1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 1992, 1993, 1994, 1995
5 * Remy Card (card@masi.ibp.fr)
6 * Laboratoire MASI - Institut Blaise Pascal
7 * Universite Pierre et Marie Curie (Paris VI)
11 * linux/fs/minix/dir.c
14 * Copyright (C) 1991, 1992 Linus Torvalds
16 * ext2 directory handling functions
18 * Big-endian to little-endian byte-swapping/bitmaps by
19 * David S. Miller (davem@caip.rutgers.edu), 1995
21 * All code that works with directory layout had been switched to pagecache
24 * Adapted for Lustre Light
25 * Copyright (C) 2002-2003, Cluster File Systems, Inc.
30 #include <linux/ext2_fs.h>
31 #include <linux/pagemap.h>
33 #include <linux/version.h>
34 #include <linux/smp_lock.h>
35 #include <asm/uaccess.h>
36 #include <linux/file.h>
37 #include <linux/kmod.h>
38 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
39 # include <linux/locks.h> // for wait_on_buffer
41 # include <linux/buffer_head.h> // for wait_on_buffer
44 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <linux/obd_support.h>
47 #include <linux/obd_class.h>
48 #include <linux/lustre_lib.h>
49 #include <linux/lustre_idl.h>
50 #include <linux/lustre_mds.h>
51 #include <linux/lustre_lite.h>
52 #include <linux/lustre_dlm.h>
53 #include "llite_internal.h"
55 typedef struct ext2_dir_entry_2 ext2_dirent;
57 #define PageChecked(page) test_bit(PG_checked, &(page)->flags)
58 #define SetPageChecked(page) set_bit(PG_checked, &(page)->flags)
60 /* returns the page unlocked, but with a reference */
61 static int ll_dir_readpage(struct file *file, struct page *page)
63 struct inode *inode = page->mapping->host;
64 struct ptlrpc_request *request;
65 struct mds_body *body;
71 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
72 inode->i_generation, inode);
74 mdc_pack_id(&id, inode->i_ino, inode->i_generation,
75 S_IFDIR, id_group(&ll_i2info(inode)->lli_id),
76 id_fid(&ll_i2info(inode)->lli_id));
78 offset = page->index << PAGE_SHIFT;
79 rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, &id, offset,
82 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
83 LASSERT (body != NULL); /* checked by md_readpage() */
84 LASSERT_REPSWABBED (request, 0); /* swabbed by md_readpage() */
86 #warning "FIXME ASAP!"
87 //inode->i_size = body->size;
88 SetPageUptodate(page);
90 ptlrpc_req_finished(request);
96 struct address_space_operations ll_dir_aops = {
97 .readpage = ll_dir_readpage,
101 * ext2 uses block-sized chunks. Arguably, sector-sized ones would be
102 * more robust, but we have what we have
104 static inline unsigned ext2_chunk_size(struct inode *inode)
106 return inode->i_sb->s_blocksize;
109 static inline void ext2_put_page(struct page *page)
112 page_cache_release(page);
115 static inline unsigned long dir_pages(struct inode *inode)
117 return (inode->i_size+PAGE_CACHE_SIZE-1)>>PAGE_CACHE_SHIFT;
120 static void ext2_check_page(struct page *page)
122 struct inode *dir = page->mapping->host;
123 unsigned chunk_size = ext2_chunk_size(dir);
124 char *kaddr = page_address(page);
125 // u32 max_inumber = le32_to_cpu(sb->u.ext2_sb.s_es->s_inodes_count);
126 unsigned offs, rec_len;
127 unsigned limit = PAGE_CACHE_SIZE;
131 if ((dir->i_size >> PAGE_CACHE_SHIFT) == page->index) {
132 limit = dir->i_size & ~PAGE_CACHE_MASK;
133 if (limit & (chunk_size - 1)) {
134 CERROR("limit %d dir size %lld index %ld\n",
135 limit, dir->i_size, page->index);
138 for (offs = limit; offs<PAGE_CACHE_SIZE; offs += chunk_size) {
139 ext2_dirent *p = (ext2_dirent*)(kaddr + offs);
140 p->rec_len = cpu_to_le16(chunk_size);
147 for (offs = 0; offs <= limit - EXT2_DIR_REC_LEN(1); offs += rec_len) {
148 p = (ext2_dirent *)(kaddr + offs);
149 rec_len = le16_to_cpu(p->rec_len);
151 if (rec_len < EXT2_DIR_REC_LEN(1))
155 if (rec_len < EXT2_DIR_REC_LEN(p->name_len))
157 if (((offs + rec_len - 1) ^ offs) & ~(chunk_size-1))
159 // if (le32_to_cpu(p->inode) > max_inumber)
165 SetPageChecked(page);
168 /* Too bad, we had an error */
171 CERROR("ext2_check_page"
172 "size of directory #%lu is not a multiple of chunk size\n",
177 error = "rec_len is smaller than minimal";
180 error = "unaligned directory entry";
183 error = "rec_len is too small for name_len";
186 error = "directory entry across blocks";
189 // error = "inode out of bounds";
191 CERROR("ext2_check_page: bad entry in directory #%lu: %s - "
192 "offset=%lu+%u, inode=%lu, rec_len=%d, name_len=%d",
193 dir->i_ino, error, (page->index<<PAGE_CACHE_SHIFT), offs,
194 (unsigned long) le32_to_cpu(p->inode),
195 rec_len, p->name_len);
198 p = (ext2_dirent *)(kaddr + offs);
199 CERROR("ext2_check_page"
200 "entry in directory #%lu spans the page boundary"
201 "offset=%lu, inode=%lu",
202 dir->i_ino, (page->index<<PAGE_CACHE_SHIFT)+offs,
203 (unsigned long) le32_to_cpu(p->inode));
205 SetPageChecked(page);
209 static struct page *ll_get_dir_page(struct inode *dir, unsigned long n)
211 struct ll_inode_info *li = ll_i2info(dir);
212 struct ldlm_res_id res_id =
213 { .name = { id_fid(&li->lli_id), id_group(&li->lli_id)} };
214 struct lustre_handle lockh;
215 struct obd_device *obddev = class_exp2obd(ll_i2sbi(dir)->ll_md_exp);
216 struct address_space *mapping = dir->i_mapping;
218 ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_UPDATE } };
221 obddev = md_get_real_obd(ll_i2sbi(dir)->ll_md_exp, &li->lli_id);
222 rc = ldlm_lock_match(obddev->obd_namespace, LDLM_FL_BLOCK_GRANTED,
223 &res_id, LDLM_IBITS, &policy, LCK_PR, &lockh);
225 struct lookup_intent it = { .it_op = IT_READDIR };
226 struct ptlrpc_request *request;
227 struct mdc_op_data *op_data;
229 OBD_ALLOC(op_data, sizeof(*op_data));
231 return ERR_PTR(-ENOMEM);
233 ll_prepare_mdc_data(op_data, dir, NULL, NULL, 0, 0);
235 rc = ll_intent_alloc(&it);
239 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, LDLM_IBITS, &it,
240 LCK_PR, op_data, &lockh, NULL, 0,
241 ldlm_completion_ast, ll_mdc_blocking_ast, dir);
242 OBD_FREE(op_data, sizeof(*op_data));
244 request = (struct ptlrpc_request *)LUSTRE_IT(&it)->it_data;
248 ptlrpc_req_finished(request);
250 CERROR("lock enqueue: rc: %d\n", rc);
254 ldlm_lock_dump_handle(D_OTHER, &lockh);
256 page = read_cache_page(mapping, n,
257 (filler_t*)mapping->a_ops->readpage, NULL);
261 if (!PageUptodate(page))
263 if (!PageChecked(page))
264 ext2_check_page(page);
270 ldlm_lock_decref(&lockh, LCK_PR);
275 page = ERR_PTR(-EIO);
280 * p is at least 6 bytes before the end of page
282 static inline ext2_dirent *ext2_next_entry(ext2_dirent *p)
284 return (ext2_dirent *)((char*)p + le16_to_cpu(p->rec_len));
287 static inline unsigned
288 ext2_validate_entry(char *base, unsigned offset, unsigned mask)
290 ext2_dirent *de = (ext2_dirent*)(base + offset);
291 ext2_dirent *p = (ext2_dirent*)(base + (offset&mask));
292 while ((char*)p < (char*)de)
293 p = ext2_next_entry(p);
294 return (char *)p - base;
297 static unsigned char ext2_filetype_table[EXT2_FT_MAX] = {
298 [EXT2_FT_UNKNOWN] DT_UNKNOWN,
299 [EXT2_FT_REG_FILE] DT_REG,
300 [EXT2_FT_DIR] DT_DIR,
301 [EXT2_FT_CHRDEV] DT_CHR,
302 [EXT2_FT_BLKDEV] DT_BLK,
303 [EXT2_FT_FIFO] DT_FIFO,
304 [EXT2_FT_SOCK] DT_SOCK,
305 [EXT2_FT_SYMLINK] DT_LNK,
309 int ll_readdir(struct file * filp, void * dirent, filldir_t filldir)
311 struct inode *inode = filp->f_dentry->d_inode;
312 loff_t pos = filp->f_pos;
313 // XXX struct super_block *sb = inode->i_sb;
314 unsigned offset = pos & ~PAGE_CACHE_MASK;
315 unsigned long n = pos >> PAGE_CACHE_SHIFT;
316 unsigned long npages = dir_pages(inode);
317 unsigned chunk_mask = ~(ext2_chunk_size(inode)-1);
318 unsigned char *types = ext2_filetype_table;
319 int need_revalidate = (filp->f_version != inode->i_version);
323 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %llu/%llu\n",
324 inode->i_ino, inode->i_generation, inode, pos, inode->i_size);
326 if (pos > inode->i_size - EXT2_DIR_REC_LEN(1))
329 for ( ; n < npages; n++, offset = 0) {
334 CDEBUG(D_EXT2,"read %lu of dir %lu/%u page %lu/%lu size %llu\n",
335 PAGE_CACHE_SIZE, inode->i_ino, inode->i_generation,
336 n, npages, inode->i_size);
337 page = ll_get_dir_page(inode, n);
339 /* size might have been updated by md_readpage() */
340 npages = dir_pages(inode);
344 CERROR("error reading dir %lu/%u page %lu: rc %d\n",
345 inode->i_ino, inode->i_generation, n, rc);
349 kaddr = page_address(page);
350 if (need_revalidate) {
351 offset = ext2_validate_entry(kaddr, offset, chunk_mask);
354 de = (ext2_dirent *)(kaddr+offset);
355 limit = kaddr + PAGE_CACHE_SIZE - EXT2_DIR_REC_LEN(1);
356 for ( ;(char*)de <= limit; de = ext2_next_entry(de)) {
360 rc = 0; /* no error if we return something */
362 offset = (char *)de - kaddr;
363 over = filldir(dirent, de->name, de->name_len,
364 (n<<PAGE_CACHE_SHIFT) | offset,
365 le32_to_cpu(de->inode),
366 types[de->file_type &
380 filp->f_pos = (n << PAGE_CACHE_SHIFT) | offset;
381 filp->f_version = inode->i_version;
386 static int ll_mkdir_stripe(struct inode *inode, unsigned long arg)
388 struct ptlrpc_request *request = NULL;
389 struct ll_sb_info *sbi = ll_i2sbi(inode);
390 struct ll_user_mkdir_stripe lums;
391 struct mdc_op_data *op_data;
398 if (copy_from_user(&lums, (void *)arg, sizeof(lums)))
401 if (lums.lums_namelen <= 0)
403 OBD_ALLOC(name, lums.lums_namelen);
407 if (copy_from_user(name, lums.lums_name, lums.lums_namelen))
408 GOTO(out, err = -EFAULT);
410 CDEBUG(D_VFSTRACE, "ioctl Op:name=%s,dir=%lu/%u(%p)\n",
411 name, inode->i_ino, inode->i_generation, inode);
412 nstripes = lums.lums_nstripes;
414 mode = lums.lums_mode;
415 mode = (mode & (S_IRWXUGO|S_ISVTX) & ~current->fs->umask) | S_IFDIR;
417 OBD_ALLOC(op_data, sizeof(*op_data));
419 GOTO(out, err = -ENOMEM);
420 ll_prepare_mdc_data(op_data, inode, NULL, name,lums.lums_namelen,0);
421 err = md_create(sbi->ll_md_exp, op_data, &nstripes, sizeof(nstripes),
422 mode, current->fsuid, current->fsgid, 0, &request);
423 OBD_FREE(op_data, sizeof(*op_data));
424 ptlrpc_req_finished(request);
427 OBD_FREE(name, lums.lums_namelen);
431 static int ll_dir_ioctl(struct inode *inode, struct file *file,
432 unsigned int cmd, unsigned long arg)
434 struct ll_sb_info *sbi = ll_i2sbi(inode);
435 struct obd_ioctl_data *data;
438 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
439 inode->i_ino, inode->i_generation, inode, cmd);
441 if (_IOC_TYPE(cmd) == 'T') /* tty ioctls */
444 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
446 case EXT3_IOC_GETFLAGS:
447 case EXT3_IOC_SETFLAGS:
448 RETURN(ll_iocontrol(inode, file, cmd, arg));
449 case IOC_MDC_LOOKUP: {
450 struct ptlrpc_request *request = NULL;
454 int namelen, rc, len = 0;
457 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
462 filename = data->ioc_inlbuf1;
463 namelen = data->ioc_inllen1;
466 CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
467 GOTO(out, rc = -EINVAL);
471 ll_inode2id(&id, inode);
472 rc = md_getattr_lock(sbi->ll_md_exp, &id,
473 filename, namelen, valid, 0, &request);
475 CDEBUG(D_INFO, "md_getattr_lock: %d\n", rc);
479 ptlrpc_req_finished(request);
483 obd_ioctl_freedata(buf, len);
486 case LL_IOC_MDC_MKDIRSTRIPE:
487 RETURN(ll_mkdir_stripe(inode, arg));
488 case LL_IOC_LOV_SETSTRIPE: {
489 struct ptlrpc_request *request = NULL;
490 struct mdc_op_data *op_data;
491 struct iattr attr = { 0 };
492 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
495 LASSERT(sizeof(lum) == sizeof(*lump));
496 LASSERT(sizeof(lum.lmm_objects[0]) ==
497 sizeof(lump->lmm_objects[0]));
498 rc = copy_from_user(&lum, lump, sizeof(lum));
502 if (lum.lmm_magic != LOV_USER_MAGIC)
505 OBD_ALLOC(op_data, sizeof(*op_data));
509 ll_prepare_mdc_data(op_data, inode, NULL, NULL, 0, 0);
511 rc = md_setattr(sbi->ll_md_exp, op_data, &attr, &lum,
512 sizeof(lum), NULL, 0, &request);
513 OBD_FREE(op_data, sizeof(*op_data));
514 ptlrpc_req_finished(request);
517 if (rc != -EPERM && rc != -EACCES)
518 CERROR("md_setattr fails: rc = %d\n", rc);
522 case LL_IOC_LOV_GETSTRIPE: {
523 struct ptlrpc_request *request = NULL;
524 struct lov_user_md *lump = (struct lov_user_md *)arg;
525 struct lov_mds_md *lmm;
527 struct mds_body *body;
531 valid |= OBD_MD_FLDIREA;
533 ll_inode2id(&id, inode);
534 rc = md_getattr(sbi->ll_md_exp, &id, valid, NULL, 0,
535 obd_size_diskmd(sbi->ll_dt_exp, NULL),
538 CDEBUG(D_INFO, "md_getattr failed: rc = %d\n", rc);
542 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof(*body));
543 LASSERT(body != NULL); /* checked by md_getattr_lock */
544 LASSERT_REPSWABBED(request, 0);/* swabbed by md_getattr_lock */
546 lmmsize = body->eadatasize;
548 GOTO(out_get, rc = -ENODATA);
550 lmm = lustre_msg_buf(request->rq_repmsg, 1, lmmsize);
551 LASSERT(lmm != NULL);
552 LASSERT_REPSWABBED(request, 1);
553 rc = copy_to_user(lump, lmm, lmmsize);
555 GOTO(out_get, rc = -EFAULT);
559 ptlrpc_req_finished(request);
562 case IOC_MDC_SHOWFID: {
563 struct lustre_id *idp = (struct lustre_id *)arg;
568 filename = getname((const char *)arg);
569 if (IS_ERR(filename))
570 RETURN(PTR_ERR(filename));
572 ll_inode2id(&id, inode);
574 rc = ll_get_fid(sbi->ll_md_exp, &id, filename, &id);
576 GOTO(out_filename, rc);
578 rc = copy_to_user(idp, &id, sizeof(*idp));
580 GOTO(out_filename, rc = -EFAULT);
587 case IOC_MDC_GETSTRIPE: {
588 struct lov_user_md *lump = (struct lov_user_md *)arg;
589 struct ptlrpc_request *request = NULL;
590 struct lov_mds_md *lmm;
591 struct mds_body *body;
596 filename = getname((const char *)arg);
597 if (IS_ERR(filename))
598 RETURN(PTR_ERR(filename));
600 ll_inode2id(&id, inode);
601 rc = md_getattr_lock(sbi->ll_md_exp, &id, filename,
602 strlen(filename) + 1, OBD_MD_FLEASIZE,
603 obd_size_diskmd(sbi->ll_dt_exp, NULL),
606 CDEBUG(D_INFO, "md_getattr_lock failed on %s: rc %d\n",
611 body = lustre_msg_buf(request->rq_repmsg, 0, sizeof (*body));
612 LASSERT(body != NULL); /* checked by md_getattr_lock */
613 LASSERT_REPSWABBED(request, 0);/* swabbed by md_getattr_lock */
615 lmmsize = body->eadatasize;
617 if (!(body->valid & OBD_MD_FLEASIZE) || lmmsize == 0)
618 GOTO(out_req, rc = -ENODATA);
621 GOTO(out_req, rc = -EFBIG);
623 lmm = lustre_msg_buf(request->rq_repmsg, 1, lmmsize);
624 LASSERT(lmm != NULL);
625 LASSERT_REPSWABBED(request, 1);
627 rc = copy_to_user(lump, lmm, lmmsize);
629 GOTO(out_req, rc = -EFAULT);
633 ptlrpc_req_finished(request);
639 struct ptlrpc_request *req = NULL;
642 struct client_obd *cli;
643 struct obd_device *obd;
645 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
650 obd = class_name2obd(data->ioc_inlbuf1);
653 GOTO(out_ping, rc = -ENODEV);
655 if (!obd->obd_attached) {
656 CERROR("Device %d not attached\n", obd->obd_minor);
657 GOTO(out_ping, rc = -ENODEV);
659 if (!obd->obd_set_up) {
660 CERROR("Device %d still not setup\n", obd->obd_minor);
661 GOTO(out_ping, rc = -ENODEV);
664 req = ptlrpc_prep_req(cli->cl_import, LUSTRE_OBD_VERSION,
665 OBD_PING, 0, NULL, NULL);
667 GOTO(out_ping, rc = -ENOMEM);
669 req->rq_replen = lustre_msg_size(0, NULL);
670 req->rq_send_state = LUSTRE_IMP_FULL;
672 rc = ptlrpc_queue_wait(req);
674 ptlrpc_req_finished(req);
676 obd_ioctl_freedata(buf, len);
679 case OBD_IOC_LLOG_CATINFO: {
680 struct ptlrpc_request *req = NULL;
686 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
691 if (!data->ioc_inlbuf1) {
692 obd_ioctl_freedata(buf, len);
696 lens[0] = data->ioc_inllen1;
697 bufs[0] = data->ioc_inlbuf1;
698 if (data->ioc_inllen2) {
699 lens[1] = data->ioc_inllen2;
700 bufs[1] = data->ioc_inlbuf2;
705 size = data->ioc_plen1;
706 req = ptlrpc_prep_req(sbi2md(sbi)->cl_import,
707 LUSTRE_LOG_VERSION, LLOG_CATINFO,
710 GOTO(out_catinfo, rc = -ENOMEM);
712 req->rq_replen = lustre_msg_size(1, &size);
714 rc = ptlrpc_queue_wait(req);
715 str = lustre_msg_string(req->rq_repmsg, 0, data->ioc_plen1);
717 rc = copy_to_user(data->ioc_pbuf1, str,
719 ptlrpc_req_finished(req);
721 obd_ioctl_freedata(buf, len);
725 return obd_iocontrol(cmd, sbi->ll_dt_exp, 0,
730 int ll_dir_open(struct inode *inode, struct file *file)
732 return ll_file_open(inode, file);
735 int ll_dir_release(struct inode *inode, struct file *file)
737 return ll_file_release(inode, file);
740 struct file_operations ll_dir_operations = {
742 .release = ll_dir_release,
743 .read = generic_read_dir,
744 .readdir = ll_readdir,
745 .ioctl = ll_dir_ioctl