1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Directory code for lustre client.
42 #include <linux/pagemap.h>
44 #include <linux/version.h>
45 #include <linux/smp_lock.h>
46 #include <asm/uaccess.h>
47 #include <linux/buffer_head.h> // for wait_on_buffer
49 #define DEBUG_SUBSYSTEM S_LLITE
51 #include <obd_support.h>
52 #include <obd_class.h>
53 #include <lustre_lib.h>
54 #include <lustre/lustre_idl.h>
55 #include <lustre_lite.h>
56 #include <lustre_dlm.h>
57 #include <lustre_fid.h>
58 #include "llite_internal.h"
60 #ifndef HAVE_PAGE_CHECKED
61 #ifdef HAVE_PG_FS_MISC
62 #define PageChecked(page) test_bit(PG_fs_misc, &(page)->flags)
63 #define SetPageChecked(page) set_bit(PG_fs_misc, &(page)->flags)
65 #error PageChecked or PageFsMisc not defined in kernel
70 * (new) readdir implementation overview.
72 * Original lustre readdir implementation cached exact copy of raw directory
73 * pages on the client. These pages were indexed in client page cache by
74 * logical offset in the directory file. This design, while very simple and
75 * intuitive had some inherent problems:
77 * . it implies that byte offset to the directory entry serves as a
78 * telldir(3)/seekdir(3) cookie, but that offset is not stable: in
79 * ext3/htree directory entries may move due to splits, and more
82 * . it is incompatible with the design of split directories for cmd3,
83 * that assumes that names are distributed across nodes based on their
84 * hash, and so readdir should be done in hash order.
86 * New readdir implementation does readdir in hash order, and uses hash of a
87 * file name as a telldir/seekdir cookie. This led to number of complications:
89 * . hash is not unique, so it cannot be used to index cached directory
90 * pages on the client (note, that it requires a whole pageful of hash
91 * collided entries to cause two pages to have identical hashes);
93 * . hash is not unique, so it cannot, strictly speaking, be used as an
94 * entry cookie. ext3/htree has the same problem and lustre implementation
95 * mimics their solution: seekdir(hash) positions directory at the first
96 * entry with the given hash.
102 * Client caches directory pages using hash of the first entry as an index. As
103 * noted above hash is not unique, so this solution doesn't work as is:
104 * special processing is needed for "page hash chains" (i.e., sequences of
105 * pages filled with entries all having the same hash value).
107 * First, such chains have to be detected. To this end, server returns to the
108 * client the hash of the first entry on the page next to one returned. When
109 * client detects that this hash is the same as hash of the first entry on the
110 * returned page, page hash collision has to be handled. Pages in the
111 * hash chain, except first one, are termed "overflow pages".
113 * Solution to index uniqueness problem is to not cache overflow
114 * pages. Instead, when page hash collision is detected, all overflow pages
115 * from emerging chain are immediately requested from the server and placed in
116 * a special data structure (struct ll_dir_chain). This data structure is used
117 * by ll_readdir() to process entries from overflow pages. When readdir
118 * invocation finishes, overflow pages are discarded. If page hash collision
119 * chain weren't completely processed, next call to readdir will again detect
120 * page hash collision, again read overflow pages in, process next portion of
121 * entries and again discard the pages. This is not as wasteful as it looks,
122 * because, given reasonable hash, page hash collisions are extremely rare.
124 * 1. directory positioning
126 * When seekdir(hash) is called, original
137 * identification of and access to overflow pages
147 /* returns the page unlocked, but with a reference */
148 static int ll_dir_readpage(struct file *file, struct page *page)
150 struct inode *inode = page->mapping->host;
151 struct ptlrpc_request *request;
152 struct mdt_body *body;
159 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
161 hash = fd->fd_dir.lfd_next;
163 struct ll_inode_info *lli = ll_i2info(inode);
165 cfs_spin_lock(&lli->lli_sa_lock);
167 LASSERT(lli->lli_sai->sai_pid == cfs_curproc_pid());
169 LASSERT(lli->lli_opendir_pid == cfs_curproc_pid());
170 hash = lli->lli_sa_pos;
171 cfs_spin_unlock(&lli->lli_sa_lock);
173 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
174 inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
176 oc = ll_mdscapa_get(inode);
177 rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
178 oc, hash, page, &request);
181 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
182 /* Checked by mdc_readpage() */
183 LASSERT(body != NULL);
185 if (body->valid & OBD_MD_FLSIZE)
186 cl_isize_write(inode, body->size);
187 SetPageUptodate(page);
189 ptlrpc_req_finished(request);
196 #ifndef MS_HAS_NEW_AOPS
197 struct address_space_operations ll_dir_aops = {
198 .readpage = ll_dir_readpage,
201 struct address_space_operations_ext ll_dir_aops = {
202 .orig_aops.readpage = ll_dir_readpage,
206 static void ll_check_page(struct inode *dir, struct page *page)
208 /* XXX: check page format later */
209 SetPageChecked(page);
212 static void ll_release_page(struct page *page, __u64 hash,
213 __u64 start, __u64 end)
217 if (likely(page->mapping != NULL)) {
218 ll_truncate_complete_page(page);
222 CWARN("NULL mapping page %p, truncated by others: "
223 "hash("LPX64") | start("LPX64") | end("LPX64")\n",
224 page, hash, start, end);
226 page_cache_release(page);
230 * Find, kmap and return page that contains given hash.
232 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
233 __u64 *start, __u64 *end)
235 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
236 struct address_space *mapping = dir->i_mapping;
238 * Complement of hash is used as an index so that
239 * radix_tree_gang_lookup() can be used to find a page with starting
240 * hash _smaller_ than one we are looking for.
242 unsigned long offset = hash_x_index(*hash, hash64);
246 TREE_READ_LOCK_IRQ(mapping);
247 found = radix_tree_gang_lookup(&mapping->page_tree,
248 (void **)&page, offset, 1);
250 struct lu_dirpage *dp;
252 page_cache_get(page);
253 TREE_READ_UNLOCK_IRQ(mapping);
255 * In contrast to find_lock_page() we are sure that directory
256 * page cannot be truncated (while DLM lock is held) and,
257 * hence, can avoid restart.
259 * In fact, page cannot be locked here at all, because
260 * ll_dir_readpage() does synchronous io.
263 if (PageUptodate(page)) {
265 if (BITS_PER_LONG == 32 && hash64) {
266 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
267 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
270 *start = le64_to_cpu(dp->ldp_hash_start);
271 *end = le64_to_cpu(dp->ldp_hash_end);
273 LASSERTF(*start <= *hash, "start = "LPX64",end = "
274 LPX64",hash = "LPX64"\n", *start, *end, *hash);
275 if (*hash > *end || (*end != *start && *hash == *end)) {
276 ll_release_page(page, *hash, *start, *end);
280 page_cache_release(page);
281 page = ERR_PTR(-EIO);
285 TREE_READ_UNLOCK_IRQ(mapping);
291 struct page *ll_get_dir_page(struct file *filp, struct inode *dir, __u64 hash,
292 int exact, struct ll_dir_chain *chain)
294 ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
295 struct address_space *mapping = dir->i_mapping;
296 struct lustre_handle lockh;
297 struct lu_dirpage *dp;
304 struct ll_inode_info *lli = ll_i2info(dir);
305 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
308 rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
309 ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
311 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
312 ll_md_blocking_ast, ldlm_completion_ast,
314 struct lookup_intent it = { .it_op = IT_READDIR };
315 struct ptlrpc_request *request;
316 struct md_op_data *op_data;
318 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
319 LUSTRE_OPC_ANY, NULL);
321 return (void *)op_data;
323 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
324 op_data, &lockh, NULL, 0, NULL, 0);
326 ll_finish_md_op_data(op_data);
328 request = (struct ptlrpc_request *)it.d.lustre.it_data;
330 ptlrpc_req_finished(request);
332 CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
333 PFID(ll_inode2fid(dir)), hash, rc);
337 /* for cross-ref object, l_ast_data of the lock may not be set,
338 * we reset it here */
339 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
342 ldlm_lock_dump_handle(D_OTHER, &lockh);
344 cfs_down(&lli->lli_readdir_sem);
345 page = ll_dir_page_locate(dir, &lhash, &start, &end);
347 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
348 PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
349 GOTO(out_unlock, page);
354 * XXX nikita: not entirely correct handling of a corner case:
355 * suppose hash chain of entries with hash value HASH crosses
356 * border between pages P0 and P1. First both P0 and P1 are
357 * cached, seekdir() is called for some entry from the P0 part
358 * of the chain. Later P0 goes out of cache. telldir(HASH)
359 * happens and finds P1, as it starts with matching hash
360 * value. Remaining entries from P0 part of the chain are
361 * skipped. (Is that really a bug?)
363 * Possible solutions: 0. don't cache P1 is such case, handle
364 * it as an "overflow" page. 1. invalidate all pages at
365 * once. 2. use HASH|1 as an index for P1.
367 if (exact && lhash != start) {
369 * readdir asked for a page starting _exactly_ from
370 * given hash, but cache contains stale page, with
371 * entries with smaller hash values. Stale page should
372 * be invalidated, and new one fetched.
374 CDEBUG(D_OTHER, "Stale readpage page %p: "
375 "start = "LPX64",end = "LPX64"hash ="LPX64"\n",
376 page, start, end, lhash);
377 ll_release_page(page, lhash, start, end);
379 GOTO(hash_collision, page);
383 page = read_cache_page(mapping, hash_x_index(hash, hash64),
384 (filler_t*)mapping->a_ops->readpage, filp);
386 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
387 PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
388 GOTO(out_unlock, page);
393 if (!PageUptodate(page)) {
394 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
395 PFID(ll_inode2fid(dir)), hash, -5);
398 if (!PageChecked(page))
399 ll_check_page(dir, page);
400 if (PageError(page)) {
401 CERROR("page error: "DFID" at "LPU64": rc %d\n",
402 PFID(ll_inode2fid(dir)), hash, -5);
406 dp = page_address(page);
407 if (BITS_PER_LONG == 32 && hash64) {
408 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
409 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
412 start = le64_to_cpu(dp->ldp_hash_start);
413 end = le64_to_cpu(dp->ldp_hash_end);
417 LASSERT(start == lhash);
418 CWARN("Page-wide hash collision: "LPU64"\n", end);
419 if (BITS_PER_LONG == 32 && hash64)
420 CWARN("Real page-wide hash collision at ["LPU64" "LPU64
421 "] with hash "LPU64"\n",
422 le64_to_cpu(dp->ldp_hash_start),
423 le64_to_cpu(dp->ldp_hash_end), hash);
425 * Fetch whole overflow chain...
432 cfs_up(&lli->lli_readdir_sem);
433 ldlm_lock_decref(&lockh, mode);
438 page = ERR_PTR(-EIO);
442 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
444 struct inode *inode = filp->f_dentry->d_inode;
445 struct ll_inode_info *info = ll_i2info(inode);
446 struct ll_sb_info *sbi = ll_i2sbi(inode);
447 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
448 __u64 pos = fd->fd_dir.lfd_pos;
449 int api32 = ll_need_32bit_api(sbi);
450 int hash64 = sbi->ll_flags & LL_SBI_64BIT_HASH;
452 struct ll_dir_chain chain;
458 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu 32bit_api %d\n",
459 inode->i_ino, inode->i_generation, inode,
460 (unsigned long)pos, i_size_read(inode), api32);
462 if (pos == MDS_DIR_END_OFF)
471 ll_dir_chain_init(&chain);
473 fd->fd_dir.lfd_next = pos;
474 page = ll_get_dir_page(filp, inode, pos, 0, &chain);
476 while (rc == 0 && !done) {
477 struct lu_dirpage *dp;
478 struct lu_dirent *ent;
482 * If page is empty (end of directory is reached),
485 __u64 hash = MDS_DIR_END_OFF;
488 dp = page_address(page);
489 for (ent = lu_dirent_start(dp); ent != NULL && !done;
490 ent = lu_dirent_next(ent)) {
498 * XXX: implement correct swabbing here.
501 hash = le64_to_cpu(ent->lde_hash);
504 * Skip until we find target hash
509 namelen = le16_to_cpu(ent->lde_namelen);
520 fid_le_to_cpu(&fid, &ent->lde_fid);
521 ino = cl_fid_build_ino(&fid, api32);
522 type = ll_dirent_type_get(ent);
523 /* For 'll_nfs_get_name_filldir()', it will try
524 * to access the 'ent' through its 'lde_name',
525 * so the parameter 'name' for 'filldir()' must
526 * be part of the 'ent'. */
527 done = filldir(cookie, ent->lde_name, namelen,
530 next = le64_to_cpu(dp->ldp_hash_end);
534 if (pos == MDS_DIR_END_OFF) {
536 * End of directory reached.
539 } else if (1 /* chain is exhausted*/) {
541 * Normal case: continue to the next
544 fd->fd_dir.lfd_next = pos;
545 page = ll_get_dir_page(filp, inode, pos,
549 * go into overflow page.
557 CERROR("error reading dir "DFID" at %lu: rc %d\n",
558 PFID(&info->lli_fid), (unsigned long)pos, rc);
562 fd->fd_dir.lfd_pos = pos;
563 if (pos == MDS_DIR_END_OFF) {
565 filp->f_pos = LL_DIR_END_OFF_32BIT;
567 filp->f_pos = LL_DIR_END_OFF;
570 filp->f_pos = pos >> 32;
574 filp->f_version = inode->i_version;
575 touch_atime(filp->f_vfsmnt, filp->f_dentry);
577 ll_dir_chain_fini(&chain);
582 int ll_send_mgc_param(struct obd_export *mgc, char *string)
584 struct mgs_send_param *msp;
591 strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
592 rc = obd_set_info_async(mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
593 sizeof(struct mgs_send_param), msp, NULL);
595 CERROR("Failed to set parameter: %d\n", rc);
601 char *ll_get_fsname(struct inode *inode)
603 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
607 OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
608 len = strlen(lsi->lsi_lmd->lmd_profile);
609 ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
610 if (ptr && (strcmp(ptr, "-client") == 0))
612 strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
618 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
621 struct ll_sb_info *sbi = ll_i2sbi(inode);
622 struct md_op_data *op_data;
623 struct ptlrpc_request *req = NULL;
625 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
626 struct obd_device *mgc = lsi->lsi_mgc;
627 char *fsname = NULL, *param = NULL;
632 * This is coming from userspace, so should be in
633 * local endian. But the MDS would like it in little
634 * endian, so we swab it before we send it.
636 switch (lump->lmm_magic) {
637 case LOV_USER_MAGIC_V1: {
638 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
639 lustre_swab_lov_user_md_v1(lump);
640 lum_size = sizeof(struct lov_user_md_v1);
643 case LOV_USER_MAGIC_V3: {
644 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
645 lustre_swab_lov_user_md_v3(
646 (struct lov_user_md_v3 *)lump);
647 lum_size = sizeof(struct lov_user_md_v3);
651 CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
652 " %#08x != %#08x nor %#08x\n",
653 lump->lmm_magic, LOV_USER_MAGIC_V1,
659 lum_size = sizeof(struct lov_user_md_v1);
662 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
663 LUSTRE_OPC_ANY, NULL);
665 RETURN(PTR_ERR(op_data));
667 /* swabbing is done in lov_setstripe() on server side */
668 rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
669 NULL, 0, &req, NULL);
670 ll_finish_md_op_data(op_data);
671 ptlrpc_req_finished(req);
673 if (rc != -EPERM && rc != -EACCES)
674 CERROR("mdc_setattr fails: rc = %d\n", rc);
677 /* In the following we use the fact that LOV_USER_MAGIC_V1 and
678 LOV_USER_MAGIC_V3 have the same initial fields so we do not
679 need the make the distiction between the 2 versions */
680 if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
681 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
683 /* Get fsname and assume devname to be -MDT0000. */
684 fsname = ll_get_fsname(inode);
685 /* Set root stripesize */
686 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
687 lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
688 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
692 /* Set root stripecount */
693 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
694 lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
695 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
699 /* Set root stripeoffset */
700 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
701 lump ? le16_to_cpu(lump->lmm_stripe_offset) :
702 (typeof(lump->lmm_stripe_offset))(-1));
703 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
708 OBD_FREE(fsname, MGS_PARAM_MAXLEN);
710 OBD_FREE(param, MGS_PARAM_MAXLEN);
715 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
716 int *lmm_size, struct ptlrpc_request **request)
718 struct ll_sb_info *sbi = ll_i2sbi(inode);
719 struct mdt_body *body;
720 struct lov_mds_md *lmm = NULL;
721 struct ptlrpc_request *req = NULL;
723 struct md_op_data *op_data;
725 rc = ll_get_max_mdsize(sbi, &lmmsize);
729 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
730 0, lmmsize, LUSTRE_OPC_ANY,
735 op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
736 rc = md_getattr(sbi->ll_md_exp, op_data, &req);
737 ll_finish_md_op_data(op_data);
739 CDEBUG(D_INFO, "md_getattr failed on inode "
740 "%lu/%u: rc %d\n", inode->i_ino,
741 inode->i_generation, rc);
745 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
746 LASSERT(body != NULL);
748 lmmsize = body->eadatasize;
750 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
752 GOTO(out, rc = -ENODATA);
755 lmm = req_capsule_server_sized_get(&req->rq_pill,
756 &RMF_MDT_MD, lmmsize);
757 LASSERT(lmm != NULL);
760 * This is coming from the MDS, so is probably in
761 * little endian. We convert it to host endian before
762 * passing it to userspace.
764 /* We don't swab objects for directories */
765 switch (le32_to_cpu(lmm->lmm_magic)) {
767 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
768 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
771 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
772 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
775 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
786 * Get MDT index for the inode.
788 int ll_get_mdt_idx(struct inode *inode)
790 struct ll_sb_info *sbi = ll_i2sbi(inode);
791 struct md_op_data *op_data;
795 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
796 0, LUSTRE_OPC_ANY, NULL);
800 op_data->op_valid |= OBD_MD_MDTIDX;
801 rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
802 mdtidx = op_data->op_mds;
803 ll_finish_md_op_data(op_data);
805 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
811 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
819 if (cfs_copy_from_user(ptr, data, len)) {
823 rc = obd_iocontrol(cmd, exp, len, data, NULL);
828 static int ll_dir_ioctl(struct inode *inode, struct file *file,
829 unsigned int cmd, unsigned long arg)
831 struct ll_sb_info *sbi = ll_i2sbi(inode);
832 struct obd_ioctl_data *data;
836 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
837 inode->i_ino, inode->i_generation, inode, cmd);
839 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
840 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
843 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
845 case FSFILT_IOC_GETFLAGS:
846 case FSFILT_IOC_SETFLAGS:
847 RETURN(ll_iocontrol(inode, file, cmd, arg));
848 case FSFILT_IOC_GETVERSION_OLD:
849 case FSFILT_IOC_GETVERSION:
850 RETURN(put_user(inode->i_generation, (int *)arg));
851 /* We need to special case any other ioctls we want to handle,
852 * to send them to the MDS/OST as appropriate and to properly
853 * network encode the arg field.
854 case FSFILT_IOC_SETVERSION_OLD:
855 case FSFILT_IOC_SETVERSION:
857 case LL_IOC_GET_MDTIDX: {
860 mdtidx = ll_get_mdt_idx(inode);
864 if (put_user((int)mdtidx, (int*)arg))
869 case IOC_MDC_LOOKUP: {
870 struct ptlrpc_request *request = NULL;
871 int namelen, len = 0;
874 struct md_op_data *op_data;
876 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
881 filename = data->ioc_inlbuf1;
882 namelen = strlen(filename);
885 CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
886 GOTO(out_free, rc = -EINVAL);
889 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
890 0, LUSTRE_OPC_ANY, NULL);
892 GOTO(out_free, rc = -ENOMEM);
894 op_data->op_valid = OBD_MD_FLID;
895 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
896 ll_finish_md_op_data(op_data);
898 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
901 ptlrpc_req_finished(request);
904 obd_ioctl_freedata(buf, len);
907 case LL_IOC_LOV_SETSTRIPE: {
908 struct lov_user_md_v3 lumv3;
909 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
910 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
911 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
915 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
916 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
917 sizeof(lumv3p->lmm_objects[0]));
918 /* first try with v1 which is smaller than v3 */
919 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
922 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
923 if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
927 if (inode->i_sb->s_root == file->f_dentry)
930 /* in v1 and v3 cases lumv1 points to data */
931 rc = ll_dir_setstripe(inode, lumv1, set_default);
935 case LL_IOC_OBD_STATFS:
936 RETURN(ll_obd_statfs(inode, (void *)arg));
937 case LL_IOC_LOV_GETSTRIPE:
938 case LL_IOC_MDC_GETINFO:
939 case IOC_MDC_GETFILEINFO:
940 case IOC_MDC_GETFILESTRIPE: {
941 struct ptlrpc_request *request = NULL;
942 struct lov_user_md *lump;
943 struct lov_mds_md *lmm = NULL;
944 struct mdt_body *body;
945 char *filename = NULL;
948 if (cmd == IOC_MDC_GETFILEINFO ||
949 cmd == IOC_MDC_GETFILESTRIPE) {
950 filename = getname((const char *)arg);
951 if (IS_ERR(filename))
952 RETURN(PTR_ERR(filename));
954 rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
957 rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
961 body = req_capsule_server_get(&request->rq_pill,
963 LASSERT(body != NULL);
969 if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
970 cmd == LL_IOC_MDC_GETINFO))
971 GOTO(skip_lmm, rc = 0);
976 if (cmd == IOC_MDC_GETFILESTRIPE ||
977 cmd == LL_IOC_LOV_GETSTRIPE) {
978 lump = (struct lov_user_md *)arg;
980 struct lov_user_mds_data *lmdp;
981 lmdp = (struct lov_user_mds_data *)arg;
982 lump = &lmdp->lmd_lmm;
984 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
985 if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
986 GOTO(out_req, rc = -EFAULT);
990 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
991 struct lov_user_mds_data *lmdp;
994 st.st_dev = inode->i_sb->s_dev;
995 st.st_mode = body->mode;
996 st.st_nlink = body->nlink;
997 st.st_uid = body->uid;
998 st.st_gid = body->gid;
999 st.st_rdev = body->rdev;
1000 st.st_size = body->size;
1001 st.st_blksize = CFS_PAGE_SIZE;
1002 st.st_blocks = body->blocks;
1003 st.st_atime = body->atime;
1004 st.st_mtime = body->mtime;
1005 st.st_ctime = body->ctime;
1006 st.st_ino = inode->i_ino;
1008 lmdp = (struct lov_user_mds_data *)arg;
1009 if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1010 GOTO(out_req, rc = -EFAULT);
1015 ptlrpc_req_finished(request);
1020 case IOC_LOV_GETINFO: {
1021 struct lov_user_mds_data *lumd;
1022 struct lov_stripe_md *lsm;
1023 struct lov_user_md *lum;
1024 struct lov_mds_md *lmm;
1028 lumd = (struct lov_user_mds_data *)arg;
1029 lum = &lumd->lmd_lmm;
1031 rc = ll_get_max_mdsize(sbi, &lmmsize);
1035 OBD_ALLOC_LARGE(lmm, lmmsize);
1036 if (cfs_copy_from_user(lmm, lum, lmmsize))
1037 GOTO(free_lmm, rc = -EFAULT);
1039 switch (lmm->lmm_magic) {
1040 case LOV_USER_MAGIC_V1:
1041 if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1043 /* swab objects first so that stripes num will be sane */
1044 lustre_swab_lov_user_md_objects(
1045 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1046 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1047 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1049 case LOV_USER_MAGIC_V3:
1050 if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1052 /* swab objects first so that stripes num will be sane */
1053 lustre_swab_lov_user_md_objects(
1054 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1055 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1056 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1059 GOTO(free_lmm, rc = -EINVAL);
1062 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1064 GOTO(free_lmm, rc = -ENOMEM);
1066 /* Perform glimpse_size operation. */
1067 memset(&st, 0, sizeof(st));
1069 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1073 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1074 GOTO(free_lsm, rc = -EFAULT);
1078 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1080 OBD_FREE_LARGE(lmm, lmmsize);
1083 case OBD_IOC_LLOG_CATINFO: {
1084 struct ptlrpc_request *req = NULL;
1089 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1094 if (!data->ioc_inlbuf1) {
1095 obd_ioctl_freedata(buf, len);
1099 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
1102 GOTO(out_catinfo, rc = -ENOMEM);
1104 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
1106 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
1109 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
1111 ptlrpc_request_free(req);
1112 GOTO(out_catinfo, rc);
1115 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
1116 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
1117 if (data->ioc_inllen2) {
1118 str = req_capsule_client_get(&req->rq_pill,
1120 memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
1123 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
1125 ptlrpc_request_set_replen(req);
1127 rc = ptlrpc_queue_wait(req);
1129 str = req_capsule_server_get(&req->rq_pill,
1131 if (cfs_copy_to_user(data->ioc_pbuf1, str,
1135 ptlrpc_req_finished(req);
1137 obd_ioctl_freedata(buf, len);
1140 case OBD_IOC_QUOTACHECK: {
1141 struct obd_quotactl *oqctl;
1144 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1145 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1148 OBD_ALLOC_PTR(oqctl);
1151 oqctl->qc_type = arg;
1152 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1154 CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1158 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1160 CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1162 OBD_FREE_PTR(oqctl);
1165 case OBD_IOC_POLL_QUOTACHECK: {
1166 struct if_quotacheck *check;
1168 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1169 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1172 OBD_ALLOC_PTR(check);
1176 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1179 CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1180 if (cfs_copy_to_user((void *)arg, check,
1182 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1186 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1189 CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1190 if (cfs_copy_to_user((void *)arg, check,
1192 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1196 OBD_FREE_PTR(check);
1199 case OBD_IOC_QUOTACTL: {
1200 struct if_quotactl *qctl;
1201 int cmd, type, id, valid;
1203 OBD_ALLOC_PTR(qctl);
1207 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1208 GOTO(out_quotactl, rc = -EFAULT);
1211 type = qctl->qc_type;
1213 valid = qctl->qc_valid;
1216 case LUSTRE_Q_INVALIDATE:
1217 case LUSTRE_Q_FINVALIDATE:
1222 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1223 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1224 GOTO(out_quotactl, rc = -EPERM);
1227 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
1228 (type == GRPQUOTA && !in_egroup_p(id))) &&
1229 (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1230 sbi->ll_flags & LL_SBI_RMT_CLIENT))
1231 GOTO(out_quotactl, rc = -EPERM);
1236 CERROR("unsupported quotactl op: %#x\n", cmd);
1237 GOTO(out_quotactl, rc = -ENOTTY);
1240 if (valid != QC_GENERAL) {
1241 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
1242 GOTO(out_quotactl, rc = -EOPNOTSUPP);
1244 if (cmd == Q_GETINFO)
1245 qctl->qc_cmd = Q_GETOINFO;
1246 else if (cmd == Q_GETQUOTA)
1247 qctl->qc_cmd = Q_GETOQUOTA;
1249 GOTO(out_quotactl, rc = -EINVAL);
1253 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1255 sizeof(*qctl), qctl, NULL);
1258 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1260 sizeof(*qctl), qctl, NULL);
1263 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1265 sizeof(*qctl), qctl, NULL);
1267 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
1269 sizeof(*qctl), qctl,
1278 GOTO(out_quotactl, rc);
1282 struct obd_quotactl *oqctl;
1284 OBD_ALLOC_PTR(oqctl);
1286 GOTO(out_quotactl, rc = -ENOMEM);
1288 QCTL_COPY(oqctl, qctl);
1289 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1291 if (rc != -EALREADY && cmd == Q_QUOTAON) {
1292 oqctl->qc_cmd = Q_QUOTAOFF;
1293 obd_quotactl(sbi->ll_md_exp, oqctl);
1295 OBD_FREE_PTR(oqctl);
1296 GOTO(out_quotactl, rc);
1298 QCTL_COPY(qctl, oqctl);
1299 OBD_FREE_PTR(oqctl);
1303 if (cfs_copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1310 case OBD_IOC_GETNAME: {
1311 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1314 if (cfs_copy_to_user((void *)arg, obd->obd_name,
1315 strlen(obd->obd_name) + 1))
1319 case LL_IOC_FLUSHCTX:
1320 RETURN(ll_flush_ctx(inode));
1321 #ifdef CONFIG_FS_POSIX_ACL
1322 case LL_IOC_RMTACL: {
1323 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1324 inode == inode->i_sb->s_root->d_inode) {
1325 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1327 LASSERT(fd != NULL);
1328 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1330 fd->fd_flags |= LL_FILE_RMTACL;
1336 case LL_IOC_GETOBDCOUNT: {
1339 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1344 struct lov_obd *lov = &sbi->ll_dt_exp->exp_obd->u.lov;
1345 count = lov->desc.ld_tgt_count;
1348 struct lmv_obd *lmv = &sbi->ll_md_exp->exp_obd->u.lmv;
1349 count = lmv->desc.ld_tgt_count;
1352 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1357 case LL_IOC_PATH2FID:
1358 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1359 sizeof(struct lu_fid)))
1362 case LL_IOC_GET_CONNECT_FLAGS: {
1363 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1365 case OBD_IOC_CHANGELOG_SEND:
1366 case OBD_IOC_CHANGELOG_CLEAR:
1367 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1368 sizeof(struct ioc_changelog));
1370 case OBD_IOC_FID2PATH:
1371 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1372 case LL_IOC_HSM_CT_START:
1373 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1374 sizeof(struct lustre_kernelcomm));
1378 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1382 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1384 struct inode *inode = file->f_mapping->host;
1385 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1386 struct ll_sb_info *sbi = ll_i2sbi(inode);
1387 int api32 = ll_need_32bit_api(sbi);
1388 loff_t ret = -EINVAL;
1391 cfs_mutex_lock(&inode->i_mutex);
1396 offset += file->f_pos;
1402 offset += LL_DIR_END_OFF_32BIT;
1404 offset += LL_DIR_END_OFF;
1411 ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1412 (!api32 && offset <= LL_DIR_END_OFF))) {
1413 if (offset != file->f_pos) {
1414 if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1415 (!api32 && offset == LL_DIR_END_OFF))
1416 fd->fd_dir.lfd_pos = MDS_DIR_END_OFF;
1417 else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1418 fd->fd_dir.lfd_pos = offset << 32;
1420 fd->fd_dir.lfd_pos = offset;
1421 file->f_pos = offset;
1422 file->f_version = 0;
1429 cfs_mutex_unlock(&inode->i_mutex);
1433 int ll_dir_open(struct inode *inode, struct file *file)
1436 RETURN(ll_file_open(inode, file));
1439 int ll_dir_release(struct inode *inode, struct file *file)
1442 RETURN(ll_file_release(inode, file));
1445 struct file_operations ll_dir_operations = {
1446 .llseek = ll_dir_seek,
1447 .open = ll_dir_open,
1448 .release = ll_dir_release,
1449 .read = generic_read_dir,
1450 .readdir = ll_readdir,
1451 .ioctl = ll_dir_ioctl