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 struct ll_inode_info *lli = ll_i2info(inode);
184 cfs_spin_lock(&lli->lli_sa_lock);
186 LASSERT(lli->lli_sai->sai_pid == cfs_curproc_pid());
188 LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
189 hash = lli->lli_sa_pos;
190 cfs_spin_unlock(&lli->lli_sa_lock);
192 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) hash "LPU64"\n",
193 inode->i_ino, inode->i_generation, inode, hash);
195 LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
197 OBD_ALLOC(page_pool, sizeof(page) * max_pages);
198 if (page_pool != NULL) {
199 page_pool[0] = page0;
204 for (npages = 1; npages < max_pages; npages++) {
205 page = page_cache_alloc_cold(inode->i_mapping);
208 page_pool[npages] = page;
211 oc = ll_mdscapa_get(inode);
212 rc = md_readpage(exp, ll_inode2fid(inode), oc, hash, page_pool, npages,
216 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
217 /* Checked by mdc_readpage() */
218 LASSERT(body != NULL);
220 if (body->valid & OBD_MD_FLSIZE)
221 cl_isize_write(inode, body->size);
223 nrdpgs = (request->rq_bulk->bd_nob_transferred+CFS_PAGE_SIZE-1)
225 SetPageUptodate(page0);
228 ptlrpc_req_finished(request);
230 CDEBUG(D_VFSTRACE, "read %d/%d pages\n", nrdpgs, npages);
232 ll_pagevec_init(&lru_pvec, 0);
233 for (i = 1; i < npages; i++) {
234 unsigned long offset;
239 if (rc < 0 || i >= nrdpgs) {
240 page_cache_release(page);
244 SetPageUptodate(page);
247 hash = le64_to_cpu(dp->ldp_hash_start);
250 offset = hash_x_index(hash, hash64);
252 prefetchw(&page->flags);
253 ret = ll_add_to_page_cache_lru(page, inode->i_mapping, offset,
257 page_cache_get(page);
258 if (ll_pagevec_add(&lru_pvec, page) == 0)
259 ll_pagevec_lru_add_file(&lru_pvec);
261 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
262 " %d\n", offset, ret);
264 page_cache_release(page);
266 ll_pagevec_lru_add_file(&lru_pvec);
268 if (page_pool != &page0)
269 OBD_FREE(page_pool, sizeof(struct page *) * max_pages);
274 #ifndef MS_HAS_NEW_AOPS
275 struct address_space_operations ll_dir_aops = {
276 .readpage = ll_dir_readpage,
279 struct address_space_operations_ext ll_dir_aops = {
280 .orig_aops.readpage = ll_dir_readpage,
284 static void ll_check_page(struct inode *dir, struct page *page)
286 /* XXX: check page format later */
287 SetPageChecked(page);
290 static void ll_release_page(struct page *page, __u64 hash,
291 __u64 start, __u64 end)
295 if (likely(page->mapping != NULL)) {
296 ll_truncate_complete_page(page);
300 CWARN("NULL mapping page %p, truncated by others: "
301 "hash("LPX64") | start("LPX64") | end("LPX64")\n",
302 page, hash, start, end);
304 page_cache_release(page);
308 * Find, kmap and return page that contains given hash.
310 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
311 __u64 *start, __u64 *end)
313 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
314 struct address_space *mapping = dir->i_mapping;
316 * Complement of hash is used as an index so that
317 * radix_tree_gang_lookup() can be used to find a page with starting
318 * hash _smaller_ than one we are looking for.
320 unsigned long offset = hash_x_index(*hash, hash64);
324 TREE_READ_LOCK_IRQ(mapping);
325 found = radix_tree_gang_lookup(&mapping->page_tree,
326 (void **)&page, offset, 1);
328 struct lu_dirpage *dp;
330 page_cache_get(page);
331 TREE_READ_UNLOCK_IRQ(mapping);
333 * In contrast to find_lock_page() we are sure that directory
334 * page cannot be truncated (while DLM lock is held) and,
335 * hence, can avoid restart.
337 * In fact, page cannot be locked here at all, because
338 * ll_dir_readpage() does synchronous io.
341 if (PageUptodate(page)) {
343 if (BITS_PER_LONG == 32 && hash64) {
344 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
345 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
348 *start = le64_to_cpu(dp->ldp_hash_start);
349 *end = le64_to_cpu(dp->ldp_hash_end);
351 LASSERTF(*start <= *hash, "start = "LPX64",end = "
352 LPX64",hash = "LPX64"\n", *start, *end, *hash);
353 CDEBUG(D_VFSTRACE, "page %lu [%llu %llu], hash "LPU64"\n",
354 offset, *start, *end, *hash);
355 if (*hash > *end || (*end != *start && *hash == *end)) {
357 * upon hash collision, remove this page,
358 * otherwise put page reference, and
359 * ll_get_dir_page() will issue RPC to fetch
362 if (dp->ldp_flags & cpu_to_le32(LDF_COLLIDE)) {
363 ll_release_page(page, *hash, *start, *end);
366 page_cache_release(page);
371 page_cache_release(page);
372 page = ERR_PTR(-EIO);
376 TREE_READ_UNLOCK_IRQ(mapping);
382 struct page *ll_get_dir_page(struct file *filp, struct inode *dir, __u64 hash,
383 int exact, struct ll_dir_chain *chain)
385 ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
386 struct address_space *mapping = dir->i_mapping;
387 struct lustre_handle lockh;
388 struct lu_dirpage *dp;
395 struct ll_inode_info *lli = ll_i2info(dir);
396 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
399 rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
400 ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
402 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
403 ll_md_blocking_ast, ldlm_completion_ast,
405 struct lookup_intent it = { .it_op = IT_READDIR };
406 struct ptlrpc_request *request;
407 struct md_op_data *op_data;
409 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
410 LUSTRE_OPC_ANY, NULL);
412 return (void *)op_data;
414 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
415 op_data, &lockh, NULL, 0, NULL, 0);
417 ll_finish_md_op_data(op_data);
419 request = (struct ptlrpc_request *)it.d.lustre.it_data;
421 ptlrpc_req_finished(request);
423 CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
424 PFID(ll_inode2fid(dir)), hash, rc);
428 /* for cross-ref object, l_ast_data of the lock may not be set,
429 * we reset it here */
430 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
433 ldlm_lock_dump_handle(D_OTHER, &lockh);
435 cfs_down(&lli->lli_readdir_sem);
436 page = ll_dir_page_locate(dir, &lhash, &start, &end);
438 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
439 PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
440 GOTO(out_unlock, page);
445 * XXX nikita: not entirely correct handling of a corner case:
446 * suppose hash chain of entries with hash value HASH crosses
447 * border between pages P0 and P1. First both P0 and P1 are
448 * cached, seekdir() is called for some entry from the P0 part
449 * of the chain. Later P0 goes out of cache. telldir(HASH)
450 * happens and finds P1, as it starts with matching hash
451 * value. Remaining entries from P0 part of the chain are
452 * skipped. (Is that really a bug?)
454 * Possible solutions: 0. don't cache P1 is such case, handle
455 * it as an "overflow" page. 1. invalidate all pages at
456 * once. 2. use HASH|1 as an index for P1.
458 if (exact && lhash != start) {
460 * readdir asked for a page starting _exactly_ from
461 * given hash, but cache contains stale page, with
462 * entries with smaller hash values. Stale page should
463 * be invalidated, and new one fetched.
465 CDEBUG(D_OTHER, "Stale readpage page %p: "
466 "start = "LPX64",end = "LPX64"hash ="LPX64"\n",
467 page, start, end, lhash);
468 ll_release_page(page, lhash, start, end);
470 GOTO(hash_collision, page);
474 page = read_cache_page(mapping, hash_x_index(hash, hash64),
475 (filler_t*)mapping->a_ops->readpage, filp);
477 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
478 PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
479 GOTO(out_unlock, page);
484 if (!PageUptodate(page)) {
485 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
486 PFID(ll_inode2fid(dir)), hash, -5);
489 if (!PageChecked(page))
490 ll_check_page(dir, page);
491 if (PageError(page)) {
492 CERROR("page error: "DFID" at "LPU64": rc %d\n",
493 PFID(ll_inode2fid(dir)), hash, -5);
497 dp = page_address(page);
498 if (BITS_PER_LONG == 32 && hash64) {
499 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
500 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
503 start = le64_to_cpu(dp->ldp_hash_start);
504 end = le64_to_cpu(dp->ldp_hash_end);
508 LASSERT(start == lhash);
509 CWARN("Page-wide hash collision: "LPU64"\n", end);
510 if (BITS_PER_LONG == 32 && hash64)
511 CWARN("Real page-wide hash collision at ["LPU64" "LPU64
512 "] with hash "LPU64"\n",
513 le64_to_cpu(dp->ldp_hash_start),
514 le64_to_cpu(dp->ldp_hash_end), hash);
516 * Fetch whole overflow chain...
523 cfs_up(&lli->lli_readdir_sem);
524 ldlm_lock_decref(&lockh, mode);
529 page = ERR_PTR(-EIO);
533 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
535 struct inode *inode = filp->f_dentry->d_inode;
536 struct ll_inode_info *info = ll_i2info(inode);
537 struct ll_sb_info *sbi = ll_i2sbi(inode);
538 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
539 __u64 pos = fd->fd_dir.lfd_pos;
540 int api32 = ll_need_32bit_api(sbi);
541 int hash64 = sbi->ll_flags & LL_SBI_64BIT_HASH;
543 struct ll_dir_chain chain;
549 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu 32bit_api %d\n",
550 inode->i_ino, inode->i_generation, inode,
551 (unsigned long)pos, i_size_read(inode), api32);
553 if (pos == MDS_DIR_END_OFF)
562 ll_dir_chain_init(&chain);
564 fd->fd_dir.lfd_next = pos;
565 page = ll_get_dir_page(filp, inode, pos, 0, &chain);
567 while (rc == 0 && !done) {
568 struct lu_dirpage *dp;
569 struct lu_dirent *ent;
573 * If page is empty (end of directory is reached),
576 __u64 hash = MDS_DIR_END_OFF;
579 dp = page_address(page);
580 for (ent = lu_dirent_start(dp); ent != NULL && !done;
581 ent = lu_dirent_next(ent)) {
589 * XXX: implement correct swabbing here.
592 hash = le64_to_cpu(ent->lde_hash);
595 * Skip until we find target hash
600 namelen = le16_to_cpu(ent->lde_namelen);
611 fid_le_to_cpu(&fid, &ent->lde_fid);
612 ino = cl_fid_build_ino(&fid, api32);
613 type = ll_dirent_type_get(ent);
614 /* For 'll_nfs_get_name_filldir()', it will try
615 * to access the 'ent' through its 'lde_name',
616 * so the parameter 'name' for 'filldir()' must
617 * be part of the 'ent'. */
618 done = filldir(cookie, ent->lde_name, namelen,
621 next = le64_to_cpu(dp->ldp_hash_end);
625 if (pos == MDS_DIR_END_OFF) {
627 * End of directory reached.
630 } else if (1 /* chain is exhausted*/) {
632 * Normal case: continue to the next
635 fd->fd_dir.lfd_next = pos;
636 page = ll_get_dir_page(filp, inode, pos,
640 * go into overflow page.
648 CERROR("error reading dir "DFID" at %lu: rc %d\n",
649 PFID(&info->lli_fid), (unsigned long)pos, rc);
653 fd->fd_dir.lfd_pos = pos;
654 if (pos == MDS_DIR_END_OFF) {
656 filp->f_pos = LL_DIR_END_OFF_32BIT;
658 filp->f_pos = LL_DIR_END_OFF;
661 filp->f_pos = pos >> 32;
665 filp->f_version = inode->i_version;
666 touch_atime(filp->f_vfsmnt, filp->f_dentry);
668 ll_dir_chain_fini(&chain);
673 int ll_send_mgc_param(struct obd_export *mgc, char *string)
675 struct mgs_send_param *msp;
682 strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
683 rc = obd_set_info_async(mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
684 sizeof(struct mgs_send_param), msp, NULL);
686 CERROR("Failed to set parameter: %d\n", rc);
692 char *ll_get_fsname(struct inode *inode)
694 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
698 OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
699 len = strlen(lsi->lsi_lmd->lmd_profile);
700 ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
701 if (ptr && (strcmp(ptr, "-client") == 0))
703 strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
709 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
712 struct ll_sb_info *sbi = ll_i2sbi(inode);
713 struct md_op_data *op_data;
714 struct ptlrpc_request *req = NULL;
716 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
717 struct obd_device *mgc = lsi->lsi_mgc;
718 char *fsname = NULL, *param = NULL;
723 * This is coming from userspace, so should be in
724 * local endian. But the MDS would like it in little
725 * endian, so we swab it before we send it.
727 switch (lump->lmm_magic) {
728 case LOV_USER_MAGIC_V1: {
729 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
730 lustre_swab_lov_user_md_v1(lump);
731 lum_size = sizeof(struct lov_user_md_v1);
734 case LOV_USER_MAGIC_V3: {
735 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
736 lustre_swab_lov_user_md_v3(
737 (struct lov_user_md_v3 *)lump);
738 lum_size = sizeof(struct lov_user_md_v3);
742 CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
743 " %#08x != %#08x nor %#08x\n",
744 lump->lmm_magic, LOV_USER_MAGIC_V1,
750 lum_size = sizeof(struct lov_user_md_v1);
753 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
754 LUSTRE_OPC_ANY, NULL);
756 RETURN(PTR_ERR(op_data));
758 /* swabbing is done in lov_setstripe() on server side */
759 rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
760 NULL, 0, &req, NULL);
761 ll_finish_md_op_data(op_data);
762 ptlrpc_req_finished(req);
764 if (rc != -EPERM && rc != -EACCES)
765 CERROR("mdc_setattr fails: rc = %d\n", rc);
768 /* In the following we use the fact that LOV_USER_MAGIC_V1 and
769 LOV_USER_MAGIC_V3 have the same initial fields so we do not
770 need the make the distiction between the 2 versions */
771 if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
772 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
774 /* Get fsname and assume devname to be -MDT0000. */
775 fsname = ll_get_fsname(inode);
776 /* Set root stripesize */
777 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
778 lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
779 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
783 /* Set root stripecount */
784 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
785 lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
786 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
790 /* Set root stripeoffset */
791 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
792 lump ? le16_to_cpu(lump->lmm_stripe_offset) :
793 (typeof(lump->lmm_stripe_offset))(-1));
794 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
799 OBD_FREE(fsname, MGS_PARAM_MAXLEN);
801 OBD_FREE(param, MGS_PARAM_MAXLEN);
806 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
807 int *lmm_size, struct ptlrpc_request **request)
809 struct ll_sb_info *sbi = ll_i2sbi(inode);
810 struct mdt_body *body;
811 struct lov_mds_md *lmm = NULL;
812 struct ptlrpc_request *req = NULL;
814 struct md_op_data *op_data;
816 rc = ll_get_max_mdsize(sbi, &lmmsize);
820 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
821 0, lmmsize, LUSTRE_OPC_ANY,
826 op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
827 rc = md_getattr(sbi->ll_md_exp, op_data, &req);
828 ll_finish_md_op_data(op_data);
830 CDEBUG(D_INFO, "md_getattr failed on inode "
831 "%lu/%u: rc %d\n", inode->i_ino,
832 inode->i_generation, rc);
836 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
837 LASSERT(body != NULL);
839 lmmsize = body->eadatasize;
841 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
843 GOTO(out, rc = -ENODATA);
846 lmm = req_capsule_server_sized_get(&req->rq_pill,
847 &RMF_MDT_MD, lmmsize);
848 LASSERT(lmm != NULL);
851 * This is coming from the MDS, so is probably in
852 * little endian. We convert it to host endian before
853 * passing it to userspace.
855 /* We don't swab objects for directories */
856 switch (le32_to_cpu(lmm->lmm_magic)) {
858 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
859 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
862 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
863 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
866 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
877 * Get MDT index for the inode.
879 int ll_get_mdt_idx(struct inode *inode)
881 struct ll_sb_info *sbi = ll_i2sbi(inode);
882 struct md_op_data *op_data;
886 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
887 0, LUSTRE_OPC_ANY, NULL);
891 op_data->op_valid |= OBD_MD_MDTIDX;
892 rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
893 mdtidx = op_data->op_mds;
894 ll_finish_md_op_data(op_data);
896 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
902 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
910 if (cfs_copy_from_user(ptr, data, len)) {
914 rc = obd_iocontrol(cmd, exp, len, data, NULL);
919 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
921 int cmd = qctl->qc_cmd;
922 int type = qctl->qc_type;
923 int id = qctl->qc_id;
924 int valid = qctl->qc_valid;
929 case LUSTRE_Q_INVALIDATE:
930 case LUSTRE_Q_FINVALIDATE:
935 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
936 sbi->ll_flags & LL_SBI_RMT_CLIENT)
940 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
941 (type == GRPQUOTA && !in_egroup_p(id))) &&
942 (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
943 sbi->ll_flags & LL_SBI_RMT_CLIENT))
949 CERROR("unsupported quotactl op: %#x\n", cmd);
953 if (valid != QC_GENERAL) {
954 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
957 if (cmd == Q_GETINFO)
958 qctl->qc_cmd = Q_GETOINFO;
959 else if (cmd == Q_GETQUOTA)
960 qctl->qc_cmd = Q_GETOQUOTA;
966 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
967 sizeof(*qctl), qctl, NULL);
970 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
971 sizeof(*qctl), qctl, NULL);
974 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
975 sizeof(*qctl), qctl, NULL);
977 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
979 sizeof(*qctl), qctl, NULL);
991 struct obd_quotactl *oqctl;
993 OBD_ALLOC_PTR(oqctl);
997 QCTL_COPY(oqctl, qctl);
998 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1000 if (rc != -EALREADY && cmd == Q_QUOTAON) {
1001 oqctl->qc_cmd = Q_QUOTAOFF;
1002 obd_quotactl(sbi->ll_md_exp, oqctl);
1004 OBD_FREE_PTR(oqctl);
1008 QCTL_COPY(qctl, oqctl);
1009 OBD_FREE_PTR(oqctl);
1015 static int ll_dir_ioctl(struct inode *inode, struct file *file,
1016 unsigned int cmd, unsigned long arg)
1018 struct ll_sb_info *sbi = ll_i2sbi(inode);
1019 struct obd_ioctl_data *data;
1023 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1024 inode->i_ino, inode->i_generation, inode, cmd);
1026 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1027 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1030 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1032 case FSFILT_IOC_GETFLAGS:
1033 case FSFILT_IOC_SETFLAGS:
1034 RETURN(ll_iocontrol(inode, file, cmd, arg));
1035 case FSFILT_IOC_GETVERSION_OLD:
1036 case FSFILT_IOC_GETVERSION:
1037 RETURN(put_user(inode->i_generation, (int *)arg));
1038 /* We need to special case any other ioctls we want to handle,
1039 * to send them to the MDS/OST as appropriate and to properly
1040 * network encode the arg field.
1041 case FSFILT_IOC_SETVERSION_OLD:
1042 case FSFILT_IOC_SETVERSION:
1044 case LL_IOC_GET_MDTIDX: {
1047 mdtidx = ll_get_mdt_idx(inode);
1051 if (put_user((int)mdtidx, (int*)arg))
1056 case IOC_MDC_LOOKUP: {
1057 struct ptlrpc_request *request = NULL;
1058 int namelen, len = 0;
1061 struct md_op_data *op_data;
1063 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1068 filename = data->ioc_inlbuf1;
1069 namelen = strlen(filename);
1072 CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1073 GOTO(out_free, rc = -EINVAL);
1076 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1077 0, LUSTRE_OPC_ANY, NULL);
1078 if (op_data == NULL)
1079 GOTO(out_free, rc = -ENOMEM);
1081 op_data->op_valid = OBD_MD_FLID;
1082 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1083 ll_finish_md_op_data(op_data);
1085 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1088 ptlrpc_req_finished(request);
1091 obd_ioctl_freedata(buf, len);
1094 case LL_IOC_LOV_SETSTRIPE: {
1095 struct lov_user_md_v3 lumv3;
1096 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1097 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1098 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1100 int set_default = 0;
1102 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1103 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1104 sizeof(lumv3p->lmm_objects[0]));
1105 /* first try with v1 which is smaller than v3 */
1106 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1109 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1110 if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1114 if (inode->i_sb->s_root == file->f_dentry)
1117 /* in v1 and v3 cases lumv1 points to data */
1118 rc = ll_dir_setstripe(inode, lumv1, set_default);
1122 case LL_IOC_OBD_STATFS:
1123 RETURN(ll_obd_statfs(inode, (void *)arg));
1124 case LL_IOC_LOV_GETSTRIPE:
1125 case LL_IOC_MDC_GETINFO:
1126 case IOC_MDC_GETFILEINFO:
1127 case IOC_MDC_GETFILESTRIPE: {
1128 struct ptlrpc_request *request = NULL;
1129 struct lov_user_md *lump;
1130 struct lov_mds_md *lmm = NULL;
1131 struct mdt_body *body;
1132 char *filename = NULL;
1135 if (cmd == IOC_MDC_GETFILEINFO ||
1136 cmd == IOC_MDC_GETFILESTRIPE) {
1137 filename = getname((const char *)arg);
1138 if (IS_ERR(filename))
1139 RETURN(PTR_ERR(filename));
1141 rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1142 &lmmsize, &request);
1144 rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1148 body = req_capsule_server_get(&request->rq_pill,
1150 LASSERT(body != NULL);
1156 if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1157 cmd == LL_IOC_MDC_GETINFO))
1158 GOTO(skip_lmm, rc = 0);
1163 if (cmd == IOC_MDC_GETFILESTRIPE ||
1164 cmd == LL_IOC_LOV_GETSTRIPE) {
1165 lump = (struct lov_user_md *)arg;
1167 struct lov_user_mds_data *lmdp;
1168 lmdp = (struct lov_user_mds_data *)arg;
1169 lump = &lmdp->lmd_lmm;
1171 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1172 if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1173 GOTO(out_req, rc = -EFAULT);
1177 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1178 struct lov_user_mds_data *lmdp;
1181 st.st_dev = inode->i_sb->s_dev;
1182 st.st_mode = body->mode;
1183 st.st_nlink = body->nlink;
1184 st.st_uid = body->uid;
1185 st.st_gid = body->gid;
1186 st.st_rdev = body->rdev;
1187 st.st_size = body->size;
1188 st.st_blksize = CFS_PAGE_SIZE;
1189 st.st_blocks = body->blocks;
1190 st.st_atime = body->atime;
1191 st.st_mtime = body->mtime;
1192 st.st_ctime = body->ctime;
1193 st.st_ino = inode->i_ino;
1195 lmdp = (struct lov_user_mds_data *)arg;
1196 if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1197 GOTO(out_req, rc = -EFAULT);
1202 ptlrpc_req_finished(request);
1207 case IOC_LOV_GETINFO: {
1208 struct lov_user_mds_data *lumd;
1209 struct lov_stripe_md *lsm;
1210 struct lov_user_md *lum;
1211 struct lov_mds_md *lmm;
1215 lumd = (struct lov_user_mds_data *)arg;
1216 lum = &lumd->lmd_lmm;
1218 rc = ll_get_max_mdsize(sbi, &lmmsize);
1222 OBD_ALLOC_LARGE(lmm, lmmsize);
1223 if (cfs_copy_from_user(lmm, lum, lmmsize))
1224 GOTO(free_lmm, rc = -EFAULT);
1226 switch (lmm->lmm_magic) {
1227 case LOV_USER_MAGIC_V1:
1228 if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1230 /* swab objects first so that stripes num will be sane */
1231 lustre_swab_lov_user_md_objects(
1232 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1233 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1234 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1236 case LOV_USER_MAGIC_V3:
1237 if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1239 /* swab objects first so that stripes num will be sane */
1240 lustre_swab_lov_user_md_objects(
1241 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1242 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1243 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1246 GOTO(free_lmm, rc = -EINVAL);
1249 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1251 GOTO(free_lmm, rc = -ENOMEM);
1253 /* Perform glimpse_size operation. */
1254 memset(&st, 0, sizeof(st));
1256 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1260 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1261 GOTO(free_lsm, rc = -EFAULT);
1265 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1267 OBD_FREE_LARGE(lmm, lmmsize);
1270 case OBD_IOC_LLOG_CATINFO: {
1271 struct ptlrpc_request *req = NULL;
1276 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1281 if (!data->ioc_inlbuf1) {
1282 obd_ioctl_freedata(buf, len);
1286 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
1289 GOTO(out_catinfo, rc = -ENOMEM);
1291 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
1293 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
1296 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
1298 ptlrpc_request_free(req);
1299 GOTO(out_catinfo, rc);
1302 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
1303 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
1304 if (data->ioc_inllen2) {
1305 str = req_capsule_client_get(&req->rq_pill,
1307 memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
1310 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
1312 ptlrpc_request_set_replen(req);
1314 rc = ptlrpc_queue_wait(req);
1316 str = req_capsule_server_get(&req->rq_pill,
1318 if (cfs_copy_to_user(data->ioc_pbuf1, str,
1322 ptlrpc_req_finished(req);
1324 obd_ioctl_freedata(buf, len);
1327 case OBD_IOC_QUOTACHECK: {
1328 struct obd_quotactl *oqctl;
1331 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1332 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1335 OBD_ALLOC_PTR(oqctl);
1338 oqctl->qc_type = arg;
1339 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1341 CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1345 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1347 CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1349 OBD_FREE_PTR(oqctl);
1352 case OBD_IOC_POLL_QUOTACHECK: {
1353 struct if_quotacheck *check;
1355 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1356 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1359 OBD_ALLOC_PTR(check);
1363 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1366 CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1367 if (cfs_copy_to_user((void *)arg, check,
1369 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1373 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1376 CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1377 if (cfs_copy_to_user((void *)arg, check,
1379 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1383 OBD_FREE_PTR(check);
1386 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
1387 case LL_IOC_QUOTACTL_18: {
1388 /* copy the old 1.x quota struct for internal use, then copy
1389 * back into old format struct. For 1.8 compatibility. */
1390 struct if_quotactl_18 *qctl_18;
1391 struct if_quotactl *qctl_20;
1393 OBD_ALLOC_PTR(qctl_18);
1397 OBD_ALLOC_PTR(qctl_20);
1399 GOTO(out_quotactl_18, rc = -ENOMEM);
1401 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1402 GOTO(out_quotactl_20, rc = -ENOMEM);
1404 QCTL_COPY(qctl_20, qctl_18);
1405 qctl_20->qc_idx = 0;
1407 /* XXX: dqb_valid was borrowed as a flag to mark that
1408 * only mds quota is wanted */
1409 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1410 qctl_18->qc_dqblk.dqb_valid) {
1411 qctl_20->qc_valid = QC_MDTIDX;
1412 qctl_20->qc_dqblk.dqb_valid = 0;
1413 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1414 qctl_20->qc_valid = QC_UUID;
1415 qctl_20->obd_uuid = qctl_18->obd_uuid;
1417 qctl_20->qc_valid = QC_GENERAL;
1420 rc = quotactl_ioctl(sbi, qctl_20);
1423 QCTL_COPY(qctl_18, qctl_20);
1424 qctl_18->obd_uuid = qctl_20->obd_uuid;
1426 if (cfs_copy_to_user((void *)arg, qctl_18,
1432 OBD_FREE_PTR(qctl_20);
1434 OBD_FREE_PTR(qctl_18);
1438 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1439 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0) */
1440 case LL_IOC_QUOTACTL: {
1441 struct if_quotactl *qctl;
1443 OBD_ALLOC_PTR(qctl);
1447 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1448 GOTO(out_quotactl, rc = -EFAULT);
1450 rc = quotactl_ioctl(sbi, qctl);
1452 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1459 case OBD_IOC_GETNAME: {
1460 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1463 if (cfs_copy_to_user((void *)arg, obd->obd_name,
1464 strlen(obd->obd_name) + 1))
1468 case LL_IOC_FLUSHCTX:
1469 RETURN(ll_flush_ctx(inode));
1470 #ifdef CONFIG_FS_POSIX_ACL
1471 case LL_IOC_RMTACL: {
1472 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1473 inode == inode->i_sb->s_root->d_inode) {
1474 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1476 LASSERT(fd != NULL);
1477 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1479 fd->fd_flags |= LL_FILE_RMTACL;
1485 case LL_IOC_GETOBDCOUNT: {
1488 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1493 struct lov_obd *lov = &sbi->ll_dt_exp->exp_obd->u.lov;
1494 count = lov->desc.ld_tgt_count;
1497 struct lmv_obd *lmv = &sbi->ll_md_exp->exp_obd->u.lmv;
1498 count = lmv->desc.ld_tgt_count;
1501 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1506 case LL_IOC_PATH2FID:
1507 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1508 sizeof(struct lu_fid)))
1511 case LL_IOC_GET_CONNECT_FLAGS: {
1512 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1514 case OBD_IOC_CHANGELOG_SEND:
1515 case OBD_IOC_CHANGELOG_CLEAR:
1516 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1517 sizeof(struct ioc_changelog));
1519 case OBD_IOC_FID2PATH:
1520 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1521 case LL_IOC_HSM_CT_START:
1522 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1523 sizeof(struct lustre_kernelcomm));
1527 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1531 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1533 struct inode *inode = file->f_mapping->host;
1534 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1535 struct ll_sb_info *sbi = ll_i2sbi(inode);
1536 int api32 = ll_need_32bit_api(sbi);
1537 loff_t ret = -EINVAL;
1540 cfs_mutex_lock(&inode->i_mutex);
1545 offset += file->f_pos;
1551 offset += LL_DIR_END_OFF_32BIT;
1553 offset += LL_DIR_END_OFF;
1560 ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1561 (!api32 && offset <= LL_DIR_END_OFF))) {
1562 if (offset != file->f_pos) {
1563 if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1564 (!api32 && offset == LL_DIR_END_OFF))
1565 fd->fd_dir.lfd_pos = MDS_DIR_END_OFF;
1566 else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1567 fd->fd_dir.lfd_pos = offset << 32;
1569 fd->fd_dir.lfd_pos = offset;
1570 file->f_pos = offset;
1571 file->f_version = 0;
1578 cfs_mutex_unlock(&inode->i_mutex);
1582 int ll_dir_open(struct inode *inode, struct file *file)
1585 RETURN(ll_file_open(inode, file));
1588 int ll_dir_release(struct inode *inode, struct file *file)
1591 RETURN(ll_file_release(inode, file));
1594 struct file_operations ll_dir_operations = {
1595 .llseek = ll_dir_seek,
1596 .open = ll_dir_open,
1597 .release = ll_dir_release,
1598 .read = generic_read_dir,
1599 .readdir = ll_readdir,
1600 .ioctl = ll_dir_ioctl