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;
158 hash = (__u64)hash_x_index(page->index);
159 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
160 inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
162 oc = ll_mdscapa_get(inode);
163 rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
164 oc, hash, page, &request);
167 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
168 /* Checked by mdc_readpage() */
169 LASSERT(body != NULL);
171 if (body->valid & OBD_MD_FLSIZE)
172 cl_isize_write(inode, body->size);
173 SetPageUptodate(page);
175 ptlrpc_req_finished(request);
182 struct address_space_operations ll_dir_aops = {
183 .readpage = ll_dir_readpage,
186 static void ll_check_page(struct inode *dir, struct page *page)
188 /* XXX: check page format later */
189 SetPageChecked(page);
192 static void ll_release_page(struct page *page, __u64 hash,
193 __u64 start, __u64 end)
197 if (likely(page->mapping != NULL)) {
198 ll_truncate_complete_page(page);
202 CWARN("NULL mapping page %p, truncated by others: "
203 "hash("LPX64") | start("LPX64") | end("LPX64")\n",
204 page, hash, start, end);
206 page_cache_release(page);
210 * Find, kmap and return page that contains given hash.
212 static struct page *ll_dir_page_locate(struct inode *dir, __u64 hash,
213 __u64 *start, __u64 *end)
215 struct address_space *mapping = dir->i_mapping;
217 * Complement of hash is used as an index so that
218 * radix_tree_gang_lookup() can be used to find a page with starting
219 * hash _smaller_ than one we are looking for.
221 unsigned long offset = hash_x_index((unsigned long)hash);
225 TREE_READ_LOCK_IRQ(mapping);
226 found = radix_tree_gang_lookup(&mapping->page_tree,
227 (void **)&page, offset, 1);
229 struct lu_dirpage *dp;
231 page_cache_get(page);
232 TREE_READ_UNLOCK_IRQ(mapping);
234 * In contrast to find_lock_page() we are sure that directory
235 * page cannot be truncated (while DLM lock is held) and,
236 * hence, can avoid restart.
238 * In fact, page cannot be locked here at all, because
239 * ll_dir_readpage() does synchronous io.
242 if (PageUptodate(page)) {
244 *start = le64_to_cpu(dp->ldp_hash_start);
245 *end = le64_to_cpu(dp->ldp_hash_end);
246 LASSERT(*start <= hash);
247 if (hash > *end || (*end != *start && hash == *end)) {
248 ll_release_page(page, hash, *start, *end);
252 page_cache_release(page);
253 page = ERR_PTR(-EIO);
257 TREE_READ_UNLOCK_IRQ(mapping);
263 struct page *ll_get_dir_page(struct inode *dir, __u64 hash, int exact,
264 struct ll_dir_chain *chain)
266 ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
267 struct address_space *mapping = dir->i_mapping;
268 struct lustre_handle lockh;
269 struct lu_dirpage *dp;
277 rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
278 ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
280 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
281 ll_md_blocking_ast, ldlm_completion_ast,
283 struct lookup_intent it = { .it_op = IT_READDIR };
284 struct ptlrpc_request *request;
285 struct md_op_data *op_data;
287 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
288 LUSTRE_OPC_ANY, NULL);
290 return (void *)op_data;
292 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
293 op_data, &lockh, NULL, 0, NULL, 0);
295 ll_finish_md_op_data(op_data);
297 request = (struct ptlrpc_request *)it.d.lustre.it_data;
299 ptlrpc_req_finished(request);
301 CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
302 PFID(ll_inode2fid(dir)), hash, rc);
306 /* for cross-ref object, l_ast_data of the lock may not be set,
307 * we reset it here */
308 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
311 ldlm_lock_dump_handle(D_OTHER, &lockh);
313 page = ll_dir_page_locate(dir, hash, &start, &end);
315 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
316 PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
317 GOTO(out_unlock, page);
322 * XXX nikita: not entirely correct handling of a corner case:
323 * suppose hash chain of entries with hash value HASH crosses
324 * border between pages P0 and P1. First both P0 and P1 are
325 * cached, seekdir() is called for some entry from the P0 part
326 * of the chain. Later P0 goes out of cache. telldir(HASH)
327 * happens and finds P1, as it starts with matching hash
328 * value. Remaining entries from P0 part of the chain are
329 * skipped. (Is that really a bug?)
331 * Possible solutions: 0. don't cache P1 is such case, handle
332 * it as an "overflow" page. 1. invalidate all pages at
333 * once. 2. use HASH|1 as an index for P1.
335 if (exact && hash != start) {
337 * readdir asked for a page starting _exactly_ from
338 * given hash, but cache contains stale page, with
339 * entries with smaller hash values. Stale page should
340 * be invalidated, and new one fetched.
342 CDEBUG(D_OTHER, "Stale readpage page %p: "LPX64" != "LPX64"\n",
344 ll_release_page(page, hash, start, end);
346 GOTO(hash_collision, page);
350 page = read_cache_page(mapping, hash_x_index((unsigned long)hash),
351 (filler_t*)mapping->a_ops->readpage, NULL);
353 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
354 PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
355 GOTO(out_unlock, page);
360 if (!PageUptodate(page)) {
361 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
362 PFID(ll_inode2fid(dir)), hash, -5);
365 if (!PageChecked(page))
366 ll_check_page(dir, page);
367 if (PageError(page)) {
368 CERROR("page error: "DFID" at "LPU64": rc %d\n",
369 PFID(ll_inode2fid(dir)), hash, -5);
373 dp = page_address(page);
375 start = le64_to_cpu(dp->ldp_hash_start);
376 end = le64_to_cpu(dp->ldp_hash_end);
378 LASSERT(start == hash);
379 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
381 * Fetch whole overflow chain...
388 ldlm_lock_decref(&lockh, mode);
393 page = ERR_PTR(-EIO);
397 #define READDIR_ROLLBACK(dp, cur_page, prev_page, cross, cur_ent, first_ent) \
399 LASSERT(first_ent != NULL); \
400 cur_ent = first_ent; \
403 LASSERT(prev_page != NULL); \
404 ll_put_page(cur_page); \
405 cur_page = prev_page; \
407 dp = page_address(cur_page); \
411 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
413 struct inode *inode = filp->f_dentry->d_inode;
414 struct ll_inode_info *info = ll_i2info(inode);
415 __u64 pos = filp->f_pos;
417 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
418 struct lu_dirent *first = NULL;
419 struct page *prev_page = NULL;
420 struct lu_dirent *prev_ent = NULL;
422 struct ll_dir_chain chain;
429 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu\n",
430 inode->i_ino, inode->i_generation, inode,
431 (unsigned long)pos, i_size_read(inode));
433 /* If someone changed f_pos out of readdir, ignore fd_dir */
434 if (unlikely(fd->fd_dir.lfd_valid && fd->fd_dir.lfd_hash != pos)) {
435 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) someone changed "
436 "pos from %llu to %llu, maybe cause misposition\n",
437 inode->i_ino, inode->i_generation, inode,
438 fd->fd_dir.lfd_hash, pos);
439 fd->fd_dir.lfd_valid = 0;
442 if (pos == DIR_END_OFF)
451 ll_dir_chain_init(&chain);
453 page = ll_get_dir_page(inode, pos, 0, &chain);
455 while (rc == 0 && !done) {
456 struct lu_dirpage *dp;
457 struct lu_dirent *ent;
461 * If page is empty (end of directory is reached),
464 __u64 hash = DIR_END_OFF;
466 dp = page_address(page);
467 for (ent = lu_dirent_start(dp); ent != NULL && !done;
468 ent = lu_dirent_next(ent)) {
475 * XXX: implement correct swabbing here.
479 hash = le64_to_cpu(ent->lde_hash);
482 * Skip until we find target hash
487 namelen = le16_to_cpu(ent->lde_namelen);
494 name = ent->lde_name;
495 fid_le_to_cpu(&fid, &ent->lde_fid);
497 if (unlikely(fd->fd_dir.lfd_valid &&
499 if (fd->fd_dir.lfd_namelen != namelen ||
500 memcmp(fd->fd_dir.lfd_name, name,
505 * It is not the one we want.
511 fd->fd_dir.lfd_valid = 0;
513 &fd->fd_dir.lfd_fid)) {
517 CDEBUG(D_VFSTRACE,"VFS Op:inode"
518 "=%lu/%u(%p) someone "
519 "re-add at pos %llu\n",
533 if (unlikely(fd->fd_dir.lfd_valid &&
535 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u"
536 "(%p) miss to find pos %llu\n",
537 inode->i_ino,inode->i_generation,
539 fd->fd_dir.lfd_valid = 0;
541 READDIR_ROLLBACK(dp, page,
549 if (cfs_curproc_is_32bit())
550 ino = cl_fid_build_ino32(&fid);
552 ino = cl_fid_build_ino(&fid);
553 type = ll_dirent_type_get(ent);
555 if (prev_ent != NULL) {
556 if (unlikely(namelen ==
557 le16_to_cpu(prev_ent->lde_namelen)&&
558 memcmp(name, prev_ent->lde_name,
560 struct lu_fid prev_fid;
562 fid_le_to_cpu(&prev_fid,
564 CERROR("Pack two entries with "
565 "the same name [%.*s] "
567 "(%p): entry1 [%p], "
572 "FID2 ["DFID"]. (C)\n",
580 ent, hash, PFID(&fid));
581 LASSERT(ent != prev_ent);
585 done = filldir(cookie, name, namelen,
586 (loff_t)hash, ino, type);
592 if (unlikely(fd->fd_dir.lfd_valid && first)) {
597 * At here, we fall into full
598 * page hash collision.
599 * Currently, we can not handle
600 * full page hash collision yet.
601 * Give up search, just pack
604 CWARN("Full page hash collision"
605 " for inode=%lu/%u(%p) "
606 "with value ["LPU64"]\n",
610 fd->fd_dir.lfd_valid = 0;
611 READDIR_ROLLBACK(dp, page,
618 if (prev_page != NULL)
619 ll_put_page(prev_page);
622 pos = le64_to_cpu(dp->ldp_hash_end);
623 if (pos == DIR_END_OFF)
625 * End of directory reached.
628 else if (1 /* chain is exhausted*/)
630 * Normal case: continue to the next
633 page = ll_get_dir_page(inode, pos, 1,
637 * go into overflow page.
642 if (unlikely(prev_ent != NULL &&
643 le64_to_cpu(prev_ent->lde_hash) == hash)) {
644 struct lu_fid prev_fid;
647 le16_to_cpu(ent->lde_namelen);
649 fid_le_to_cpu(&prev_fid,
653 /* This should replace D_WARNING with
654 * D_VFSTRACE after bug 20581 fixed. */
655 CDEBUG(D_WARNING, "Hash collision for "
656 "inode=%lu/%u(%p) with the "
657 "same hash value ["LPU64"] to "
658 "be returned to user space "
659 "through multiple readdir: "
660 "entry1 [%p], name1 [%.*s], "
662 "entry2 [%p], name2 [%.*s], "
666 inode, hash, prev_ent,
668 prev_ent->lde_namelen),
674 LASSERT(ent != prev_ent);
675 fd->fd_dir.lfd_fid = fid;
676 fd->fd_dir.lfd_hash = hash;
677 fd->fd_dir.lfd_namelen = namelen;
678 memcpy(fd->fd_dir.lfd_name,
679 ent->lde_name, namelen);
680 fd->fd_dir.lfd_valid = 1;
683 if (prev_page != NULL)
684 ll_put_page(prev_page);
689 CERROR("error reading dir "DFID" at %lu: rc %d\n",
690 PFID(&info->lli_fid), (unsigned long)pos, rc);
694 if (prev_page != NULL)
695 ll_put_page(prev_page);
697 filp->f_pos = (loff_t)pos;
698 filp->f_version = inode->i_version;
699 touch_atime(filp->f_vfsmnt, filp->f_dentry);
701 ll_dir_chain_fini(&chain);
706 int ll_send_mgc_param(struct obd_export *mgc, char *string)
708 struct mgs_send_param *msp;
715 strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
716 rc = obd_set_info_async(mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
717 sizeof(struct mgs_send_param), msp, NULL);
719 CERROR("Failed to set parameter: %d\n", rc);
725 char *ll_get_fsname(struct inode *inode)
727 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
731 OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
732 len = strlen(lsi->lsi_lmd->lmd_profile);
733 ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
734 if (ptr && (strcmp(ptr, "-client") == 0))
736 strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
742 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
745 struct ll_sb_info *sbi = ll_i2sbi(inode);
746 struct md_op_data *op_data;
747 struct ptlrpc_request *req = NULL;
749 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
750 struct obd_device *mgc = lsi->lsi_mgc;
751 char *fsname = NULL, *param = NULL;
756 * This is coming from userspace, so should be in
757 * local endian. But the MDS would like it in little
758 * endian, so we swab it before we send it.
760 switch (lump->lmm_magic) {
761 case LOV_USER_MAGIC_V1: {
762 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
763 lustre_swab_lov_user_md_v1(lump);
764 lum_size = sizeof(struct lov_user_md_v1);
767 case LOV_USER_MAGIC_V3: {
768 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
769 lustre_swab_lov_user_md_v3(
770 (struct lov_user_md_v3 *)lump);
771 lum_size = sizeof(struct lov_user_md_v3);
775 CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
776 " %#08x != %#08x nor %#08x\n",
777 lump->lmm_magic, LOV_USER_MAGIC_V1,
783 lum_size = sizeof(struct lov_user_md_v1);
786 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
787 LUSTRE_OPC_ANY, NULL);
789 RETURN(PTR_ERR(op_data));
791 /* swabbing is done in lov_setstripe() on server side */
792 rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
793 NULL, 0, &req, NULL);
794 ll_finish_md_op_data(op_data);
795 ptlrpc_req_finished(req);
797 if (rc != -EPERM && rc != -EACCES)
798 CERROR("mdc_setattr fails: rc = %d\n", rc);
801 /* In the following we use the fact that LOV_USER_MAGIC_V1 and
802 LOV_USER_MAGIC_V3 have the same initial fields so we do not
803 need the make the distiction between the 2 versions */
804 if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
805 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
807 /* Get fsname and assume devname to be -MDT0000. */
808 fsname = ll_get_fsname(inode);
809 /* Set root stripesize */
810 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
811 lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
812 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
816 /* Set root stripecount */
817 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
818 lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
819 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
823 /* Set root stripeoffset */
824 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
825 lump ? le16_to_cpu(lump->lmm_stripe_offset) :
826 (typeof(lump->lmm_stripe_offset))(-1));
827 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
832 OBD_FREE(fsname, MGS_PARAM_MAXLEN);
834 OBD_FREE(param, MGS_PARAM_MAXLEN);
839 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
840 int *lmm_size, struct ptlrpc_request **request)
842 struct ll_sb_info *sbi = ll_i2sbi(inode);
843 struct mdt_body *body;
844 struct lov_mds_md *lmm = NULL;
845 struct ptlrpc_request *req = NULL;
847 struct md_op_data *op_data;
849 rc = ll_get_max_mdsize(sbi, &lmmsize);
853 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
854 0, lmmsize, LUSTRE_OPC_ANY,
859 op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
860 rc = md_getattr(sbi->ll_md_exp, op_data, &req);
861 ll_finish_md_op_data(op_data);
863 CDEBUG(D_INFO, "md_getattr failed on inode "
864 "%lu/%u: rc %d\n", inode->i_ino,
865 inode->i_generation, rc);
869 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
870 LASSERT(body != NULL);
872 lmmsize = body->eadatasize;
874 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
876 GOTO(out, rc = -ENODATA);
879 lmm = req_capsule_server_sized_get(&req->rq_pill,
880 &RMF_MDT_MD, lmmsize);
881 LASSERT(lmm != NULL);
884 * This is coming from the MDS, so is probably in
885 * little endian. We convert it to host endian before
886 * passing it to userspace.
888 /* We don't swab objects for directories */
889 switch (le32_to_cpu(lmm->lmm_magic)) {
891 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
892 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
895 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
896 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
899 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
910 * Get MDT index for the inode.
912 int ll_get_mdt_idx(struct inode *inode)
914 struct ll_sb_info *sbi = ll_i2sbi(inode);
915 struct md_op_data *op_data;
919 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
920 0, LUSTRE_OPC_ANY, NULL);
924 op_data->op_valid |= OBD_MD_MDTIDX;
925 rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
926 mdtidx = op_data->op_mds;
927 ll_finish_md_op_data(op_data);
929 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
935 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
943 if (cfs_copy_from_user(ptr, data, len)) {
947 rc = obd_iocontrol(cmd, exp, len, data, NULL);
952 static int ll_dir_ioctl(struct inode *inode, struct file *file,
953 unsigned int cmd, unsigned long arg)
955 struct ll_sb_info *sbi = ll_i2sbi(inode);
956 struct obd_ioctl_data *data;
960 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
961 inode->i_ino, inode->i_generation, inode, cmd);
963 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
964 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
967 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
969 case FSFILT_IOC_GETFLAGS:
970 case FSFILT_IOC_SETFLAGS:
971 RETURN(ll_iocontrol(inode, file, cmd, arg));
972 case FSFILT_IOC_GETVERSION_OLD:
973 case FSFILT_IOC_GETVERSION:
974 RETURN(put_user(inode->i_generation, (int *)arg));
975 /* We need to special case any other ioctls we want to handle,
976 * to send them to the MDS/OST as appropriate and to properly
977 * network encode the arg field.
978 case FSFILT_IOC_SETVERSION_OLD:
979 case FSFILT_IOC_SETVERSION:
981 case LL_IOC_GET_MDTIDX: {
984 mdtidx = ll_get_mdt_idx(inode);
988 if (put_user((int)mdtidx, (int*)arg))
993 case IOC_MDC_LOOKUP: {
994 struct ptlrpc_request *request = NULL;
995 int namelen, len = 0;
998 struct md_op_data *op_data;
1000 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1005 filename = data->ioc_inlbuf1;
1006 namelen = strlen(filename);
1009 CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1010 GOTO(out_free, rc = -EINVAL);
1013 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1014 0, LUSTRE_OPC_ANY, NULL);
1015 if (op_data == NULL)
1016 GOTO(out_free, rc = -ENOMEM);
1018 op_data->op_valid = OBD_MD_FLID;
1019 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1020 ll_finish_md_op_data(op_data);
1022 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1025 ptlrpc_req_finished(request);
1028 obd_ioctl_freedata(buf, len);
1031 case LL_IOC_LOV_SETSTRIPE: {
1032 struct lov_user_md_v3 lumv3;
1033 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1034 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1035 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1037 int set_default = 0;
1039 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1040 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1041 sizeof(lumv3p->lmm_objects[0]));
1042 /* first try with v1 which is smaller than v3 */
1043 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1046 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1047 if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1051 if (inode->i_sb->s_root == file->f_dentry)
1054 /* in v1 and v3 cases lumv1 points to data */
1055 rc = ll_dir_setstripe(inode, lumv1, set_default);
1059 case LL_IOC_OBD_STATFS:
1060 RETURN(ll_obd_statfs(inode, (void *)arg));
1061 case LL_IOC_LOV_GETSTRIPE:
1062 case LL_IOC_MDC_GETINFO:
1063 case IOC_MDC_GETFILEINFO:
1064 case IOC_MDC_GETFILESTRIPE: {
1065 struct ptlrpc_request *request = NULL;
1066 struct lov_user_md *lump;
1067 struct lov_mds_md *lmm = NULL;
1068 struct mdt_body *body;
1069 char *filename = NULL;
1072 if (cmd == IOC_MDC_GETFILEINFO ||
1073 cmd == IOC_MDC_GETFILESTRIPE) {
1074 filename = getname((const char *)arg);
1075 if (IS_ERR(filename))
1076 RETURN(PTR_ERR(filename));
1078 rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1079 &lmmsize, &request);
1081 rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1085 body = req_capsule_server_get(&request->rq_pill,
1087 LASSERT(body != NULL);
1093 if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1094 cmd == LL_IOC_MDC_GETINFO))
1095 GOTO(skip_lmm, rc = 0);
1100 if (cmd == IOC_MDC_GETFILESTRIPE ||
1101 cmd == LL_IOC_LOV_GETSTRIPE) {
1102 lump = (struct lov_user_md *)arg;
1104 struct lov_user_mds_data *lmdp;
1105 lmdp = (struct lov_user_mds_data *)arg;
1106 lump = &lmdp->lmd_lmm;
1108 if (cfs_copy_to_user(lump, lmm, lmmsize))
1109 GOTO(out_req, rc = -EFAULT);
1111 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1112 struct lov_user_mds_data *lmdp;
1115 st.st_dev = inode->i_sb->s_dev;
1116 st.st_mode = body->mode;
1117 st.st_nlink = body->nlink;
1118 st.st_uid = body->uid;
1119 st.st_gid = body->gid;
1120 st.st_rdev = body->rdev;
1121 st.st_size = body->size;
1122 st.st_blksize = CFS_PAGE_SIZE;
1123 st.st_blocks = body->blocks;
1124 st.st_atime = body->atime;
1125 st.st_mtime = body->mtime;
1126 st.st_ctime = body->ctime;
1127 st.st_ino = inode->i_ino;
1129 lmdp = (struct lov_user_mds_data *)arg;
1130 if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1131 GOTO(out_req, rc = -EFAULT);
1136 ptlrpc_req_finished(request);
1141 case IOC_LOV_GETINFO: {
1142 struct lov_user_mds_data *lumd;
1143 struct lov_stripe_md *lsm;
1144 struct lov_user_md *lum;
1145 struct lov_mds_md *lmm;
1149 lumd = (struct lov_user_mds_data *)arg;
1150 lum = &lumd->lmd_lmm;
1152 rc = ll_get_max_mdsize(sbi, &lmmsize);
1156 OBD_ALLOC(lmm, lmmsize);
1157 if (cfs_copy_from_user(lmm, lum, lmmsize))
1158 GOTO(free_lmm, rc = -EFAULT);
1160 switch (lmm->lmm_magic) {
1161 case LOV_USER_MAGIC_V1:
1162 if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1164 /* swab objects first so that stripes num will be sane */
1165 lustre_swab_lov_user_md_objects(
1166 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1167 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1168 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1170 case LOV_USER_MAGIC_V3:
1171 if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1173 /* swab objects first so that stripes num will be sane */
1174 lustre_swab_lov_user_md_objects(
1175 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1176 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1177 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1180 GOTO(free_lmm, rc = -EINVAL);
1183 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1185 GOTO(free_lmm, rc = -ENOMEM);
1187 /* Perform glimpse_size operation. */
1188 memset(&st, 0, sizeof(st));
1190 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1194 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1195 GOTO(free_lsm, rc = -EFAULT);
1199 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1201 OBD_FREE(lmm, lmmsize);
1204 case OBD_IOC_LLOG_CATINFO: {
1205 struct ptlrpc_request *req = NULL;
1210 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1215 if (!data->ioc_inlbuf1) {
1216 obd_ioctl_freedata(buf, len);
1220 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
1223 GOTO(out_catinfo, rc = -ENOMEM);
1225 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
1227 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
1230 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
1232 ptlrpc_request_free(req);
1233 GOTO(out_catinfo, rc);
1236 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
1237 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
1238 if (data->ioc_inllen2) {
1239 str = req_capsule_client_get(&req->rq_pill,
1241 memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
1244 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
1246 ptlrpc_request_set_replen(req);
1248 rc = ptlrpc_queue_wait(req);
1250 str = req_capsule_server_get(&req->rq_pill,
1252 if (cfs_copy_to_user(data->ioc_pbuf1, str,
1256 ptlrpc_req_finished(req);
1258 obd_ioctl_freedata(buf, len);
1261 case OBD_IOC_QUOTACHECK: {
1262 struct obd_quotactl *oqctl;
1265 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1266 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1269 OBD_ALLOC_PTR(oqctl);
1272 oqctl->qc_type = arg;
1273 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1275 CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1279 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1281 CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1283 OBD_FREE_PTR(oqctl);
1286 case OBD_IOC_POLL_QUOTACHECK: {
1287 struct if_quotacheck *check;
1289 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1290 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1293 OBD_ALLOC_PTR(check);
1297 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1300 CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1301 if (cfs_copy_to_user((void *)arg, check,
1307 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1310 CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1311 if (cfs_copy_to_user((void *)arg, check,
1317 OBD_FREE_PTR(check);
1320 case OBD_IOC_QUOTACTL: {
1321 struct if_quotactl *qctl;
1322 int cmd, type, id, valid;
1324 OBD_ALLOC_PTR(qctl);
1328 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1329 GOTO(out_quotactl, rc = -EFAULT);
1332 type = qctl->qc_type;
1334 valid = qctl->qc_valid;
1337 case LUSTRE_Q_INVALIDATE:
1338 case LUSTRE_Q_FINVALIDATE:
1343 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1344 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1345 GOTO(out_quotactl, rc = -EPERM);
1348 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
1349 (type == GRPQUOTA && !in_egroup_p(id))) &&
1350 (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1351 sbi->ll_flags & LL_SBI_RMT_CLIENT))
1352 GOTO(out_quotactl, rc = -EPERM);
1357 CERROR("unsupported quotactl op: %#x\n", cmd);
1358 GOTO(out_quotactl, rc = -ENOTTY);
1361 if (valid != QC_GENERAL) {
1362 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
1363 GOTO(out_quotactl, rc = -EOPNOTSUPP);
1365 if (cmd == Q_GETINFO)
1366 qctl->qc_cmd = Q_GETOINFO;
1367 else if (cmd == Q_GETQUOTA)
1368 qctl->qc_cmd = Q_GETOQUOTA;
1370 GOTO(out_quotactl, rc = -EINVAL);
1374 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1376 sizeof(*qctl), qctl, NULL);
1379 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1381 sizeof(*qctl), qctl, NULL);
1384 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1386 sizeof(*qctl), qctl, NULL);
1388 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1390 sizeof(*qctl), qctl,
1399 GOTO(out_quotactl, rc);
1403 struct obd_quotactl *oqctl;
1405 OBD_ALLOC_PTR(oqctl);
1407 GOTO(out_quotactl, rc = -ENOMEM);
1409 QCTL_COPY(oqctl, qctl);
1410 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1412 if (rc != -EALREADY && cmd == Q_QUOTAON) {
1413 oqctl->qc_cmd = Q_QUOTAOFF;
1414 obd_quotactl(sbi->ll_md_exp, oqctl);
1416 OBD_FREE_PTR(oqctl);
1417 GOTO(out_quotactl, rc);
1419 QCTL_COPY(qctl, oqctl);
1420 OBD_FREE_PTR(oqctl);
1424 if (cfs_copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1431 case OBD_IOC_GETNAME: {
1432 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1435 if (cfs_copy_to_user((void *)arg, obd->obd_name,
1436 strlen(obd->obd_name) + 1))
1440 case LL_IOC_FLUSHCTX:
1441 RETURN(ll_flush_ctx(inode));
1442 #ifdef CONFIG_FS_POSIX_ACL
1443 case LL_IOC_RMTACL: {
1444 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1445 inode == inode->i_sb->s_root->d_inode) {
1446 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1448 LASSERT(fd != NULL);
1449 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1451 fd->fd_flags |= LL_FILE_RMTACL;
1457 case LL_IOC_GETOBDCOUNT: {
1460 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1465 struct lov_obd *lov = &sbi->ll_dt_exp->exp_obd->u.lov;
1466 count = lov->desc.ld_tgt_count;
1469 struct lmv_obd *lmv = &sbi->ll_md_exp->exp_obd->u.lmv;
1470 count = lmv->desc.ld_tgt_count;
1473 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1478 case LL_IOC_PATH2FID:
1479 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1480 sizeof(struct lu_fid)))
1483 case LL_IOC_GET_CONNECT_FLAGS: {
1484 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1486 case OBD_IOC_CHANGELOG_SEND:
1487 case OBD_IOC_CHANGELOG_CLEAR:
1488 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1489 sizeof(struct ioc_changelog));
1491 case OBD_IOC_FID2PATH:
1492 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1493 case LL_IOC_HSM_CT_START:
1494 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1495 sizeof(struct lustre_kernelcomm));
1499 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1503 int ll_dir_open(struct inode *inode, struct file *file)
1506 RETURN(ll_file_open(inode, file));
1509 int ll_dir_release(struct inode *inode, struct file *file)
1512 RETURN(ll_file_release(inode, file));
1515 struct file_operations ll_dir_operations = {
1516 .open = ll_dir_open,
1517 .release = ll_dir_release,
1518 .read = generic_read_dir,
1519 .readdir = ll_readdir,
1520 .ioctl = ll_dir_ioctl