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 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
41 * Directory code for lustre client.
45 #include <linux/pagemap.h>
47 #include <linux/version.h>
48 #include <linux/smp_lock.h>
49 #include <asm/uaccess.h>
50 #include <linux/buffer_head.h> // for wait_on_buffer
51 #include <linux/pagevec.h>
53 #define DEBUG_SUBSYSTEM S_LLITE
55 #include <obd_support.h>
56 #include <obd_class.h>
57 #include <lustre_lib.h>
58 #include <lustre/lustre_idl.h>
59 #include <lustre_lite.h>
60 #include <lustre_dlm.h>
61 #include <lustre_fid.h>
62 #include "llite_internal.h"
64 #ifndef HAVE_PAGE_CHECKED
65 #ifdef HAVE_PG_FS_MISC
66 #define PageChecked(page) test_bit(PG_fs_misc, &(page)->flags)
67 #define SetPageChecked(page) set_bit(PG_fs_misc, &(page)->flags)
69 #error PageChecked or PageFsMisc not defined in kernel
74 * (new) readdir implementation overview.
76 * Original lustre readdir implementation cached exact copy of raw directory
77 * pages on the client. These pages were indexed in client page cache by
78 * logical offset in the directory file. This design, while very simple and
79 * intuitive had some inherent problems:
81 * . it implies that byte offset to the directory entry serves as a
82 * telldir(3)/seekdir(3) cookie, but that offset is not stable: in
83 * ext3/htree directory entries may move due to splits, and more
86 * . it is incompatible with the design of split directories for cmd3,
87 * that assumes that names are distributed across nodes based on their
88 * hash, and so readdir should be done in hash order.
90 * New readdir implementation does readdir in hash order, and uses hash of a
91 * file name as a telldir/seekdir cookie. This led to number of complications:
93 * . hash is not unique, so it cannot be used to index cached directory
94 * pages on the client (note, that it requires a whole pageful of hash
95 * collided entries to cause two pages to have identical hashes);
97 * . hash is not unique, so it cannot, strictly speaking, be used as an
98 * entry cookie. ext3/htree has the same problem and lustre implementation
99 * mimics their solution: seekdir(hash) positions directory at the first
100 * entry with the given hash.
106 * Client caches directory pages using hash of the first entry as an index. As
107 * noted above hash is not unique, so this solution doesn't work as is:
108 * special processing is needed for "page hash chains" (i.e., sequences of
109 * pages filled with entries all having the same hash value).
111 * First, such chains have to be detected. To this end, server returns to the
112 * client the hash of the first entry on the page next to one returned. When
113 * client detects that this hash is the same as hash of the first entry on the
114 * returned page, page hash collision has to be handled. Pages in the
115 * hash chain, except first one, are termed "overflow pages".
117 * Solution to index uniqueness problem is to not cache overflow
118 * pages. Instead, when page hash collision is detected, all overflow pages
119 * from emerging chain are immediately requested from the server and placed in
120 * a special data structure (struct ll_dir_chain). This data structure is used
121 * by ll_readdir() to process entries from overflow pages. When readdir
122 * invocation finishes, overflow pages are discarded. If page hash collision
123 * chain weren't completely processed, next call to readdir will again detect
124 * page hash collision, again read overflow pages in, process next portion of
125 * entries and again discard the pages. This is not as wasteful as it looks,
126 * because, given reasonable hash, page hash collisions are extremely rare.
128 * 1. directory positioning
130 * When seekdir(hash) is called, original
141 * identification of and access to overflow pages
145 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
146 * a header lu_dirpage which describes the start/end hash, and whether this
147 * page is empty (contains no dir entry) or hash collide with next page.
148 * After client receives reply, several pages will be integrated into dir page
149 * in CFS_PAGE_SIZE (if CFS_PAGE_SIZE greater than LU_PAGE_SIZE), and the
150 * lu_dirpage for this integrated page will be adjusted.
154 /* returns the page unlocked, but with a reference */
155 static int ll_dir_readpage(struct file *file, struct page *page0)
157 struct inode *inode = page0->mapping->host;
158 int hash64 = ll_i2sbi(inode)->ll_flags & LL_SBI_64BIT_HASH;
159 struct obd_export *exp = ll_i2sbi(inode)->ll_md_exp;
160 struct ptlrpc_request *request;
161 struct mdt_body *body;
164 struct page **page_pool;
166 #ifndef HAVE_ADD_TO_PAGE_CACHE_LRU
167 struct pagevec lru_pvec;
169 struct lu_dirpage *dp;
170 int max_pages = ll_i2sbi(inode)->ll_md_brw_size >> CFS_PAGE_SHIFT;
171 int nrdpgs = 0; /* number of pages read actually */
178 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
180 hash = fd->fd_dir.lfd_next;
182 hash = ll_i2info(inode)->lli_sa_pos;
184 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) hash "LPU64"\n",
185 inode->i_ino, inode->i_generation, inode, hash);
187 LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
189 OBD_ALLOC(page_pool, sizeof(page) * max_pages);
190 if (page_pool != NULL) {
191 page_pool[0] = page0;
196 for (npages = 1; npages < max_pages; npages++) {
197 page = page_cache_alloc_cold(inode->i_mapping);
200 page_pool[npages] = page;
203 oc = ll_mdscapa_get(inode);
204 rc = md_readpage(exp, ll_inode2fid(inode), oc, hash, page_pool, npages,
208 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
209 /* Checked by mdc_readpage() */
210 LASSERT(body != NULL);
212 if (body->valid & OBD_MD_FLSIZE)
213 cl_isize_write(inode, body->size);
215 nrdpgs = (request->rq_bulk->bd_nob_transferred+CFS_PAGE_SIZE-1)
217 SetPageUptodate(page0);
220 ptlrpc_req_finished(request);
222 CDEBUG(D_VFSTRACE, "read %d/%d pages\n", nrdpgs, npages);
224 ll_pagevec_init(&lru_pvec, 0);
225 for (i = 1; i < npages; i++) {
226 unsigned long offset;
231 if (rc < 0 || i >= nrdpgs) {
232 page_cache_release(page);
236 SetPageUptodate(page);
239 hash = le64_to_cpu(dp->ldp_hash_start);
242 offset = hash_x_index(hash, hash64);
244 prefetchw(&page->flags);
245 ret = ll_add_to_page_cache_lru(page, inode->i_mapping, offset,
249 page_cache_get(page);
250 if (ll_pagevec_add(&lru_pvec, page) == 0)
251 ll_pagevec_lru_add_file(&lru_pvec);
253 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
254 " %d\n", offset, ret);
256 page_cache_release(page);
258 ll_pagevec_lru_add_file(&lru_pvec);
260 if (page_pool != &page0)
261 OBD_FREE(page_pool, sizeof(struct page *) * max_pages);
266 #ifndef MS_HAS_NEW_AOPS
267 struct address_space_operations ll_dir_aops = {
268 .readpage = ll_dir_readpage,
271 struct address_space_operations_ext ll_dir_aops = {
272 .orig_aops.readpage = ll_dir_readpage,
276 static void ll_check_page(struct inode *dir, struct page *page)
278 /* XXX: check page format later */
279 SetPageChecked(page);
282 void ll_release_page(struct page *page, int remove)
287 if (likely(page->mapping != NULL))
288 truncate_complete_page(page->mapping, page);
291 page_cache_release(page);
295 * Find, kmap and return page that contains given hash.
297 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
298 __u64 *start, __u64 *end)
300 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
301 struct address_space *mapping = dir->i_mapping;
303 * Complement of hash is used as an index so that
304 * radix_tree_gang_lookup() can be used to find a page with starting
305 * hash _smaller_ than one we are looking for.
307 unsigned long offset = hash_x_index(*hash, hash64);
311 TREE_READ_LOCK_IRQ(mapping);
312 found = radix_tree_gang_lookup(&mapping->page_tree,
313 (void **)&page, offset, 1);
315 struct lu_dirpage *dp;
317 page_cache_get(page);
318 TREE_READ_UNLOCK_IRQ(mapping);
320 * In contrast to find_lock_page() we are sure that directory
321 * page cannot be truncated (while DLM lock is held) and,
322 * hence, can avoid restart.
324 * In fact, page cannot be locked here at all, because
325 * ll_dir_readpage() does synchronous io.
328 if (PageUptodate(page)) {
330 if (BITS_PER_LONG == 32 && hash64) {
331 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
332 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
335 *start = le64_to_cpu(dp->ldp_hash_start);
336 *end = le64_to_cpu(dp->ldp_hash_end);
338 LASSERTF(*start <= *hash, "start = "LPX64",end = "
339 LPX64",hash = "LPX64"\n", *start, *end, *hash);
340 CDEBUG(D_VFSTRACE, "page %lu [%llu %llu], hash "LPU64"\n",
341 offset, *start, *end, *hash);
343 ll_release_page(page, 0);
345 } else if (*end != *start && *hash == *end) {
347 * upon hash collision, remove this page,
348 * otherwise put page reference, and
349 * ll_get_dir_page() will issue RPC to fetch
352 ll_release_page(page,
353 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
357 page_cache_release(page);
358 page = ERR_PTR(-EIO);
362 TREE_READ_UNLOCK_IRQ(mapping);
368 struct page *ll_get_dir_page(struct file *filp, struct inode *dir, __u64 hash,
369 struct ll_dir_chain *chain)
371 ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
372 struct address_space *mapping = dir->i_mapping;
373 struct lustre_handle lockh;
374 struct lu_dirpage *dp;
381 struct ll_inode_info *lli = ll_i2info(dir);
382 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
385 rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
386 ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
388 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
389 ll_md_blocking_ast, ldlm_completion_ast,
391 struct lookup_intent it = { .it_op = IT_READDIR };
392 struct ptlrpc_request *request;
393 struct md_op_data *op_data;
395 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
396 LUSTRE_OPC_ANY, NULL);
398 return (void *)op_data;
400 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
401 op_data, &lockh, NULL, 0, NULL, 0);
403 ll_finish_md_op_data(op_data);
405 request = (struct ptlrpc_request *)it.d.lustre.it_data;
407 ptlrpc_req_finished(request);
409 CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
410 PFID(ll_inode2fid(dir)), hash, rc);
414 /* for cross-ref object, l_ast_data of the lock may not be set,
415 * we reset it here */
416 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
419 ldlm_lock_dump_handle(D_OTHER, &lockh);
421 cfs_down(&lli->lli_readdir_sem);
422 page = ll_dir_page_locate(dir, &lhash, &start, &end);
424 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
425 PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
426 GOTO(out_unlock, page);
427 } else if (page != NULL) {
429 * XXX nikita: not entirely correct handling of a corner case:
430 * suppose hash chain of entries with hash value HASH crosses
431 * border between pages P0 and P1. First both P0 and P1 are
432 * cached, seekdir() is called for some entry from the P0 part
433 * of the chain. Later P0 goes out of cache. telldir(HASH)
434 * happens and finds P1, as it starts with matching hash
435 * value. Remaining entries from P0 part of the chain are
436 * skipped. (Is that really a bug?)
438 * Possible solutions: 0. don't cache P1 is such case, handle
439 * it as an "overflow" page. 1. invalidate all pages at
440 * once. 2. use HASH|1 as an index for P1.
442 GOTO(hash_collision, page);
445 page = read_cache_page(mapping, hash_x_index(hash, hash64),
446 (filler_t*)mapping->a_ops->readpage, filp);
448 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
449 PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
450 GOTO(out_unlock, page);
455 if (!PageUptodate(page)) {
456 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
457 PFID(ll_inode2fid(dir)), hash, -5);
460 if (!PageChecked(page))
461 ll_check_page(dir, page);
462 if (PageError(page)) {
463 CERROR("page error: "DFID" at "LPU64": rc %d\n",
464 PFID(ll_inode2fid(dir)), hash, -5);
468 dp = page_address(page);
469 if (BITS_PER_LONG == 32 && hash64) {
470 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
471 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
474 start = le64_to_cpu(dp->ldp_hash_start);
475 end = le64_to_cpu(dp->ldp_hash_end);
479 LASSERT(start == lhash);
480 CWARN("Page-wide hash collision: "LPU64"\n", end);
481 if (BITS_PER_LONG == 32 && hash64)
482 CWARN("Real page-wide hash collision at ["LPU64" "LPU64
483 "] with hash "LPU64"\n",
484 le64_to_cpu(dp->ldp_hash_start),
485 le64_to_cpu(dp->ldp_hash_end), hash);
487 * Fetch whole overflow chain...
494 cfs_up(&lli->lli_readdir_sem);
495 ldlm_lock_decref(&lockh, mode);
499 ll_release_page(page, 1);
500 page = ERR_PTR(-EIO);
504 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
506 struct inode *inode = filp->f_dentry->d_inode;
507 struct ll_inode_info *info = ll_i2info(inode);
508 struct ll_sb_info *sbi = ll_i2sbi(inode);
509 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
510 __u64 pos = fd->fd_dir.lfd_pos;
511 int api32 = ll_need_32bit_api(sbi);
512 int hash64 = sbi->ll_flags & LL_SBI_64BIT_HASH;
514 struct ll_dir_chain chain;
520 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu 32bit_api %d\n",
521 inode->i_ino, inode->i_generation, inode,
522 (unsigned long)pos, i_size_read(inode), api32);
524 if (pos == MDS_DIR_END_OFF)
533 ll_dir_chain_init(&chain);
535 fd->fd_dir.lfd_next = pos;
536 page = ll_get_dir_page(filp, inode, pos, &chain);
538 while (rc == 0 && !done) {
539 struct lu_dirpage *dp;
540 struct lu_dirent *ent;
544 * If page is empty (end of directory is reached),
547 __u64 hash = MDS_DIR_END_OFF;
550 dp = page_address(page);
551 for (ent = lu_dirent_start(dp); ent != NULL && !done;
552 ent = lu_dirent_next(ent)) {
560 * XXX: implement correct swabbing here.
563 hash = le64_to_cpu(ent->lde_hash);
566 * Skip until we find target hash
571 namelen = le16_to_cpu(ent->lde_namelen);
582 fid_le_to_cpu(&fid, &ent->lde_fid);
583 ino = cl_fid_build_ino(&fid, api32);
584 type = ll_dirent_type_get(ent);
585 /* For 'll_nfs_get_name_filldir()', it will try
586 * to access the 'ent' through its 'lde_name',
587 * so the parameter 'name' for 'filldir()' must
588 * be part of the 'ent'. */
589 done = filldir(cookie, ent->lde_name, namelen,
592 next = le64_to_cpu(dp->ldp_hash_end);
595 if (pos == MDS_DIR_END_OFF) {
597 * End of directory reached.
600 ll_release_page(page, 0);
601 } else if (1 /* chain is exhausted*/) {
603 * Normal case: continue to the next
606 ll_release_page(page,
607 le32_to_cpu(dp->ldp_flags) &
609 fd->fd_dir.lfd_next = pos;
610 page = ll_get_dir_page(filp, inode, pos,
614 * go into overflow page.
616 LASSERT(le32_to_cpu(dp->ldp_flags) &
618 ll_release_page(page, 1);
622 ll_release_page(page, 0);
626 CERROR("error reading dir "DFID" at %lu: rc %d\n",
627 PFID(&info->lli_fid), (unsigned long)pos, rc);
631 fd->fd_dir.lfd_pos = pos;
632 if (pos == MDS_DIR_END_OFF) {
634 filp->f_pos = LL_DIR_END_OFF_32BIT;
636 filp->f_pos = LL_DIR_END_OFF;
639 filp->f_pos = pos >> 32;
643 filp->f_version = inode->i_version;
644 touch_atime(filp->f_vfsmnt, filp->f_dentry);
646 ll_dir_chain_fini(&chain);
651 int ll_send_mgc_param(struct obd_export *mgc, char *string)
653 struct mgs_send_param *msp;
660 strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
661 rc = obd_set_info_async(mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
662 sizeof(struct mgs_send_param), msp, NULL);
664 CERROR("Failed to set parameter: %d\n", rc);
670 char *ll_get_fsname(struct inode *inode)
672 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
676 OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
677 len = strlen(lsi->lsi_lmd->lmd_profile);
678 ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
679 if (ptr && (strcmp(ptr, "-client") == 0))
681 strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
687 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
690 struct ll_sb_info *sbi = ll_i2sbi(inode);
691 struct md_op_data *op_data;
692 struct ptlrpc_request *req = NULL;
694 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
695 struct obd_device *mgc = lsi->lsi_mgc;
696 char *fsname = NULL, *param = NULL;
701 * This is coming from userspace, so should be in
702 * local endian. But the MDS would like it in little
703 * endian, so we swab it before we send it.
705 switch (lump->lmm_magic) {
706 case LOV_USER_MAGIC_V1: {
707 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
708 lustre_swab_lov_user_md_v1(lump);
709 lum_size = sizeof(struct lov_user_md_v1);
712 case LOV_USER_MAGIC_V3: {
713 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
714 lustre_swab_lov_user_md_v3(
715 (struct lov_user_md_v3 *)lump);
716 lum_size = sizeof(struct lov_user_md_v3);
720 CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
721 " %#08x != %#08x nor %#08x\n",
722 lump->lmm_magic, LOV_USER_MAGIC_V1,
728 lum_size = sizeof(struct lov_user_md_v1);
731 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
732 LUSTRE_OPC_ANY, NULL);
734 RETURN(PTR_ERR(op_data));
736 /* swabbing is done in lov_setstripe() on server side */
737 rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
738 NULL, 0, &req, NULL);
739 ll_finish_md_op_data(op_data);
740 ptlrpc_req_finished(req);
742 if (rc != -EPERM && rc != -EACCES)
743 CERROR("mdc_setattr fails: rc = %d\n", rc);
746 /* In the following we use the fact that LOV_USER_MAGIC_V1 and
747 LOV_USER_MAGIC_V3 have the same initial fields so we do not
748 need the make the distiction between the 2 versions */
749 if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
750 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
752 /* Get fsname and assume devname to be -MDT0000. */
753 fsname = ll_get_fsname(inode);
754 /* Set root stripesize */
755 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
756 lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
757 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
761 /* Set root stripecount */
762 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
763 lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
764 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
768 /* Set root stripeoffset */
769 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
770 lump ? le16_to_cpu(lump->lmm_stripe_offset) :
771 (typeof(lump->lmm_stripe_offset))(-1));
772 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
777 OBD_FREE(fsname, MGS_PARAM_MAXLEN);
779 OBD_FREE(param, MGS_PARAM_MAXLEN);
784 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
785 int *lmm_size, struct ptlrpc_request **request)
787 struct ll_sb_info *sbi = ll_i2sbi(inode);
788 struct mdt_body *body;
789 struct lov_mds_md *lmm = NULL;
790 struct ptlrpc_request *req = NULL;
792 struct md_op_data *op_data;
794 rc = ll_get_max_mdsize(sbi, &lmmsize);
798 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
799 0, lmmsize, LUSTRE_OPC_ANY,
804 op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
805 rc = md_getattr(sbi->ll_md_exp, op_data, &req);
806 ll_finish_md_op_data(op_data);
808 CDEBUG(D_INFO, "md_getattr failed on inode "
809 "%lu/%u: rc %d\n", inode->i_ino,
810 inode->i_generation, rc);
814 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
815 LASSERT(body != NULL);
817 lmmsize = body->eadatasize;
819 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
821 GOTO(out, rc = -ENODATA);
824 lmm = req_capsule_server_sized_get(&req->rq_pill,
825 &RMF_MDT_MD, lmmsize);
826 LASSERT(lmm != NULL);
829 * This is coming from the MDS, so is probably in
830 * little endian. We convert it to host endian before
831 * passing it to userspace.
833 /* We don't swab objects for directories */
834 switch (le32_to_cpu(lmm->lmm_magic)) {
836 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
837 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
840 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
841 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
844 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
855 * Get MDT index for the inode.
857 int ll_get_mdt_idx(struct inode *inode)
859 struct ll_sb_info *sbi = ll_i2sbi(inode);
860 struct md_op_data *op_data;
864 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
865 0, LUSTRE_OPC_ANY, NULL);
869 op_data->op_valid |= OBD_MD_MDTIDX;
870 rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
871 mdtidx = op_data->op_mds;
872 ll_finish_md_op_data(op_data);
874 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
880 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
888 if (cfs_copy_from_user(ptr, data, len)) {
892 rc = obd_iocontrol(cmd, exp, len, data, NULL);
897 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
899 int cmd = qctl->qc_cmd;
900 int type = qctl->qc_type;
901 int id = qctl->qc_id;
902 int valid = qctl->qc_valid;
907 case LUSTRE_Q_INVALIDATE:
908 case LUSTRE_Q_FINVALIDATE:
913 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
914 sbi->ll_flags & LL_SBI_RMT_CLIENT)
918 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
919 (type == GRPQUOTA && !in_egroup_p(id))) &&
920 (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
921 sbi->ll_flags & LL_SBI_RMT_CLIENT))
927 CERROR("unsupported quotactl op: %#x\n", cmd);
931 if (valid != QC_GENERAL) {
932 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
935 if (cmd == Q_GETINFO)
936 qctl->qc_cmd = Q_GETOINFO;
937 else if (cmd == Q_GETQUOTA)
938 qctl->qc_cmd = Q_GETOQUOTA;
944 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
945 sizeof(*qctl), qctl, NULL);
948 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
949 sizeof(*qctl), qctl, NULL);
952 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
953 sizeof(*qctl), qctl, NULL);
955 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
957 sizeof(*qctl), qctl, NULL);
969 struct obd_quotactl *oqctl;
971 OBD_ALLOC_PTR(oqctl);
975 QCTL_COPY(oqctl, qctl);
976 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
978 if (rc != -EALREADY && cmd == Q_QUOTAON) {
979 oqctl->qc_cmd = Q_QUOTAOFF;
980 obd_quotactl(sbi->ll_md_exp, oqctl);
986 QCTL_COPY(qctl, oqctl);
993 static int ll_dir_ioctl(struct inode *inode, struct file *file,
994 unsigned int cmd, unsigned long arg)
996 struct ll_sb_info *sbi = ll_i2sbi(inode);
997 struct obd_ioctl_data *data;
1001 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1002 inode->i_ino, inode->i_generation, inode, cmd);
1004 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1005 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1008 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1010 case FSFILT_IOC_GETFLAGS:
1011 case FSFILT_IOC_SETFLAGS:
1012 RETURN(ll_iocontrol(inode, file, cmd, arg));
1013 case FSFILT_IOC_GETVERSION_OLD:
1014 case FSFILT_IOC_GETVERSION:
1015 RETURN(put_user(inode->i_generation, (int *)arg));
1016 /* We need to special case any other ioctls we want to handle,
1017 * to send them to the MDS/OST as appropriate and to properly
1018 * network encode the arg field.
1019 case FSFILT_IOC_SETVERSION_OLD:
1020 case FSFILT_IOC_SETVERSION:
1022 case LL_IOC_GET_MDTIDX: {
1025 mdtidx = ll_get_mdt_idx(inode);
1029 if (put_user((int)mdtidx, (int*)arg))
1034 case IOC_MDC_LOOKUP: {
1035 struct ptlrpc_request *request = NULL;
1036 int namelen, len = 0;
1039 struct md_op_data *op_data;
1041 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1046 filename = data->ioc_inlbuf1;
1047 namelen = strlen(filename);
1050 CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1051 GOTO(out_free, rc = -EINVAL);
1054 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1055 0, LUSTRE_OPC_ANY, NULL);
1056 if (op_data == NULL)
1057 GOTO(out_free, rc = -ENOMEM);
1059 op_data->op_valid = OBD_MD_FLID;
1060 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1061 ll_finish_md_op_data(op_data);
1063 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1066 ptlrpc_req_finished(request);
1069 obd_ioctl_freedata(buf, len);
1072 case LL_IOC_LOV_SETSTRIPE: {
1073 struct lov_user_md_v3 lumv3;
1074 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1075 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1076 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1078 int set_default = 0;
1080 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1081 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1082 sizeof(lumv3p->lmm_objects[0]));
1083 /* first try with v1 which is smaller than v3 */
1084 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1087 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1088 if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1092 if (inode->i_sb->s_root == file->f_dentry)
1095 /* in v1 and v3 cases lumv1 points to data */
1096 rc = ll_dir_setstripe(inode, lumv1, set_default);
1100 case LL_IOC_OBD_STATFS:
1101 RETURN(ll_obd_statfs(inode, (void *)arg));
1102 case LL_IOC_LOV_GETSTRIPE:
1103 case LL_IOC_MDC_GETINFO:
1104 case IOC_MDC_GETFILEINFO:
1105 case IOC_MDC_GETFILESTRIPE: {
1106 struct ptlrpc_request *request = NULL;
1107 struct lov_user_md *lump;
1108 struct lov_mds_md *lmm = NULL;
1109 struct mdt_body *body;
1110 char *filename = NULL;
1113 if (cmd == IOC_MDC_GETFILEINFO ||
1114 cmd == IOC_MDC_GETFILESTRIPE) {
1115 filename = getname((const char *)arg);
1116 if (IS_ERR(filename))
1117 RETURN(PTR_ERR(filename));
1119 rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1120 &lmmsize, &request);
1122 rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1126 body = req_capsule_server_get(&request->rq_pill,
1128 LASSERT(body != NULL);
1134 if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1135 cmd == LL_IOC_MDC_GETINFO))
1136 GOTO(skip_lmm, rc = 0);
1141 if (cmd == IOC_MDC_GETFILESTRIPE ||
1142 cmd == LL_IOC_LOV_GETSTRIPE) {
1143 lump = (struct lov_user_md *)arg;
1145 struct lov_user_mds_data *lmdp;
1146 lmdp = (struct lov_user_mds_data *)arg;
1147 lump = &lmdp->lmd_lmm;
1149 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1150 if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1151 GOTO(out_req, rc = -EFAULT);
1155 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1156 struct lov_user_mds_data *lmdp;
1159 st.st_dev = inode->i_sb->s_dev;
1160 st.st_mode = body->mode;
1161 st.st_nlink = body->nlink;
1162 st.st_uid = body->uid;
1163 st.st_gid = body->gid;
1164 st.st_rdev = body->rdev;
1165 st.st_size = body->size;
1166 st.st_blksize = CFS_PAGE_SIZE;
1167 st.st_blocks = body->blocks;
1168 st.st_atime = body->atime;
1169 st.st_mtime = body->mtime;
1170 st.st_ctime = body->ctime;
1171 st.st_ino = inode->i_ino;
1173 lmdp = (struct lov_user_mds_data *)arg;
1174 if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1175 GOTO(out_req, rc = -EFAULT);
1180 ptlrpc_req_finished(request);
1185 case IOC_LOV_GETINFO: {
1186 struct lov_user_mds_data *lumd;
1187 struct lov_stripe_md *lsm;
1188 struct lov_user_md *lum;
1189 struct lov_mds_md *lmm;
1193 lumd = (struct lov_user_mds_data *)arg;
1194 lum = &lumd->lmd_lmm;
1196 rc = ll_get_max_mdsize(sbi, &lmmsize);
1200 OBD_ALLOC_LARGE(lmm, lmmsize);
1201 if (cfs_copy_from_user(lmm, lum, lmmsize))
1202 GOTO(free_lmm, rc = -EFAULT);
1204 switch (lmm->lmm_magic) {
1205 case LOV_USER_MAGIC_V1:
1206 if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1208 /* swab objects first so that stripes num will be sane */
1209 lustre_swab_lov_user_md_objects(
1210 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1211 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1212 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1214 case LOV_USER_MAGIC_V3:
1215 if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1217 /* swab objects first so that stripes num will be sane */
1218 lustre_swab_lov_user_md_objects(
1219 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1220 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1221 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1224 GOTO(free_lmm, rc = -EINVAL);
1227 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1229 GOTO(free_lmm, rc = -ENOMEM);
1231 /* Perform glimpse_size operation. */
1232 memset(&st, 0, sizeof(st));
1234 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1238 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1239 GOTO(free_lsm, rc = -EFAULT);
1243 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1245 OBD_FREE_LARGE(lmm, lmmsize);
1248 case OBD_IOC_LLOG_CATINFO: {
1249 struct ptlrpc_request *req = NULL;
1254 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1259 if (!data->ioc_inlbuf1) {
1260 obd_ioctl_freedata(buf, len);
1264 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
1267 GOTO(out_catinfo, rc = -ENOMEM);
1269 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
1271 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
1274 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
1276 ptlrpc_request_free(req);
1277 GOTO(out_catinfo, rc);
1280 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
1281 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
1282 if (data->ioc_inllen2) {
1283 str = req_capsule_client_get(&req->rq_pill,
1285 memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
1288 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
1290 ptlrpc_request_set_replen(req);
1292 rc = ptlrpc_queue_wait(req);
1294 str = req_capsule_server_get(&req->rq_pill,
1296 if (cfs_copy_to_user(data->ioc_pbuf1, str,
1300 ptlrpc_req_finished(req);
1302 obd_ioctl_freedata(buf, len);
1305 case OBD_IOC_QUOTACHECK: {
1306 struct obd_quotactl *oqctl;
1309 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1310 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1313 OBD_ALLOC_PTR(oqctl);
1316 oqctl->qc_type = arg;
1317 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1319 CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1323 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1325 CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1327 OBD_FREE_PTR(oqctl);
1330 case OBD_IOC_POLL_QUOTACHECK: {
1331 struct if_quotacheck *check;
1333 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1334 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1337 OBD_ALLOC_PTR(check);
1341 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1344 CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1345 if (cfs_copy_to_user((void *)arg, check,
1347 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1351 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1354 CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1355 if (cfs_copy_to_user((void *)arg, check,
1357 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1361 OBD_FREE_PTR(check);
1364 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
1365 case LL_IOC_QUOTACTL_18: {
1366 /* copy the old 1.x quota struct for internal use, then copy
1367 * back into old format struct. For 1.8 compatibility. */
1368 struct if_quotactl_18 *qctl_18;
1369 struct if_quotactl *qctl_20;
1371 OBD_ALLOC_PTR(qctl_18);
1375 OBD_ALLOC_PTR(qctl_20);
1377 GOTO(out_quotactl_18, rc = -ENOMEM);
1379 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1380 GOTO(out_quotactl_20, rc = -ENOMEM);
1382 QCTL_COPY(qctl_20, qctl_18);
1383 qctl_20->qc_idx = 0;
1385 /* XXX: dqb_valid was borrowed as a flag to mark that
1386 * only mds quota is wanted */
1387 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1388 qctl_18->qc_dqblk.dqb_valid) {
1389 qctl_20->qc_valid = QC_MDTIDX;
1390 qctl_20->qc_dqblk.dqb_valid = 0;
1391 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1392 qctl_20->qc_valid = QC_UUID;
1393 qctl_20->obd_uuid = qctl_18->obd_uuid;
1395 qctl_20->qc_valid = QC_GENERAL;
1398 rc = quotactl_ioctl(sbi, qctl_20);
1401 QCTL_COPY(qctl_18, qctl_20);
1402 qctl_18->obd_uuid = qctl_20->obd_uuid;
1404 if (cfs_copy_to_user((void *)arg, qctl_18,
1410 OBD_FREE_PTR(qctl_20);
1412 OBD_FREE_PTR(qctl_18);
1416 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1417 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0) */
1418 case LL_IOC_QUOTACTL: {
1419 struct if_quotactl *qctl;
1421 OBD_ALLOC_PTR(qctl);
1425 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1426 GOTO(out_quotactl, rc = -EFAULT);
1428 rc = quotactl_ioctl(sbi, qctl);
1430 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1437 case OBD_IOC_GETNAME: {
1438 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1441 if (cfs_copy_to_user((void *)arg, obd->obd_name,
1442 strlen(obd->obd_name) + 1))
1446 case LL_IOC_FLUSHCTX:
1447 RETURN(ll_flush_ctx(inode));
1448 #ifdef CONFIG_FS_POSIX_ACL
1449 case LL_IOC_RMTACL: {
1450 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1451 inode == inode->i_sb->s_root->d_inode) {
1452 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1454 LASSERT(fd != NULL);
1455 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1457 fd->fd_flags |= LL_FILE_RMTACL;
1463 case LL_IOC_GETOBDCOUNT: {
1465 struct obd_export *exp;
1467 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1470 /* get ost count when count is zero, get mdt count otherwise */
1471 exp = count ? sbi->ll_md_exp : sbi->ll_dt_exp;
1472 vallen = sizeof(count);
1473 rc = obd_get_info(exp, sizeof(KEY_TGT_COUNT), KEY_TGT_COUNT,
1474 &vallen, &count, NULL);
1476 CERROR("get target count failed: %d\n", rc);
1480 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1485 case LL_IOC_PATH2FID:
1486 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1487 sizeof(struct lu_fid)))
1490 case LL_IOC_GET_CONNECT_FLAGS: {
1491 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1493 case OBD_IOC_CHANGELOG_SEND:
1494 case OBD_IOC_CHANGELOG_CLEAR:
1495 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1496 sizeof(struct ioc_changelog));
1498 case OBD_IOC_FID2PATH:
1499 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1500 case LL_IOC_HSM_CT_START:
1501 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1502 sizeof(struct lustre_kernelcomm));
1506 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1510 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1512 struct inode *inode = file->f_mapping->host;
1513 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1514 struct ll_sb_info *sbi = ll_i2sbi(inode);
1515 int api32 = ll_need_32bit_api(sbi);
1516 loff_t ret = -EINVAL;
1519 cfs_mutex_lock(&inode->i_mutex);
1524 offset += file->f_pos;
1530 offset += LL_DIR_END_OFF_32BIT;
1532 offset += LL_DIR_END_OFF;
1539 ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1540 (!api32 && offset <= LL_DIR_END_OFF))) {
1541 if (offset != file->f_pos) {
1542 if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1543 (!api32 && offset == LL_DIR_END_OFF))
1544 fd->fd_dir.lfd_pos = MDS_DIR_END_OFF;
1545 else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1546 fd->fd_dir.lfd_pos = offset << 32;
1548 fd->fd_dir.lfd_pos = offset;
1549 file->f_pos = offset;
1550 file->f_version = 0;
1557 cfs_mutex_unlock(&inode->i_mutex);
1561 int ll_dir_open(struct inode *inode, struct file *file)
1564 RETURN(ll_file_open(inode, file));
1567 int ll_dir_release(struct inode *inode, struct file *file)
1570 RETURN(ll_file_release(inode, file));
1573 struct file_operations ll_dir_operations = {
1574 .llseek = ll_dir_seek,
1575 .open = ll_dir_open,
1576 .release = ll_dir_release,
1577 .read = generic_read_dir,
1578 .readdir = ll_readdir,
1579 .ioctl = ll_dir_ioctl,