i_mapping;
289 struct lustre_handle lockh;
290 struct lu_dirpage *dp;
297 struct ll_inode_info *lli = ll_i2info(dir);
300 rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
301 ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
303 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
304 ll_md_blocking_ast, ldlm_completion_ast,
306 struct lookup_intent it = { .it_op = IT_READDIR };
307 struct ptlrpc_request *request;
308 struct md_op_data *op_data;
310 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
311 LUSTRE_OPC_ANY, NULL);
313 return (void *)op_data;
315 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
316 op_data, &lockh, NULL, 0, NULL, 0);
318 ll_finish_md_op_data(op_data);
320 request = (struct ptlrpc_request *)it.d.lustre.it_data;
322 ptlrpc_req_finished(request);
324 CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
325 PFID(ll_inode2fid(dir)), hash, rc);
329 /* for cross-ref object, l_ast_data of the lock may not be set,
330 * we reset it here */
331 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
334 ldlm_lock_dump_handle(D_OTHER, &lockh);
336 cfs_down(&lli->lli_readdir_sem);
337 page = ll_dir_page_locate(dir, &lhash, &start, &end);
339 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
340 PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
341 GOTO(out_unlock, page);
346 * XXX nikita: not entirely correct handling of a corner case:
347 * suppose hash chain of entries with hash value HASH crosses
348 * border between pages P0 and P1. First both P0 and P1 are
349 * cached, seekdir() is called for some entry from the P0 part
350 * of the chain. Later P0 goes out of cache. telldir(HASH)
351 * happens and finds P1, as it starts with matching hash
352 * value. Remaining entries from P0 part of the chain are
353 * skipped. (Is that really a bug?)
355 * Possible solutions: 0. don't cache P1 is such case, handle
356 * it as an "overflow" page. 1. invalidate all pages at
357 * once. 2. use HASH|1 as an index for P1.
359 if (exact && lhash != start) {
361 * readdir asked for a page starting _exactly_ from
362 * given hash, but cache contains stale page, with
363 * entries with smaller hash values. Stale page should
364 * be invalidated, and new one fetched.
366 CDEBUG(D_OTHER, "Stale readpage page %p: "
367 "start = "LPX64",end = "LPX64"hash ="LPX64"\n",
368 page, start, end, lhash);
369 ll_release_page(page, lhash, start, end);
371 GOTO(hash_collision, page);
375 page = read_cache_page(mapping, hash_x_index(hash),
376 (filler_t*)mapping->a_ops->readpage, filp);
378 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
379 PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
380 GOTO(out_unlock, page);
385 if (!PageUptodate(page)) {
386 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
387 PFID(ll_inode2fid(dir)), hash, -5);
390 if (!PageChecked(page))
391 ll_check_page(dir, page);
392 if (PageError(page)) {
393 CERROR("page error: "DFID" at "LPU64": rc %d\n",
394 PFID(ll_inode2fid(dir)), hash, -5);
398 dp = page_address(page);
399 #if BITS_PER_LONG == 32
400 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
401 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
404 start = le64_to_cpu(dp->ldp_hash_start);
405 end = le64_to_cpu(dp->ldp_hash_end);
409 LASSERT(start == lhash);
410 CWARN("Page-wide hash collision: "LPU64"\n", end);
411 #if BITS_PER_LONG == 32
412 CWARN("Real page-wide hash collision at ["LPU64" "LPU64"] with "
413 "hash "LPU64"\n", le64_to_cpu(dp->ldp_hash_start),
414 le64_to_cpu(dp->ldp_hash_end), hash);
417 * Fetch whole overflow chain...
424 cfs_up(&lli->lli_readdir_sem);
425 ldlm_lock_decref(&lockh, mode);
430 page = ERR_PTR(-EIO);
434 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
436 struct inode *inode = filp->f_dentry->d_inode;
437 struct ll_inode_info *info = ll_i2info(inode);
438 struct ll_sb_info *sbi = ll_i2sbi(inode);
439 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
440 __u64 pos = fd->fd_dir.lfd_pos;
442 struct ll_dir_chain chain;
449 need_32bit = ll_need_32bit_api(sbi);
450 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu 32bit_api %d\n",
451 inode->i_ino, inode->i_generation, inode,
452 (unsigned long)pos, i_size_read(inode), need_32bit);
454 if (pos == DIR_END_OFF)
463 ll_dir_chain_init(&chain);
465 fd->fd_dir.lfd_next = pos;
466 page = ll_get_dir_page(filp, inode, pos, 0, &chain);
468 while (rc == 0 && !done) {
469 struct lu_dirpage *dp;
470 struct lu_dirent *ent;
474 * If page is empty (end of directory is reached),
477 __u64 hash = DIR_END_OFF;
480 dp = page_address(page);
481 for (ent = lu_dirent_start(dp); ent != NULL && !done;
482 ent = lu_dirent_next(ent)) {
490 * XXX: implement correct swabbing here.
493 hash = le64_to_cpu(ent->lde_hash);
496 * Skip until we find target hash
501 namelen = le16_to_cpu(ent->lde_namelen);
508 name = ent->lde_name;
509 fid_le_to_cpu(&fid, &ent->lde_fid);
512 ino = cl_fid_build_ino32(&fid);
515 ino = cl_fid_build_ino(&fid);
517 type = ll_dirent_type_get(ent);
518 done = filldir(cookie, name, namelen,
521 next = le64_to_cpu(dp->ldp_hash_end);
525 if (pos == DIR_END_OFF) {
527 * End of directory reached.
530 } else if (1 /* chain is exhausted*/) {
532 * Normal case: continue to the next
535 fd->fd_dir.lfd_next = pos;
536 page = ll_get_dir_page(filp, inode, pos,
540 * go into overflow page.
548 CERROR("error reading dir "DFID" at %lu: rc %d\n",
549 PFID(&info->lli_fid), (unsigned long)pos, rc);
553 fd->fd_dir.lfd_pos = pos;
555 filp->f_pos = pos >> 32;
558 filp->f_version = inode->i_version;
559 touch_atime(filp->f_vfsmnt, filp->f_dentry);
561 ll_dir_chain_fini(&chain);
566 int ll_send_mgc_param(struct obd_export *mgc, char *string)
568 struct mgs_send_param *msp;
575 strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
576 rc = obd_set_info_async(mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
577 sizeof(struct mgs_send_param), msp, NULL);
579 CERROR("Failed to set parameter: %d\n", rc);
585 char *ll_get_fsname(struct inode *inode)
587 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
591 OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
592 len = strlen(lsi->lsi_lmd->lmd_profile);
593 ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
594 if (ptr && (strcmp(ptr, "-client") == 0))
596 strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
602 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
605 struct ll_sb_info *sbi = ll_i2sbi(inode);
606 struct md_op_data *op_data;
607 struct ptlrpc_request *req = NULL;
609 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
610 struct obd_device *mgc = lsi->lsi_mgc;
611 char *fsname = NULL, *param = NULL;
616 * This is coming from userspace, so should be in
617 * local endian. But the MDS would like it in little
618 * endian, so we swab it before we send it.
620 switch (lump->lmm_magic) {
621 case LOV_USER_MAGIC_V1: {
622 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
623 lustre_swab_lov_user_md_v1(lump);
624 lum_size = sizeof(struct lov_user_md_v1);
627 case LOV_USER_MAGIC_V3: {
628 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
629 lustre_swab_lov_user_md_v3(
630 (struct lov_user_md_v3 *)lump);
631 lum_size = sizeof(struct lov_user_md_v3);
635 CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
636 " %#08x != %#08x nor %#08x\n",
637 lump->lmm_magic, LOV_USER_MAGIC_V1,
643 lum_size = sizeof(struct lov_user_md_v1);
646 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
647 LUSTRE_OPC_ANY, NULL);
649 RETURN(PTR_ERR(op_data));
651 /* swabbing is done in lov_setstripe() on server side */
652 rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
653 NULL, 0, &req, NULL);
654 ll_finish_md_op_data(op_data);
655 ptlrpc_req_finished(req);
657 if (rc != -EPERM && rc != -EACCES)
658 CERROR("mdc_setattr fails: rc = %d\n", rc);
661 /* In the following we use the fact that LOV_USER_MAGIC_V1 and
662 LOV_USER_MAGIC_V3 have the same initial fields so we do not
663 need the make the distiction between the 2 versions */
664 if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
665 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
667 /* Get fsname and assume devname to be -MDT0000. */
668 fsname = ll_get_fsname(inode);
669 /* Set root stripesize */
670 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
671 lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
672 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
676 /* Set root stripecount */
677 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
678 lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
679 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
683 /* Set root stripeoffset */
684 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
685 lump ? le16_to_cpu(lump->lmm_stripe_offset) :
686 (typeof(lump->lmm_stripe_offset))(-1));
687 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
692 OBD_FREE(fsname, MGS_PARAM_MAXLEN);
694 OBD_FREE(param, MGS_PARAM_MAXLEN);
699 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
700 int *lmm_size, struct ptlrpc_request **request)
702 struct ll_sb_info *sbi = ll_i2sbi(inode);
703 struct mdt_body *body;
704 struct lov_mds_md *lmm = NULL;
705 struct ptlrpc_request *req = NULL;
707 struct md_op_data *op_data;
709 rc = ll_get_max_mdsize(sbi, &lmmsize);
713 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
714 0, lmmsize, LUSTRE_OPC_ANY,
719 op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
720 rc = md_getattr(sbi->ll_md_exp, op_data, &req);
721 ll_finish_md_op_data(op_data);
723 CDEBUG(D_INFO, "md_getattr failed on inode "
724 "%lu/%u: rc %d\n", inode->i_ino,
725 inode->i_generation, rc);
729 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
730 LASSERT(body != NULL);
732 lmmsize = body->eadatasize;
734 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
736 GOTO(out, rc = -ENODATA);
739 lmm = req_capsule_server_sized_get(&req->rq_pill,
740 &RMF_MDT_MD, lmmsize);
741 LASSERT(lmm != NULL);
744 * This is coming from the MDS, so is probably in
745 * little endian. We convert it to host endian before
746 * passing it to userspace.
748 /* We don't swab objects for directories */
749 switch (le32_to_cpu(lmm->lmm_magic)) {
751 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
752 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
755 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
756 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
759 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
770 * Get MDT index for the inode.
772 int ll_get_mdt_idx(struct inode *inode)
774 struct ll_sb_info *sbi = ll_i2sbi(inode);
775 struct md_op_data *op_data;
779 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
780 0, LUSTRE_OPC_ANY, NULL);
784 op_data->op_valid |= OBD_MD_MDTIDX;
785 rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
786 mdtidx = op_data->op_mds;
787 ll_finish_md_op_data(op_data);
789 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
795 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
803 if (cfs_copy_from_user(ptr, data, len)) {
807 rc = obd_iocontrol(cmd, exp, len, data, NULL);
812 static int ll_dir_ioctl(struct inode *inode, struct file *file,
813 unsigned int cmd, unsigned long arg)
815 struct ll_sb_info *sbi = ll_i2sbi(inode);
816 struct obd_ioctl_data *data;
820 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
821 inode->i_ino, inode->i_generation, inode, cmd);
823 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
824 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
827 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
829 case FSFILT_IOC_GETFLAGS:
830 case FSFILT_IOC_SETFLAGS:
831 RETURN(ll_iocontrol(inode, file, cmd, arg));
832 case FSFILT_IOC_GETVERSION_OLD:
833 case FSFILT_IOC_GETVERSION:
834 RETURN(put_user(inode->i_generation, (int *)arg));
835 /* We need to special case any other ioctls we want to handle,
836 * to send them to the MDS/OST as appropriate and to properly
837 * network encode the arg field.
838 case FSFILT_IOC_SETVERSION_OLD:
839 case FSFILT_IOC_SETVERSION:
841 case LL_IOC_GET_MDTIDX: {
844 mdtidx = ll_get_mdt_idx(inode);
848 if (put_user((int)mdtidx, (int*)arg))
853 case IOC_MDC_LOOKUP: {
854 struct ptlrpc_request *request = NULL;
855 int namelen, len = 0;
858 struct md_op_data *op_data;
860 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
865 filename = data->ioc_inlbuf1;
866 namelen = strlen(filename);
869 CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
870 GOTO(out_free, rc = -EINVAL);
873 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
874 0, LUSTRE_OPC_ANY, NULL);
876 GOTO(out_free, rc = -ENOMEM);
878 op_data->op_valid = OBD_MD_FLID;
879 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
880 ll_finish_md_op_data(op_data);
882 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
885 ptlrpc_req_finished(request);
888 obd_ioctl_freedata(buf, len);
891 case LL_IOC_LOV_SETSTRIPE: {
892 struct lov_user_md_v3 lumv3;
893 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
894 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
895 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
899 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
900 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
901 sizeof(lumv3p->lmm_objects[0]));
902 /* first try with v1 which is smaller than v3 */
903 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
906 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
907 if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
911 if (inode->i_sb->s_root == file->f_dentry)
914 /* in v1 and v3 cases lumv1 points to data */
915 rc = ll_dir_setstripe(inode, lumv1, set_default);
919 case LL_IOC_OBD_STATFS:
920 RETURN(ll_obd_statfs(inode, (void *)arg));
921 case LL_IOC_LOV_GETSTRIPE:
922 case LL_IOC_MDC_GETINFO:
923 case IOC_MDC_GETFILEINFO:
924 case IOC_MDC_GETFILESTRIPE: {
925 struct ptlrpc_request *request = NULL;
926 struct lov_user_md *lump;
927 struct lov_mds_md *lmm = NULL;
928 struct mdt_body *body;
929 char *filename = NULL;
932 if (cmd == IOC_MDC_GETFILEINFO ||
933 cmd == IOC_MDC_GETFILESTRIPE) {
934 filename = getname((const char *)arg);
935 if (IS_ERR(filename))
936 RETURN(PTR_ERR(filename));
938 rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
941 rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
945 body = req_capsule_server_get(&request->rq_pill,
947 LASSERT(body != NULL);
953 if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
954 cmd == LL_IOC_MDC_GETINFO))
955 GOTO(skip_lmm, rc = 0);
960 if (cmd == IOC_MDC_GETFILESTRIPE ||
961 cmd == LL_IOC_LOV_GETSTRIPE) {
962 lump = (struct lov_user_md *)arg;
964 struct lov_user_mds_data *lmdp;
965 lmdp = (struct lov_user_mds_data *)arg;
966 lump = &lmdp->lmd_lmm;
968 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
969 if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
970 GOTO(out_req, rc = -EFAULT);
974 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
975 struct lov_user_mds_data *lmdp;
978 st.st_dev = inode->i_sb->s_dev;
979 st.st_mode = body->mode;
980 st.st_nlink = body->nlink;
981 st.st_uid = body->uid;
982 st.st_gid = body->gid;
983 st.st_rdev = body->rdev;
984 st.st_size = body->size;
985 st.st_blksize = CFS_PAGE_SIZE;
986 st.st_blocks = body->blocks;
987 st.st_atime = body->atime;
988 st.st_mtime = body->mtime;
989 st.st_ctime = body->ctime;
990 st.st_ino = inode->i_ino;
992 lmdp = (struct lov_user_mds_data *)arg;
993 if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
994 GOTO(out_req, rc = -EFAULT);
999 ptlrpc_req_finished(request);
1004 case IOC_LOV_GETINFO: {
1005 struct lov_user_mds_data *lumd;
1006 struct lov_stripe_md *lsm;
1007 struct lov_user_md *lum;
1008 struct lov_mds_md *lmm;
1012 lumd = (struct lov_user_mds_data *)arg;
1013 lum = &lumd->lmd_lmm;
1015 rc = ll_get_max_mdsize(sbi, &lmmsize);
1019 OBD_ALLOC_LARGE(lmm, lmmsize);
1020 if (cfs_copy_from_user(lmm, lum, lmmsize))
1021 GOTO(free_lmm, rc = -EFAULT);
1023 switch (lmm->lmm_magic) {
1024 case LOV_USER_MAGIC_V1:
1025 if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1027 /* swab objects first so that stripes num will be sane */
1028 lustre_swab_lov_user_md_objects(
1029 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1030 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1031 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1033 case LOV_USER_MAGIC_V3:
1034 if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1036 /* swab objects first so that stripes num will be sane */
1037 lustre_swab_lov_user_md_objects(
1038 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1039 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1040 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1043 GOTO(free_lmm, rc = -EINVAL);
1046 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1048 GOTO(free_lmm, rc = -ENOMEM);
1050 /* Perform glimpse_size operation. */
1051 memset(&st, 0, sizeof(st));
1053 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1057 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1058 GOTO(free_lsm, rc = -EFAULT);
1062 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1064 OBD_FREE_LARGE(lmm, lmmsize);
1067 case OBD_IOC_LLOG_CATINFO: {
1068 struct ptlrpc_request *req = NULL;
1073 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1078 if (!data->ioc_inlbuf1) {
1079 obd_ioctl_freedata(buf, len);
1083 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
1086 GOTO(out_catinfo, rc = -ENOMEM);
1088 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
1090 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
1093 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
1095 ptlrpc_request_free(req);
1096 GOTO(out_catinfo, rc);
1099 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
1100 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
1101 if (data->ioc_inllen2) {
1102 str = req_capsule_client_get(&req->rq_pill,
1104 memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
1107 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
1109 ptlrpc_request_set_replen(req);
1111 rc = ptlrpc_queue_wait(req);
1113 str = req_capsule_server_get(&req->rq_pill,
1115 if (cfs_copy_to_user(data->ioc_pbuf1, str,
1119 ptlrpc_req_finished(req);
1121 obd_ioctl_freedata(buf, len);
1124 case OBD_IOC_QUOTACHECK: {
1125 struct obd_quotactl *oqctl;
1128 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1129 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1132 OBD_ALLOC_PTR(oqctl);
1135 oqctl->qc_type = arg;
1136 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1138 CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1142 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1144 CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1146 OBD_FREE_PTR(oqctl);
1149 case OBD_IOC_POLL_QUOTACHECK: {
1150 struct if_quotacheck *check;
1152 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1153 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1156 OBD_ALLOC_PTR(check);
1160 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1163 CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1164 if (cfs_copy_to_user((void *)arg, check,
1166 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1170 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1173 CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1174 if (cfs_copy_to_user((void *)arg, check,
1176 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1180 OBD_FREE_PTR(check);
1183 case OBD_IOC_QUOTACTL: {
1184 struct if_quotactl *qctl;
1185 int cmd, type, id, valid;
1187 OBD_ALLOC_PTR(qctl);
1191 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1192 GOTO(out_quotactl, rc = -EFAULT);
1195 type = qctl->qc_type;
1197 valid = qctl->qc_valid;
1200 case LUSTRE_Q_INVALIDATE:
1201 case LUSTRE_Q_FINVALIDATE:
1206 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1207 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1208 GOTO(out_quotactl, rc = -EPERM);
1211 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
1212 (type == GRPQUOTA && !in_egroup_p(id))) &&
1213 (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1214 sbi->ll_flags & LL_SBI_RMT_CLIENT))
1215 GOTO(out_quotactl, rc = -EPERM);
1220 CERROR("unsupported quotactl op: %#x\n", cmd);
1221 GOTO(out_quotactl, rc = -ENOTTY);
1224 if (valid != QC_GENERAL) {
1225 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
1226 GOTO(out_quotactl, rc = -EOPNOTSUPP);
1228 if (cmd == Q_GETINFO)
1229 qctl->qc_cmd = Q_GETOINFO;
1230 else if (cmd == Q_GETQUOTA)
1231 qctl->qc_cmd = Q_GETOQUOTA;
1233 GOTO(out_quotactl, rc = -EINVAL);
1237 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1239 sizeof(*qctl), qctl, NULL);
1242 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1244 sizeof(*qctl), qctl, NULL);
1247 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1249 sizeof(*qctl), qctl, NULL);
1251 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1253 sizeof(*qctl), qctl,
1262 GOTO(out_quotactl, rc);
1266 struct obd_quotactl *oqctl;
1268 OBD_ALLOC_PTR(oqctl);
1270 GOTO(out_quotactl, rc = -ENOMEM);
1272 QCTL_COPY(oqctl, qctl);
1273 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1275 if (rc != -EALREADY && cmd == Q_QUOTAON) {
1276 oqctl->qc_cmd = Q_QUOTAOFF;
1277 obd_quotactl(sbi->ll_md_exp, oqctl);
1279 OBD_FREE_PTR(oqctl);
1280 GOTO(out_quotactl, rc);
1282 QCTL_COPY(qctl, oqctl);
1283 OBD_FREE_PTR(oqctl);
1287 if (cfs_copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1294 case OBD_IOC_GETNAME: {
1295 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1298 if (cfs_copy_to_user((void *)arg, obd->obd_name,
1299 strlen(obd->obd_name) + 1))
1303 case LL_IOC_FLUSHCTX:
1304 RETURN(ll_flush_ctx(inode));
1305 #ifdef CONFIG_FS_POSIX_ACL
1306 case LL_IOC_RMTACL: {
1307 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1308 inode == inode->i_sb->s_root->d_inode) {
1309 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1311 LASSERT(fd != NULL);
1312 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1314 fd->fd_flags |= LL_FILE_RMTACL;
1320 case LL_IOC_GETOBDCOUNT: {
1323 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1328 struct lov_obd *lov = &sbi->ll_dt_exp->exp_obd->u.lov;
1329 count = lov->desc.ld_tgt_count;
1332 struct lmv_obd *lmv = &sbi->ll_md_exp->exp_obd->u.lmv;
1333 count = lmv->desc.ld_tgt_count;
1336 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1341 case LL_IOC_PATH2FID:
1342 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1343 sizeof(struct lu_fid)))
1346 case LL_IOC_GET_CONNECT_FLAGS: {
1347 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1349 case OBD_IOC_CHANGELOG_SEND:
1350 case OBD_IOC_CHANGELOG_CLEAR:
1351 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1352 sizeof(struct ioc_changelog));
1354 case OBD_IOC_FID2PATH:
1355 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1356 case LL_IOC_HSM_CT_START:
1357 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1358 sizeof(struct lustre_kernelcomm));
1362 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1366 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1368 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1369 loff_t pos = file->f_pos;
1373 if (origin == 1 && offset >= 0 && file->f_pos == DIR_END_OFF) {
1374 CWARN("end of dir hash, DIR_END_OFF(-2) is returned\n");
1375 RETURN(DIR_END_OFF);
1378 ret = default_llseek(file, offset, origin);
1380 struct ll_sb_info *sbi = ll_i2sbi(file->f_dentry->d_inode);
1382 if (ll_need_32bit_api(sbi)) {
1383 if (file->f_pos >> 32) {
1384 /* hash overflow, simple revert */
1388 fd->fd_dir.lfd_pos = file->f_pos << 32;
1391 fd->fd_dir.lfd_pos = file->f_pos;
1397 int ll_dir_open(struct inode *inode, struct file *file)
1400 RETURN(ll_file_open(inode, file));
1403 int ll_dir_release(struct inode *inode, struct file *file)
1406 RETURN(ll_file_release(inode, file));
1409 struct file_operations ll_dir_operations = {
1410 .llseek = ll_dir_seek,
1411 .open = ll_dir_open,
1412 .release = ll_dir_release,
1413 .read = generic_read_dir,
1414 .readdir = ll_readdir,
1415 .ioctl = ll_dir_ioctl