4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
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 <asm/uaccess.h>
46 #include <linux/buffer_head.h> // for wait_on_buffer
47 #include <linux/pagevec.h>
49 #define DEBUG_SUBSYSTEM S_LLITE
51 #include <lustre/lustre_idl.h>
52 #include <obd_support.h>
53 #include <obd_class.h>
54 #include <lustre_lib.h>
55 #include <lustre/lustre_idl.h>
56 #include <lustre_lite.h>
57 #include <lustre_dlm.h>
58 #include <lustre_fid.h>
59 #include "llite_internal.h"
62 * (new) readdir implementation overview.
64 * Original lustre readdir implementation cached exact copy of raw directory
65 * pages on the client. These pages were indexed in client page cache by
66 * logical offset in the directory file. This design, while very simple and
67 * intuitive had some inherent problems:
69 * . it implies that byte offset to the directory entry serves as a
70 * telldir(3)/seekdir(3) cookie, but that offset is not stable: in
71 * ext3/htree directory entries may move due to splits, and more
74 * . it is incompatible with the design of split directories for cmd3,
75 * that assumes that names are distributed across nodes based on their
76 * hash, and so readdir should be done in hash order.
78 * New readdir implementation does readdir in hash order, and uses hash of a
79 * file name as a telldir/seekdir cookie. This led to number of complications:
81 * . hash is not unique, so it cannot be used to index cached directory
82 * pages on the client (note, that it requires a whole pageful of hash
83 * collided entries to cause two pages to have identical hashes);
85 * . hash is not unique, so it cannot, strictly speaking, be used as an
86 * entry cookie. ext3/htree has the same problem and lustre implementation
87 * mimics their solution: seekdir(hash) positions directory at the first
88 * entry with the given hash.
94 * Client caches directory pages using hash of the first entry as an index. As
95 * noted above hash is not unique, so this solution doesn't work as is:
96 * special processing is needed for "page hash chains" (i.e., sequences of
97 * pages filled with entries all having the same hash value).
99 * First, such chains have to be detected. To this end, server returns to the
100 * client the hash of the first entry on the page next to one returned. When
101 * client detects that this hash is the same as hash of the first entry on the
102 * returned page, page hash collision has to be handled. Pages in the
103 * hash chain, except first one, are termed "overflow pages".
105 * Solution to index uniqueness problem is to not cache overflow
106 * pages. Instead, when page hash collision is detected, all overflow pages
107 * from emerging chain are immediately requested from the server and placed in
108 * a special data structure (struct ll_dir_chain). This data structure is used
109 * by ll_readdir() to process entries from overflow pages. When readdir
110 * invocation finishes, overflow pages are discarded. If page hash collision
111 * chain weren't completely processed, next call to readdir will again detect
112 * page hash collision, again read overflow pages in, process next portion of
113 * entries and again discard the pages. This is not as wasteful as it looks,
114 * because, given reasonable hash, page hash collisions are extremely rare.
116 * 1. directory positioning
118 * When seekdir(hash) is called, original
129 * identification of and access to overflow pages
133 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
134 * a header lu_dirpage which describes the start/end hash, and whether this
135 * page is empty (contains no dir entry) or hash collide with next page.
136 * After client receives reply, several pages will be integrated into dir page
137 * in CFS_PAGE_SIZE (if CFS_PAGE_SIZE greater than LU_PAGE_SIZE), and the
138 * lu_dirpage for this integrated page will be adjusted.
142 /* returns the page unlocked, but with a reference */
143 static int ll_dir_filler(void *_hash, struct page *page0)
145 struct inode *inode = page0->mapping->host;
146 int hash64 = ll_i2sbi(inode)->ll_flags & LL_SBI_64BIT_HASH;
147 struct obd_export *exp = ll_i2sbi(inode)->ll_md_exp;
148 struct ptlrpc_request *request;
149 struct mdt_body *body;
150 struct md_op_data *op_data;
151 __u64 hash = *((__u64 *)_hash);
152 struct page **page_pool;
154 #ifndef HAVE_ADD_TO_PAGE_CACHE_LRU
155 struct pagevec lru_pvec;
157 struct lu_dirpage *dp;
158 int max_pages = ll_i2sbi(inode)->ll_md_brw_size >> CFS_PAGE_SHIFT;
159 int nrdpgs = 0; /* number of pages read actually */
165 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) hash "LPU64"\n",
166 inode->i_ino, inode->i_generation, inode, hash);
168 LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
170 OBD_ALLOC(page_pool, sizeof(page) * max_pages);
171 if (page_pool != NULL) {
172 page_pool[0] = page0;
177 for (npages = 1; npages < max_pages; npages++) {
178 page = page_cache_alloc_cold(inode->i_mapping);
181 page_pool[npages] = page;
184 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
185 LUSTRE_OPC_ANY, NULL);
186 op_data->op_npages = npages;
187 op_data->op_offset = hash;
188 rc = md_readpage(exp, op_data, page_pool, &request);
189 ll_finish_md_op_data(op_data);
191 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
192 /* Checked by mdc_readpage() */
193 LASSERT(body != NULL);
195 if (body->valid & OBD_MD_FLSIZE)
196 cl_isize_write(inode, body->size);
198 nrdpgs = (request->rq_bulk->bd_nob_transferred+CFS_PAGE_SIZE-1)
200 SetPageUptodate(page0);
203 ptlrpc_req_finished(request);
205 CDEBUG(D_VFSTRACE, "read %d/%d pages\n", nrdpgs, npages);
207 ll_pagevec_init(&lru_pvec, 0);
208 for (i = 1; i < npages; i++) {
209 unsigned long offset;
214 if (rc < 0 || i >= nrdpgs) {
215 page_cache_release(page);
219 SetPageUptodate(page);
222 hash = le64_to_cpu(dp->ldp_hash_start);
225 offset = hash_x_index(hash, hash64);
227 prefetchw(&page->flags);
228 ret = ll_add_to_page_cache_lru(page, inode->i_mapping, offset,
232 if (ll_pagevec_add(&lru_pvec, page) == 0)
233 ll_pagevec_lru_add_file(&lru_pvec);
235 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
236 " %d\n", offset, ret);
238 page_cache_release(page);
240 ll_pagevec_lru_add_file(&lru_pvec);
242 if (page_pool != &page0)
243 OBD_FREE(page_pool, sizeof(struct page *) * max_pages);
248 static void ll_check_page(struct inode *dir, struct page *page)
250 /* XXX: check page format later */
251 SetPageChecked(page);
254 void ll_release_page(struct page *page, int remove)
259 if (likely(page->mapping != NULL))
260 truncate_complete_page(page->mapping, page);
263 page_cache_release(page);
267 * Find, kmap and return page that contains given hash.
269 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
270 __u64 *start, __u64 *end)
272 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
273 struct address_space *mapping = dir->i_mapping;
275 * Complement of hash is used as an index so that
276 * radix_tree_gang_lookup() can be used to find a page with starting
277 * hash _smaller_ than one we are looking for.
279 unsigned long offset = hash_x_index(*hash, hash64);
283 TREE_READ_LOCK_IRQ(mapping);
284 found = radix_tree_gang_lookup(&mapping->page_tree,
285 (void **)&page, offset, 1);
287 struct lu_dirpage *dp;
289 page_cache_get(page);
290 TREE_READ_UNLOCK_IRQ(mapping);
292 * In contrast to find_lock_page() we are sure that directory
293 * page cannot be truncated (while DLM lock is held) and,
294 * hence, can avoid restart.
296 * In fact, page cannot be locked here at all, because
297 * ll_dir_filler() does synchronous io.
300 if (PageUptodate(page)) {
302 if (BITS_PER_LONG == 32 && hash64) {
303 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
304 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
307 *start = le64_to_cpu(dp->ldp_hash_start);
308 *end = le64_to_cpu(dp->ldp_hash_end);
310 LASSERTF(*start <= *hash, "start = "LPX64",end = "
311 LPX64",hash = "LPX64"\n", *start, *end, *hash);
312 CDEBUG(D_VFSTRACE, "page %lu [%llu %llu], hash "LPU64"\n",
313 offset, *start, *end, *hash);
315 ll_release_page(page, 0);
317 } else if (*end != *start && *hash == *end) {
319 * upon hash collision, remove this page,
320 * otherwise put page reference, and
321 * ll_get_dir_page() will issue RPC to fetch
324 ll_release_page(page,
325 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
329 page_cache_release(page);
330 page = ERR_PTR(-EIO);
334 TREE_READ_UNLOCK_IRQ(mapping);
340 struct page *ll_get_dir_page(struct inode *dir, __u64 hash,
341 struct ll_dir_chain *chain)
343 ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
344 struct address_space *mapping = dir->i_mapping;
345 struct lustre_handle lockh;
346 struct lu_dirpage *dp;
353 struct ll_inode_info *lli = ll_i2info(dir);
354 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
357 rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
358 ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
360 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
361 ll_md_blocking_ast, ldlm_completion_ast,
363 struct lookup_intent it = { .it_op = IT_READDIR };
364 struct ptlrpc_request *request;
365 struct md_op_data *op_data;
367 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
368 LUSTRE_OPC_ANY, NULL);
370 return (void *)op_data;
372 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
373 op_data, &lockh, NULL, 0, NULL, 0);
375 ll_finish_md_op_data(op_data);
377 request = (struct ptlrpc_request *)it.d.lustre.it_data;
379 ptlrpc_req_finished(request);
381 CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
382 PFID(ll_inode2fid(dir)), hash, rc);
386 /* for cross-ref object, l_ast_data of the lock may not be set,
387 * we reset it here */
388 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
391 ldlm_lock_dump_handle(D_OTHER, &lockh);
393 cfs_mutex_lock(&lli->lli_readdir_mutex);
394 page = ll_dir_page_locate(dir, &lhash, &start, &end);
396 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
397 PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
398 GOTO(out_unlock, page);
399 } else if (page != NULL) {
401 * XXX nikita: not entirely correct handling of a corner case:
402 * suppose hash chain of entries with hash value HASH crosses
403 * border between pages P0 and P1. First both P0 and P1 are
404 * cached, seekdir() is called for some entry from the P0 part
405 * of the chain. Later P0 goes out of cache. telldir(HASH)
406 * happens and finds P1, as it starts with matching hash
407 * value. Remaining entries from P0 part of the chain are
408 * skipped. (Is that really a bug?)
410 * Possible solutions: 0. don't cache P1 is such case, handle
411 * it as an "overflow" page. 1. invalidate all pages at
412 * once. 2. use HASH|1 as an index for P1.
414 GOTO(hash_collision, page);
417 page = read_cache_page(mapping, hash_x_index(hash, hash64),
418 ll_dir_filler, &lhash);
420 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
421 PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
422 GOTO(out_unlock, page);
427 if (!PageUptodate(page)) {
428 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
429 PFID(ll_inode2fid(dir)), hash, -5);
432 if (!PageChecked(page))
433 ll_check_page(dir, page);
434 if (PageError(page)) {
435 CERROR("page error: "DFID" at "LPU64": rc %d\n",
436 PFID(ll_inode2fid(dir)), hash, -5);
440 dp = page_address(page);
441 if (BITS_PER_LONG == 32 && hash64) {
442 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
443 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
446 start = le64_to_cpu(dp->ldp_hash_start);
447 end = le64_to_cpu(dp->ldp_hash_end);
451 LASSERT(start == lhash);
452 CWARN("Page-wide hash collision: "LPU64"\n", end);
453 if (BITS_PER_LONG == 32 && hash64)
454 CWARN("Real page-wide hash collision at ["LPU64" "LPU64
455 "] with hash "LPU64"\n",
456 le64_to_cpu(dp->ldp_hash_start),
457 le64_to_cpu(dp->ldp_hash_end), hash);
459 * Fetch whole overflow chain...
466 cfs_mutex_unlock(&lli->lli_readdir_mutex);
467 ldlm_lock_decref(&lockh, mode);
471 ll_release_page(page, 1);
472 page = ERR_PTR(-EIO);
476 int ll_dir_read(struct inode *inode, __u64 *_pos, void *cookie,
479 struct ll_inode_info *info = ll_i2info(inode);
480 struct ll_sb_info *sbi = ll_i2sbi(inode);
482 int api32 = ll_need_32bit_api(sbi);
483 int hash64 = sbi->ll_flags & LL_SBI_64BIT_HASH;
485 struct ll_dir_chain chain;
490 ll_dir_chain_init(&chain);
492 page = ll_get_dir_page(inode, pos, &chain);
494 while (rc == 0 && !done) {
495 struct lu_dirpage *dp;
496 struct lu_dirent *ent;
500 * If page is empty (end of directory is reached),
503 __u64 hash = MDS_DIR_END_OFF;
506 dp = page_address(page);
507 for (ent = lu_dirent_start(dp); ent != NULL && !done;
508 ent = lu_dirent_next(ent)) {
516 * XXX: implement correct swabbing here.
519 hash = le64_to_cpu(ent->lde_hash);
522 * Skip until we find target hash
527 namelen = le16_to_cpu(ent->lde_namelen);
538 fid_le_to_cpu(&fid, &ent->lde_fid);
539 ino = cl_fid_build_ino(&fid, api32);
540 type = ll_dirent_type_get(ent);
541 /* For 'll_nfs_get_name_filldir()', it will try
542 * to access the 'ent' through its 'lde_name',
543 * so the parameter 'name' for 'filldir()' must
544 * be part of the 'ent'. */
545 done = filldir(cookie, ent->lde_name, namelen,
548 next = le64_to_cpu(dp->ldp_hash_end);
551 if (pos == MDS_DIR_END_OFF) {
553 * End of directory reached.
556 ll_release_page(page, 0);
557 } else if (1 /* chain is exhausted*/) {
559 * Normal case: continue to the next
562 ll_release_page(page,
563 le32_to_cpu(dp->ldp_flags) &
566 page = ll_get_dir_page(inode, pos,
570 * go into overflow page.
572 LASSERT(le32_to_cpu(dp->ldp_flags) &
574 ll_release_page(page, 1);
578 ll_release_page(page, 0);
582 CERROR("error reading dir "DFID" at %lu: rc %d\n",
583 PFID(&info->lli_fid), (unsigned long)pos, rc);
588 ll_dir_chain_fini(&chain);
592 static int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
594 struct inode *inode = filp->f_dentry->d_inode;
595 struct ll_file_data *lfd = LUSTRE_FPRIVATE(filp);
596 struct ll_sb_info *sbi = ll_i2sbi(inode);
597 __u64 pos = lfd->lfd_pos;
598 int hash64 = sbi->ll_flags & LL_SBI_64BIT_HASH;
599 int api32 = ll_need_32bit_api(sbi);
603 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu "
604 " 32bit_api %d\n", inode->i_ino, inode->i_generation,
605 inode, (unsigned long)pos, i_size_read(inode), api32);
607 if (pos == MDS_DIR_END_OFF)
613 rc = ll_dir_read(inode, &pos, cookie, filldir);
615 if (pos == MDS_DIR_END_OFF) {
617 filp->f_pos = LL_DIR_END_OFF_32BIT;
619 filp->f_pos = LL_DIR_END_OFF;
622 filp->f_pos = pos >> 32;
626 filp->f_version = inode->i_version;
627 touch_atime(filp->f_vfsmnt, filp->f_dentry);
631 ll_stats_ops_tally(sbi, LPROC_LL_READDIR, 1);
636 int ll_send_mgc_param(struct obd_export *mgc, char *string)
638 struct mgs_send_param *msp;
645 strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
646 rc = obd_set_info_async(NULL, mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
647 sizeof(struct mgs_send_param), msp, NULL);
649 CERROR("Failed to set parameter: %d\n", rc);
655 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
658 struct ll_sb_info *sbi = ll_i2sbi(inode);
659 struct md_op_data *op_data;
660 struct ptlrpc_request *req = NULL;
662 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
663 struct obd_device *mgc = lsi->lsi_mgc;
669 * This is coming from userspace, so should be in
670 * local endian. But the MDS would like it in little
671 * endian, so we swab it before we send it.
673 switch (lump->lmm_magic) {
674 case LOV_USER_MAGIC_V1: {
675 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
676 lustre_swab_lov_user_md_v1(lump);
677 lum_size = sizeof(struct lov_user_md_v1);
680 case LOV_USER_MAGIC_V3: {
681 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
682 lustre_swab_lov_user_md_v3(
683 (struct lov_user_md_v3 *)lump);
684 lum_size = sizeof(struct lov_user_md_v3);
688 CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
689 " %#08x != %#08x nor %#08x\n",
690 lump->lmm_magic, LOV_USER_MAGIC_V1,
696 lum_size = sizeof(struct lov_user_md_v1);
699 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
700 LUSTRE_OPC_ANY, NULL);
702 RETURN(PTR_ERR(op_data));
704 /* swabbing is done in lov_setstripe() on server side */
705 rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
706 NULL, 0, &req, NULL);
707 ll_finish_md_op_data(op_data);
708 ptlrpc_req_finished(req);
710 if (rc != -EPERM && rc != -EACCES)
711 CERROR("mdc_setattr fails: rc = %d\n", rc);
714 /* In the following we use the fact that LOV_USER_MAGIC_V1 and
715 LOV_USER_MAGIC_V3 have the same initial fields so we do not
716 need the make the distiction between the 2 versions */
717 if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
721 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
723 GOTO(end, rc = -ENOMEM);
726 /* Get fsname and assume devname to be -MDT0000. */
727 ll_get_fsname(inode->i_sb, buf, MTI_NAME_MAXLEN);
728 strcat(buf, "-MDT0000.lov");
731 /* Set root stripesize */
732 sprintf(buf, ".stripesize=%u",
733 lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
734 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
738 /* Set root stripecount */
739 sprintf(buf, ".stripecount=%hd",
740 lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
741 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
745 /* Set root stripeoffset */
746 sprintf(buf, ".stripeoffset=%hd",
747 lump ? le16_to_cpu(lump->lmm_stripe_offset) :
748 (typeof(lump->lmm_stripe_offset))(-1));
749 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
753 OBD_FREE(param, MGS_PARAM_MAXLEN);
758 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
759 int *lmm_size, struct ptlrpc_request **request)
761 struct ll_sb_info *sbi = ll_i2sbi(inode);
762 struct mdt_body *body;
763 struct lov_mds_md *lmm = NULL;
764 struct ptlrpc_request *req = NULL;
766 struct md_op_data *op_data;
768 rc = ll_get_max_mdsize(sbi, &lmmsize);
772 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
773 0, lmmsize, LUSTRE_OPC_ANY,
776 RETURN(PTR_ERR(op_data));
778 op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
779 rc = md_getattr(sbi->ll_md_exp, op_data, &req);
780 ll_finish_md_op_data(op_data);
782 CDEBUG(D_INFO, "md_getattr failed on inode "
783 "%lu/%u: rc %d\n", inode->i_ino,
784 inode->i_generation, rc);
788 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
789 LASSERT(body != NULL);
791 lmmsize = body->eadatasize;
793 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
795 GOTO(out, rc = -ENODATA);
798 lmm = req_capsule_server_sized_get(&req->rq_pill,
799 &RMF_MDT_MD, lmmsize);
800 LASSERT(lmm != NULL);
803 * This is coming from the MDS, so is probably in
804 * little endian. We convert it to host endian before
805 * passing it to userspace.
807 /* We don't swab objects for directories */
808 switch (le32_to_cpu(lmm->lmm_magic)) {
810 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
811 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
814 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
815 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
818 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
829 * Get MDT index for the inode.
831 int ll_get_mdt_idx(struct inode *inode)
833 struct ll_sb_info *sbi = ll_i2sbi(inode);
834 struct md_op_data *op_data;
838 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
839 0, LUSTRE_OPC_ANY, NULL);
841 RETURN(PTR_ERR(op_data));
843 op_data->op_flags |= MF_GET_MDT_IDX;
844 rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
845 mdtidx = op_data->op_mds;
846 ll_finish_md_op_data(op_data);
848 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
854 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
862 if (cfs_copy_from_user(ptr, data, len)) {
866 rc = obd_iocontrol(cmd, exp, len, data, NULL);
871 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
873 int cmd = qctl->qc_cmd;
874 int type = qctl->qc_type;
875 int id = qctl->qc_id;
876 int valid = qctl->qc_valid;
881 case LUSTRE_Q_INVALIDATE:
882 case LUSTRE_Q_FINVALIDATE:
887 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
888 sbi->ll_flags & LL_SBI_RMT_CLIENT)
892 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
893 (type == GRPQUOTA && !in_egroup_p(id))) &&
894 (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
895 sbi->ll_flags & LL_SBI_RMT_CLIENT))
901 CERROR("unsupported quotactl op: %#x\n", cmd);
905 if (valid != QC_GENERAL) {
906 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
909 if (cmd == Q_GETINFO)
910 qctl->qc_cmd = Q_GETOINFO;
911 else if (cmd == Q_GETQUOTA)
912 qctl->qc_cmd = Q_GETOQUOTA;
918 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
919 sizeof(*qctl), qctl, NULL);
922 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
923 sizeof(*qctl), qctl, NULL);
926 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
927 sizeof(*qctl), qctl, NULL);
929 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
931 sizeof(*qctl), qctl, NULL);
943 struct obd_quotactl *oqctl;
945 OBD_ALLOC_PTR(oqctl);
949 QCTL_COPY(oqctl, qctl);
950 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
952 if (rc != -EALREADY && cmd == Q_QUOTAON) {
953 oqctl->qc_cmd = Q_QUOTAOFF;
954 obd_quotactl(sbi->ll_md_exp, oqctl);
959 /* If QIF_SPACE is not set, client should collect the
960 * space usage from OSSs by itself */
961 if (cmd == Q_GETQUOTA &&
962 !(oqctl->qc_dqblk.dqb_valid & QIF_SPACE) &&
963 !oqctl->qc_dqblk.dqb_curspace) {
964 struct obd_quotactl *oqctl_tmp;
966 OBD_ALLOC_PTR(oqctl_tmp);
967 if (oqctl_tmp == NULL)
968 GOTO(out, rc = -ENOMEM);
970 oqctl_tmp->qc_cmd = Q_GETOQUOTA;
971 oqctl_tmp->qc_id = oqctl->qc_id;
972 oqctl_tmp->qc_type = oqctl->qc_type;
974 /* collect space usage from OSTs */
975 oqctl_tmp->qc_dqblk.dqb_curspace = 0;
976 rc = obd_quotactl(sbi->ll_dt_exp, oqctl_tmp);
977 if (!rc || rc == -EREMOTEIO) {
978 oqctl->qc_dqblk.dqb_curspace =
979 oqctl_tmp->qc_dqblk.dqb_curspace;
980 oqctl->qc_dqblk.dqb_valid |= QIF_SPACE;
983 /* collect space & inode usage from MDTs */
984 oqctl_tmp->qc_dqblk.dqb_curspace = 0;
985 oqctl_tmp->qc_dqblk.dqb_curinodes = 0;
986 rc = obd_quotactl(sbi->ll_md_exp, oqctl_tmp);
987 if (!rc || rc == -EREMOTEIO) {
988 oqctl->qc_dqblk.dqb_curspace +=
989 oqctl_tmp->qc_dqblk.dqb_curspace;
990 oqctl->qc_dqblk.dqb_curinodes =
991 oqctl_tmp->qc_dqblk.dqb_curinodes;
992 oqctl->qc_dqblk.dqb_valid |= QIF_INODES;
994 oqctl->qc_dqblk.dqb_valid &= ~QIF_SPACE;
997 OBD_FREE_PTR(oqctl_tmp);
1000 QCTL_COPY(qctl, oqctl);
1001 OBD_FREE_PTR(oqctl);
1007 static long ll_dir_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1009 struct inode *inode = file->f_dentry->d_inode;
1010 struct ll_sb_info *sbi = ll_i2sbi(inode);
1011 struct obd_ioctl_data *data;
1015 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1016 inode->i_ino, inode->i_generation, inode, cmd);
1018 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1019 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1022 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1024 case FSFILT_IOC_GETFLAGS:
1025 case FSFILT_IOC_SETFLAGS:
1026 RETURN(ll_iocontrol(inode, file, cmd, arg));
1027 case FSFILT_IOC_GETVERSION_OLD:
1028 case FSFILT_IOC_GETVERSION:
1029 RETURN(put_user(inode->i_generation, (int *)arg));
1030 /* We need to special case any other ioctls we want to handle,
1031 * to send them to the MDS/OST as appropriate and to properly
1032 * network encode the arg field.
1033 case FSFILT_IOC_SETVERSION_OLD:
1034 case FSFILT_IOC_SETVERSION:
1036 case LL_IOC_GET_MDTIDX: {
1039 mdtidx = ll_get_mdt_idx(inode);
1043 if (put_user((int)mdtidx, (int*)arg))
1048 case IOC_MDC_LOOKUP: {
1049 struct ptlrpc_request *request = NULL;
1050 int namelen, len = 0;
1053 struct md_op_data *op_data;
1055 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1060 filename = data->ioc_inlbuf1;
1061 namelen = strlen(filename);
1064 CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1065 GOTO(out_free, rc = -EINVAL);
1068 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1069 0, LUSTRE_OPC_ANY, NULL);
1070 if (IS_ERR(op_data))
1071 GOTO(out_free, rc = PTR_ERR(op_data));
1073 op_data->op_valid = OBD_MD_FLID;
1074 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1075 ll_finish_md_op_data(op_data);
1077 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1080 ptlrpc_req_finished(request);
1083 obd_ioctl_freedata(buf, len);
1086 case LL_IOC_LOV_SETSTRIPE: {
1087 struct lov_user_md_v3 lumv3;
1088 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1089 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1090 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1092 int set_default = 0;
1094 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1095 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1096 sizeof(lumv3p->lmm_objects[0]));
1097 /* first try with v1 which is smaller than v3 */
1098 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1101 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1102 if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1106 if (inode->i_sb->s_root == file->f_dentry)
1109 /* in v1 and v3 cases lumv1 points to data */
1110 rc = ll_dir_setstripe(inode, lumv1, set_default);
1114 case LL_IOC_OBD_STATFS:
1115 RETURN(ll_obd_statfs(inode, (void *)arg));
1116 case LL_IOC_LOV_GETSTRIPE:
1117 case LL_IOC_MDC_GETINFO:
1118 case IOC_MDC_GETFILEINFO:
1119 case IOC_MDC_GETFILESTRIPE: {
1120 struct ptlrpc_request *request = NULL;
1121 struct lov_user_md *lump;
1122 struct lov_mds_md *lmm = NULL;
1123 struct mdt_body *body;
1124 char *filename = NULL;
1127 if (cmd == IOC_MDC_GETFILEINFO ||
1128 cmd == IOC_MDC_GETFILESTRIPE) {
1129 filename = getname((const char *)arg);
1130 if (IS_ERR(filename))
1131 RETURN(PTR_ERR(filename));
1133 rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1134 &lmmsize, &request);
1136 rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1140 body = req_capsule_server_get(&request->rq_pill,
1142 LASSERT(body != NULL);
1148 if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1149 cmd == LL_IOC_MDC_GETINFO))
1150 GOTO(skip_lmm, rc = 0);
1155 if (cmd == IOC_MDC_GETFILESTRIPE ||
1156 cmd == LL_IOC_LOV_GETSTRIPE) {
1157 lump = (struct lov_user_md *)arg;
1159 struct lov_user_mds_data *lmdp;
1160 lmdp = (struct lov_user_mds_data *)arg;
1161 lump = &lmdp->lmd_lmm;
1163 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1164 if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1165 GOTO(out_req, rc = -EFAULT);
1169 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1170 struct lov_user_mds_data *lmdp;
1173 st.st_dev = inode->i_sb->s_dev;
1174 st.st_mode = body->mode;
1175 st.st_nlink = body->nlink;
1176 st.st_uid = body->uid;
1177 st.st_gid = body->gid;
1178 st.st_rdev = body->rdev;
1179 st.st_size = body->size;
1180 st.st_blksize = CFS_PAGE_SIZE;
1181 st.st_blocks = body->blocks;
1182 st.st_atime = body->atime;
1183 st.st_mtime = body->mtime;
1184 st.st_ctime = body->ctime;
1185 st.st_ino = inode->i_ino;
1187 lmdp = (struct lov_user_mds_data *)arg;
1188 if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1189 GOTO(out_req, rc = -EFAULT);
1194 ptlrpc_req_finished(request);
1199 case IOC_LOV_GETINFO: {
1200 struct lov_user_mds_data *lumd;
1201 struct lov_stripe_md *lsm;
1202 struct lov_user_md *lum;
1203 struct lov_mds_md *lmm;
1207 lumd = (struct lov_user_mds_data *)arg;
1208 lum = &lumd->lmd_lmm;
1210 rc = ll_get_max_mdsize(sbi, &lmmsize);
1214 OBD_ALLOC_LARGE(lmm, lmmsize);
1215 if (cfs_copy_from_user(lmm, lum, lmmsize))
1216 GOTO(free_lmm, rc = -EFAULT);
1218 switch (lmm->lmm_magic) {
1219 case LOV_USER_MAGIC_V1:
1220 if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1222 /* swab objects first so that stripes num will be sane */
1223 lustre_swab_lov_user_md_objects(
1224 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1225 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1226 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1228 case LOV_USER_MAGIC_V3:
1229 if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1231 /* swab objects first so that stripes num will be sane */
1232 lustre_swab_lov_user_md_objects(
1233 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1234 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1235 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1238 GOTO(free_lmm, rc = -EINVAL);
1241 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1243 GOTO(free_lmm, rc = -ENOMEM);
1245 /* Perform glimpse_size operation. */
1246 memset(&st, 0, sizeof(st));
1248 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1252 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1253 GOTO(free_lsm, rc = -EFAULT);
1257 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1259 OBD_FREE_LARGE(lmm, lmmsize);
1262 case OBD_IOC_LLOG_CATINFO: {
1263 RETURN(-EOPNOTSUPP);
1265 case OBD_IOC_QUOTACHECK: {
1266 struct obd_quotactl *oqctl;
1269 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1270 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1273 OBD_ALLOC_PTR(oqctl);
1276 oqctl->qc_type = arg;
1277 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1279 CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1283 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1285 CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1287 OBD_FREE_PTR(oqctl);
1290 case OBD_IOC_POLL_QUOTACHECK: {
1291 struct if_quotacheck *check;
1293 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1294 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1297 OBD_ALLOC_PTR(check);
1301 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1304 CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1305 if (cfs_copy_to_user((void *)arg, check,
1307 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1311 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1314 CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1315 if (cfs_copy_to_user((void *)arg, check,
1317 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1321 OBD_FREE_PTR(check);
1324 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
1325 case LL_IOC_QUOTACTL_18: {
1326 /* copy the old 1.x quota struct for internal use, then copy
1327 * back into old format struct. For 1.8 compatibility. */
1328 struct if_quotactl_18 *qctl_18;
1329 struct if_quotactl *qctl_20;
1331 OBD_ALLOC_PTR(qctl_18);
1335 OBD_ALLOC_PTR(qctl_20);
1337 GOTO(out_quotactl_18, rc = -ENOMEM);
1339 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1340 GOTO(out_quotactl_20, rc = -ENOMEM);
1342 QCTL_COPY(qctl_20, qctl_18);
1343 qctl_20->qc_idx = 0;
1345 /* XXX: dqb_valid was borrowed as a flag to mark that
1346 * only mds quota is wanted */
1347 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1348 qctl_18->qc_dqblk.dqb_valid) {
1349 qctl_20->qc_valid = QC_MDTIDX;
1350 qctl_20->qc_dqblk.dqb_valid = 0;
1351 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1352 qctl_20->qc_valid = QC_UUID;
1353 qctl_20->obd_uuid = qctl_18->obd_uuid;
1355 qctl_20->qc_valid = QC_GENERAL;
1358 rc = quotactl_ioctl(sbi, qctl_20);
1361 QCTL_COPY(qctl_18, qctl_20);
1362 qctl_18->obd_uuid = qctl_20->obd_uuid;
1364 if (cfs_copy_to_user((void *)arg, qctl_18,
1370 OBD_FREE_PTR(qctl_20);
1372 OBD_FREE_PTR(qctl_18);
1376 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1377 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0) */
1378 case LL_IOC_QUOTACTL: {
1379 struct if_quotactl *qctl;
1381 OBD_ALLOC_PTR(qctl);
1385 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1386 GOTO(out_quotactl, rc = -EFAULT);
1388 rc = quotactl_ioctl(sbi, qctl);
1390 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1397 case OBD_IOC_GETDTNAME:
1398 case OBD_IOC_GETMDNAME:
1399 RETURN(ll_get_obd_name(inode, cmd, arg));
1400 case LL_IOC_FLUSHCTX:
1401 RETURN(ll_flush_ctx(inode));
1402 #ifdef CONFIG_FS_POSIX_ACL
1403 case LL_IOC_RMTACL: {
1404 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1405 inode == inode->i_sb->s_root->d_inode) {
1406 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1408 LASSERT(fd != NULL);
1409 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1411 fd->fd_flags |= LL_FILE_RMTACL;
1417 case LL_IOC_GETOBDCOUNT: {
1419 struct obd_export *exp;
1421 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1424 /* get ost count when count is zero, get mdt count otherwise */
1425 exp = count ? sbi->ll_md_exp : sbi->ll_dt_exp;
1426 vallen = sizeof(count);
1427 rc = obd_get_info(NULL, exp, sizeof(KEY_TGT_COUNT),
1428 KEY_TGT_COUNT, &vallen, &count, NULL);
1430 CERROR("get target count failed: %d\n", rc);
1434 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1439 case LL_IOC_PATH2FID:
1440 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1441 sizeof(struct lu_fid)))
1444 case LL_IOC_GET_CONNECT_FLAGS: {
1445 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1447 case OBD_IOC_CHANGELOG_SEND:
1448 case OBD_IOC_CHANGELOG_CLEAR:
1449 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1450 sizeof(struct ioc_changelog));
1452 case OBD_IOC_FID2PATH:
1453 RETURN(ll_fid2path(inode, (void *)arg));
1454 case LL_IOC_HSM_CT_START:
1455 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1456 sizeof(struct lustre_kernelcomm));
1460 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1464 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1466 struct inode *inode = file->f_mapping->host;
1467 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1468 struct ll_sb_info *sbi = ll_i2sbi(inode);
1469 int api32 = ll_need_32bit_api(sbi);
1470 loff_t ret = -EINVAL;
1473 cfs_mutex_lock(&inode->i_mutex);
1478 offset += file->f_pos;
1484 offset += LL_DIR_END_OFF_32BIT;
1486 offset += LL_DIR_END_OFF;
1493 ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1494 (!api32 && offset <= LL_DIR_END_OFF))) {
1495 if (offset != file->f_pos) {
1496 if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1497 (!api32 && offset == LL_DIR_END_OFF))
1498 fd->lfd_pos = MDS_DIR_END_OFF;
1499 else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1500 fd->lfd_pos = offset << 32;
1502 fd->lfd_pos = offset;
1503 file->f_pos = offset;
1504 file->f_version = 0;
1511 cfs_mutex_unlock(&inode->i_mutex);
1515 int ll_dir_open(struct inode *inode, struct file *file)
1518 RETURN(ll_file_open(inode, file));
1521 int ll_dir_release(struct inode *inode, struct file *file)
1524 RETURN(ll_file_release(inode, file));
1527 struct file_operations ll_dir_operations = {
1528 .llseek = ll_dir_seek,
1529 .open = ll_dir_open,
1530 .release = ll_dir_release,
1531 .read = generic_read_dir,
1532 .readdir = ll_readdir,
1533 .unlocked_ioctl = ll_dir_ioctl,