X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Fdcache.c;h=4f5803b8012a9b483fb5110daedb72c74925af93;hp=0684968e64deaf6d6ddc989ff18cfb0cfba4c4f1;hb=12aa511bf0f2bd7b98cd152a3362ebacec7629ae;hpb=576c9a8212bc6607146d99e3413f7a24cbf91b5c;ds=sidebyside diff --git a/lustre/llite/dcache.c b/lustre/llite/dcache.c index 0684968..4f5803b 100644 --- a/lustre/llite/dcache.c +++ b/lustre/llite/dcache.c @@ -26,27 +26,90 @@ #define DEBUG_SUBSYSTEM S_LLITE -#include -#include -#include -#include +#include +#include +#include +#include +#include +//#include +//#include + +#include "llite_internal.h" /* should NOT be called with the dcache lock, see fs/dcache.c */ -void ll_release(struct dentry *de) +static void ll_release(struct dentry *de) { + struct ll_dentry_data *lld; ENTRY; - OBD_FREE(de->d_fsdata, sizeof(struct ll_dentry_data)); + LASSERT(de != NULL); + lld = ll_d2d(de); + if (lld == NULL) { /* NFS copies the de->d_op methods (bug 4655) */ + EXIT; + return; + } +#ifndef LUSTRE_KERNEL_VERSION + if (lld->lld_it) { + ll_intent_release(lld->lld_it); + OBD_FREE(lld->lld_it, sizeof(*lld->lld_it)); + } +#endif + LASSERT(lld->lld_cwd_count == 0); + LASSERT(lld->lld_mnt_count == 0); + OBD_FREE(de->d_fsdata, sizeof(*lld)); + EXIT; } -int ll_delete(struct dentry *de) +#ifdef LUSTRE_KERNEL_VERSION +/* Compare if two dentries are the same. Don't match if the existing dentry + * is marked DCACHE_LUSTRE_INVALID. Returns 1 if different, 0 if the same. + * + * This avoids a race where ll_lookup_it() instantiates a dentry, but we get + * an AST before calling d_revalidate_it(). The dentry still exists (marked + * INVALID) so d_lookup() matches it, but we have no lock on it (so + * lock_match() fails) and we spin around real_lookup(). */ +int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name) { - if (de->d_it != 0) { - CERROR("%s put dentry %p+%p with d_it %p\n", current->comm, - de, de->d_fsdata, de->d_it); - LBUG(); + struct dentry *dchild; + ENTRY; + + if (d_name->len != name->len) + RETURN(1); + + if (memcmp(d_name->name, name->name, name->len)) + RETURN(1); + + /* XXX: d_name must be in-dentry structure */ + dchild = container_of(d_name, struct dentry, d_name); /* ugh */ + if (dchild->d_flags & DCACHE_LUSTRE_INVALID) { + CDEBUG(D_DENTRY,"INVALID dentry %p not matched, was bug 3784\n", + dchild); + RETURN(1); } - return 0; + + RETURN(0); +} +#endif + +/* should NOT be called with the dcache lock, see fs/dcache.c */ +static int ll_ddelete(struct dentry *de) +{ + ENTRY; + LASSERT(de); +#ifndef DCACHE_LUSTRE_INVALID +#define DCACHE_LUSTRE_INVALID 0 +#endif + + CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n", + (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"), + de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode, + d_unhashed(de) ? "" : "hashed,", + list_empty(&de->d_subdirs) ? "" : "subdirs"); +#if DCACHE_LUSTRE_INVALID == 0 +#undef DCACHE_LUSTRE_INVALID +#endif + + RETURN(0); } void ll_set_dd(struct dentry *de) @@ -54,233 +117,666 @@ void ll_set_dd(struct dentry *de) ENTRY; LASSERT(de != NULL); + CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n", + de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode, + atomic_read(&de->d_count)); lock_kernel(); - if (de->d_fsdata == NULL) { OBD_ALLOC(de->d_fsdata, sizeof(struct ll_dentry_data)); - sema_init(&ll_d2d(de)->lld_it_sem, 1); } - unlock_kernel(); EXIT; } -void ll_intent_release(struct dentry *de, struct lookup_intent *it) +void ll_intent_drop_lock(struct lookup_intent *it) { struct lustre_handle *handle; - ENTRY; - if (it->it_lock_mode) { - handle = (struct lustre_handle *)it->it_lock_handle; + if (it->it_op && it->d.lustre.it_lock_mode) { + handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle; CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64 - " from it %p\n", - handle->cookie, it); - ldlm_lock_decref(handle, it->it_lock_mode); - - /* intent_release may be called multiple times, from - this thread and we don't want to double-decref this - lock (see bug 494) */ - it->it_lock_mode = 0; - } + " from it %p\n", handle->cookie, it); + ldlm_lock_decref(handle, it->d.lustre.it_lock_mode); - if (!de->d_it || it->it_op == IT_RELEASED_MAGIC) { - EXIT; - return; + /* bug 494: intent_release may be called multiple times, from + * this thread and we don't want to double-decref this lock */ + it->d.lustre.it_lock_mode = 0; } +} - if (de->d_it == it) - LL_GET_INTENT(de, it); - else - CDEBUG(D_INODE, "STRANGE intent release: %p %p\n", - de->d_it, it); +void ll_intent_release(struct lookup_intent *it) +{ + ENTRY; + CDEBUG(D_INFO, "intent %p released\n", it); + ll_intent_drop_lock(it); +#ifdef LUSTRE_KERNEL_VERSION + it->it_magic = 0; + it->it_op_release = 0; +#endif + /* We are still holding extra reference on a request, need to free it */ + if (it_disposition(it, DISP_ENQ_OPEN_REF)) /* open req for llfile_open*/ + ptlrpc_req_finished(it->d.lustre.it_data); + if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */ + ptlrpc_req_finished(it->d.lustre.it_data); + if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate + * to lookup */ + ptlrpc_req_finished(it->d.lustre.it_data); + + it->d.lustre.it_disposition = 0; + it->d.lustre.it_data = NULL; EXIT; } -extern struct dentry *ll_find_alias(struct inode *, struct dentry *); - -static int revalidate2_finish(int flag, struct ptlrpc_request *request, - struct inode *parent, struct dentry **de, - struct lookup_intent *it, int offset, obd_id ino) +/* Drop dentry if it is not used already, unhash otherwise. + Should be called with dcache lock held! + Returns: 1 if dentry was dropped, 0 if unhashed. */ +int ll_drop_dentry(struct dentry *dentry) { - struct ll_sb_info *sbi = ll_i2sbi(parent); - struct mds_body *body; - struct lov_stripe_md *lsm = NULL; - struct lov_mds_md *lmm; - int lmmsize; - int rc = 0; - ENTRY; + lock_dentry(dentry); + if (atomic_read(&dentry->d_count) == 0) { + CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p " + "inode %p\n", dentry->d_name.len, + dentry->d_name.name, dentry, dentry->d_parent, + dentry->d_inode); + dget_locked(dentry); + __d_drop(dentry); + unlock_dentry(dentry); + spin_unlock(&dcache_lock); + dput(dentry); + spin_lock(&dcache_lock); + return 1; + } + /* disconected dentry can not be find without lookup, because we + * not need his to unhash or mark invalid. */ + if (dentry->d_flags & DCACHE_DISCONNECTED) { + unlock_dentry(dentry); + RETURN (0); + } + +#ifdef LUSTRE_KERNEL_VERSION + if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) { +#else + if (!d_unhashed(dentry)) { +#endif + CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p " + "inode %p refc %d\n", dentry->d_name.len, + dentry->d_name.name, dentry, dentry->d_parent, + dentry->d_inode, atomic_read(&dentry->d_count)); + /* actually we don't unhash the dentry, rather just + * mark it inaccessible for to __d_lookup(). otherwise + * sys_getcwd() could return -ENOENT -bzzz */ +#ifdef LUSTRE_KERNEL_VERSION + dentry->d_flags |= DCACHE_LUSTRE_INVALID; +#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)) + __d_drop(dentry); + if (dentry->d_inode) { + /* Put positive dentries to orphan list */ + list_add(&dentry->d_hash, + &ll_i2sbi(dentry->d_inode)->ll_orphan_dentry_list); + } +#else + if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode)) + __d_drop(dentry); +#endif +#else + if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode)) + __d_drop(dentry); +#endif - /* NB 1 request reference will be taken away by ll_intent_lock() - * when I return */ + } + unlock_dentry(dentry); + return 0; +} - if ((flag & LL_LOOKUP_NEGATIVE) != 0) - GOTO (out, rc = -ENOENT); +void ll_unhash_aliases(struct inode *inode) +{ + struct list_head *tmp, *head; + ENTRY; - /* We only get called if the mdc_enqueue() called from - * ll_intent_lock() was successful. Therefore the mds_body is - * present and correct, and the eadata is present (but still - * opaque, so only obd_unpackmd() can check the size) */ - body = lustre_msg_buf(request->rq_repmsg, offset, sizeof (*body)); - LASSERT (body != NULL); - LASSERT_REPSWABBED (request, offset); + if (inode == NULL) { + CERROR("unexpected NULL inode, tell phil\n"); + return; + } - if (body->valid & OBD_MD_FLEASIZE) { - /* Only bother with this if inodes's LSM not set? */ + CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n", + inode->i_ino, inode->i_generation, inode); - if (body->eadatasize == 0) { - CERROR ("OBD_MD_FLEASIZE set, but eadatasize 0\n"); - GOTO (out, rc = -EPROTO); - } - lmmsize = body->eadatasize; - lmm = lustre_msg_buf (request->rq_repmsg, offset + 1, lmmsize); - LASSERT (lmm != NULL); - LASSERT_REPSWABBED (request, offset + 1); - - rc = obd_unpackmd (&sbi->ll_osc_conn, - &lsm, lmm, lmmsize); - if (rc < 0) { - CERROR ("Error %d unpacking eadata\n", rc); - LBUG(); - /* XXX don't know if I should do this... */ - GOTO (out, rc); - /* or skip the ll_update_inode but still do - * mdc_lock_set_inode() */ + head = &inode->i_dentry; + spin_lock(&dcache_lock); +restart: + tmp = head; + while ((tmp = tmp->next) != head) { + struct dentry *dentry = list_entry(tmp, struct dentry, d_alias); + + CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p " + "inode %p flags %d\n", dentry->d_name.len, + dentry->d_name.name, dentry, dentry->d_parent, + dentry->d_inode, dentry->d_flags); + + if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') { + CERROR("called on root (?) dentry=%p, inode=%p " + "ino=%lu\n", dentry, inode, inode->i_ino); + lustre_dump_dentry(dentry, 1); + libcfs_debug_dumpstack(NULL); + } else if (d_mountpoint(dentry)) { + /* For mountpoints we skip removal of the dentry + which happens solely because we have a lock on it + obtained when this dentry was not a mountpoint yet */ + CDEBUG(D_DENTRY, "Skippind mountpoint dentry removal " + "%.*s (%p) parent %p\n", + dentry->d_name.len, + dentry->d_name.name, + dentry, dentry->d_parent); + + continue; } - LASSERT (rc >= sizeof (*lsm)); - rc = 0; + + if (ll_drop_dentry(dentry)) + goto restart; } + spin_unlock(&dcache_lock); + EXIT; +} + +int ll_revalidate_it_finish(struct ptlrpc_request *request, + int offset, struct lookup_intent *it, + struct dentry *de) +{ + int rc = 0; + ENTRY; - ll_update_inode((*de)->d_inode, body, lsm); + if (!request) + RETURN(0); - if (lsm != NULL && - ll_i2info((*de)->d_inode)->lli_smd != lsm) - obd_free_memmd (&sbi->ll_osc_conn, &lsm); + if (it_disposition(it, DISP_LOOKUP_NEG)) + RETURN(-ENOENT); + + rc = ll_prep_inode(&de->d_inode, + request, offset, NULL); - ll_mdc_lock_set_inode((struct lustre_handle *)it->it_lock_handle, - (*de)->d_inode); - out: RETURN(rc); } -int ll_have_md_lock(struct dentry *de) +void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry) { - struct ll_sb_info *sbi = ll_s2sbi(de->d_sb); - struct lustre_handle lockh; - struct ldlm_res_id res_id = { .name = {0} }; - struct obd_device *obddev; - int flags; - ENTRY; + LASSERT(it != NULL); + LASSERT(dentry != NULL); - if (!de->d_inode) - RETURN(0); + if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) { + struct inode *inode = dentry->d_inode; + struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode); - obddev = class_conn2obd(&sbi->ll_mdc_conn); - res_id.name[0] = de->d_inode->i_ino; - res_id.name[1] = de->d_inode->i_generation; - - CDEBUG(D_INFO, "trying to match res "LPU64"\n", res_id.name[0]); + CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n", + inode, inode->i_ino, inode->i_generation); + md_set_lock_data(sbi->ll_md_exp, &it->d.lustre.it_lock_handle, + inode); + } - flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_MATCH_DATA; - if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN, - NULL, 0, LCK_PR, de->d_inode, &lockh)) { - ldlm_lock_decref(&lockh, LCK_PR); - RETURN(1); + /* drop lookup or getattr locks immediately */ + if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) { +#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0)) + /* on 2.6 there are situation when several lookups and + * revalidations may be requested during single operation. + * therefore, we don't release intent here -bzzz */ + ll_intent_drop_lock(it); +#else + ll_intent_release(it); +#endif } +} - if (ldlm_lock_match(obddev->obd_namespace, flags, &res_id, LDLM_PLAIN, - NULL, 0, LCK_PW, de->d_inode, &lockh)) { - ldlm_lock_decref(&lockh, LCK_PW); - RETURN(1); +void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft) +{ + struct lookup_intent *it = *itp; +#if defined(LUSTRE_KERNEL_VERSION)&&(LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0)) + if (it) { + LASSERTF(it->it_magic == INTENT_MAGIC, + "%p has bad intent magic: %x\n", + it, it->it_magic); } - RETURN(0); +#endif + + if (!it || it->it_op == IT_GETXATTR) + it = *itp = deft; + +#ifdef LUSTRE_KERNEL_VERSION + it->it_op_release = ll_intent_release; +#endif } -int ll_revalidate2(struct dentry *de, int flags, struct lookup_intent *it) +int ll_revalidate_it(struct dentry *de, int lookup_flags, + struct lookup_intent *it) { int rc; + struct md_op_data *op_data; + struct ptlrpc_request *req = NULL; + struct lookup_intent lookup_it = { .it_op = IT_LOOKUP }; + struct obd_export *exp; + struct inode *parent; + ENTRY; CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name, LL_IT2STR(it)); - /* We don't want to cache negative dentries, so return 0 immediately. - * We believe that this is safe, that negative dentries cannot be - * pinned by someone else */ if (de->d_inode == NULL) { - CDEBUG(D_INODE, "negative dentry: ret 0 to force lookup2\n"); - RETURN(0); + /* We can only use negative dentries if this is stat or lookup, + for opens and stuff we do need to query server. */ + /* If there is IT_CREAT in intent op set, then we must throw + away this negative dentry and actually do the request to + kernel to create whatever needs to be created (if possible)*/ + if (it && (it->it_op & IT_CREAT)) + RETURN(0); + +#ifdef LUSTRE_KERNEL_VERSION + if (de->d_flags & DCACHE_LUSTRE_INVALID) + RETURN(0); +#endif + + rc = ll_have_md_lock(de->d_parent->d_inode, + MDS_INODELOCK_UPDATE); + + RETURN(rc); + } + + exp = ll_i2mdexp(de->d_inode); + + /* Never execute intents for mount points. + * Attributes will be fixed up in ll_inode_revalidate_it */ + if (d_mountpoint(de)) + RETURN(1); + + /* Root of the lustre tree. Always valid. + * Attributes will be fixed up in ll_inode_revalidate_it */ + if (de == de->d_sb->s_root) + RETURN(1); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5); + ll_frob_intent(&it, &lookup_it); + LASSERT(it); + + parent = de->d_parent->d_inode; + + if (it->it_op & IT_CREAT) { + op_data = ll_prep_md_op_data(NULL, parent, NULL, + de->d_name.name, de->d_name.len, + 0, LUSTRE_OPC_CREATE, NULL); + } else { + op_data = ll_prep_md_op_data(NULL, parent, de->d_inode, + de->d_name.name, de->d_name.len, + 0, LUSTRE_OPC_ANY, NULL); } + if (IS_ERR(op_data)) + RETURN(PTR_ERR(op_data)); - if (it == NULL || it->it_op == IT_GETATTR) { - /* We could just return 1 immediately, but since we should only - * be called in revalidate2 if we already have a lock, let's - * verify that. */ + + if ((it->it_op == IT_OPEN) && de->d_inode) { struct inode *inode = de->d_inode; - struct ll_sb_info *sbi = ll_i2sbi(inode); - struct obd_device *obddev = class_conn2obd(&sbi->ll_mdc_conn); - struct ldlm_res_id res_id = - { .name = {inode->i_ino, (__u64)inode->i_generation} }; - struct lustre_handle lockh; - int flags; - flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_MATCH_DATA; - rc = ldlm_lock_match(obddev->obd_namespace, flags, &res_id, - LDLM_PLAIN, NULL, 0, LCK_PR, inode, - &lockh); - if (rc) { - de->d_flags &= ~DCACHE_LUSTRE_INVALID; - if (it && it->it_op == IT_GETATTR) { - memcpy(it->it_lock_handle, &lockh, - sizeof(lockh)); - it->it_lock_mode = LCK_PR; - LL_SAVE_INTENT(de, it); - } else { - ldlm_lock_decref(&lockh, LCK_PR); - } - RETURN(1); + struct ll_inode_info *lli = ll_i2info(inode); + struct obd_client_handle **och_p; + __u64 *och_usecount; + + /* + * We used to check for MDS_INODELOCK_OPEN here, but in fact + * just having LOOKUP lock is enough to justify inode is the + * same. And if inode is the same and we have suitable + * openhandle, then there is no point in doing another OPEN RPC + * just to throw away newly received openhandle. There are no + * security implications too, if file owner or access mode is + * change, LOOKUP lock is revoked. + */ + + + if (it->it_flags & FMODE_WRITE) { + och_p = &lli->lli_mds_write_och; + och_usecount = &lli->lli_open_fd_write_count; + } else if (it->it_flags & FMODE_EXEC) { + och_p = &lli->lli_mds_exec_och; + och_usecount = &lli->lli_open_fd_exec_count; + } else { + och_p = &lli->lli_mds_read_och; + och_usecount = &lli->lli_open_fd_read_count; } - rc = ldlm_lock_match(obddev->obd_namespace, flags, &res_id, - LDLM_PLAIN, NULL, 0, LCK_PW, inode, - &lockh); - if (rc) { - de->d_flags &= ~DCACHE_LUSTRE_INVALID; - if (it && it->it_op == IT_GETATTR) { - memcpy(it->it_lock_handle, &lockh, - sizeof(lockh)); - it->it_lock_mode = LCK_PW; - LL_SAVE_INTENT(de, it); - } else { - ldlm_lock_decref(&lockh, LCK_PW); - } + /* Check for the proper lock. */ + if (!ll_have_md_lock(inode, MDS_INODELOCK_LOOKUP)) + goto do_lock; + down(&lli->lli_och_sem); + if (*och_p) { /* Everything is open already, do nothing */ + /*(*och_usecount)++; Do not let them steal our open + handle from under us */ + /* XXX The code above was my original idea, but in case + we have the handle, but we cannot use it due to later + checks (e.g. O_CREAT|O_EXCL flags set), nobody + would decrement counter increased here. So we just + hope the lock won't be invalidated in between. But + if it would be, we'll reopen the open request to + MDS later during file open path */ + up(&lli->lli_och_sem); + ll_finish_md_op_data(op_data); RETURN(1); + } else { + up(&lli->lli_och_sem); } - if (S_ISDIR(de->d_inode->i_mode)) - ll_invalidate_inode_pages(de->d_inode); - d_unhash_aliases(de->d_inode); - RETURN(0); } - rc = ll_intent_lock(de->d_parent->d_inode, &de, it, revalidate2_finish); +do_lock: + it->it_create_mode &= ~current->fs->umask; + it->it_flags |= O_CHECK_STALE; + rc = md_intent_lock(exp, op_data, NULL, 0, it, + lookup_flags, + &req, ll_md_blocking_ast, 0); + it->it_flags &= ~O_CHECK_STALE; + ll_finish_md_op_data(op_data); + /* If req is NULL, then md_intent_lock only tried to do a lock match; + * if all was well, it will return 1 if it found locks, 0 otherwise. */ + if (req == NULL && rc >= 0) { + if (!rc) + goto do_lookup; + GOTO(out, rc); + } + if (rc < 0) { if (rc != -ESTALE) { - CERROR("ll_intent_lock: rc %d : it->it_status %d\n", rc, - it->it_status); + CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status " + "%d\n", rc, it->d.lustre.it_status); } - RETURN(0); + GOTO(out, rc = 0); } + +revalidate_finish: + rc = ll_revalidate_it_finish(req, DLM_REPLY_REC_OFF, it, de); + if (rc != 0) { + if (rc != -ESTALE && rc != -ENOENT) + ll_intent_release(it); + GOTO(out, rc = 0); + } + + if ((it->it_op & IT_OPEN) && de->d_inode && + !S_ISREG(de->d_inode->i_mode) && + !S_ISDIR(de->d_inode->i_mode)) { + ll_release_openhandle(de, it); + } + rc = 1; + /* unfortunately ll_intent_lock may cause a callback and revoke our - dentry */ + * dentry */ spin_lock(&dcache_lock); - list_del_init(&de->d_hash); + lock_dentry(de); + __d_drop(de); + unlock_dentry(de); __d_rehash(de, 0); spin_unlock(&dcache_lock); - RETURN(1); +out: + /* We do not free request as it may be reused during following lookup + * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will + * be freed in ll_lookup_it or in ll_intent_release. But if + * request was not completed, we need to free it. (bug 5154, 9903) */ + if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE)) + ptlrpc_req_finished(req); + if (rc == 0) { +#ifdef LUSTRE_KERNEL_VERSION + ll_unhash_aliases(de->d_inode); + /* done in ll_unhash_aliases() + dentry->d_flags |= DCACHE_LUSTRE_INVALID; */ +#else + /* We do not want d_invalidate to kill all child dentries too */ + d_drop(de); +#endif + } else { + CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p " + "inode %p refc %d\n", de->d_name.len, + de->d_name.name, de, de->d_parent, de->d_inode, + atomic_read(&de->d_count)); + ll_lookup_finish_locks(it, de); +#ifdef LUSTRE_KERNEL_VERSION + lock_dentry(de); + de->d_flags &= ~DCACHE_LUSTRE_INVALID; + unlock_dentry(de); +#endif + } + RETURN(rc); + + /* + * This part is here to combat evil-evil race in real_lookup on 2.6 + * kernels. The race details are: We enter do_lookup() looking for some + * name, there is nothing in dcache for this name yet and d_lookup() + * returns NULL. We proceed to real_lookup(), and while we do this, + * another process does open on the same file we looking up (most simple + * reproducer), open succeeds and the dentry is added. Now back to + * us. In real_lookup() we do d_lookup() again and suddenly find the + * dentry, so we call d_revalidate on it, but there is no lock, so + * without this code we would return 0, but unpatched real_lookup just + * returns -ENOENT in such a case instead of retrying the lookup. Once + * this is dealt with in real_lookup(), all of this ugly mess can go and + * we can just check locks in ->d_revalidate without doing any RPCs + * ever. + */ +do_lookup: + if (it != &lookup_it) { + /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */ + if (it->it_op == IT_GETATTR) + lookup_it.it_op = IT_GETATTR; + ll_lookup_finish_locks(it, de); + it = &lookup_it; + } + + /* Do real lookup here. */ + op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name, + de->d_name.len, 0, (it->it_op & IT_CREAT ? + LUSTRE_OPC_CREATE : + LUSTRE_OPC_ANY), NULL); + if (IS_ERR(op_data)) + RETURN(PTR_ERR(op_data)); + + rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req, + ll_md_blocking_ast, 0); + if (rc >= 0) { + struct mdt_body *mdt_body = lustre_msg_buf(req->rq_repmsg, + DLM_REPLY_REC_OFF, + sizeof(*mdt_body)); + struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0}; + + if (de->d_inode) + fid = *ll_inode2fid(de->d_inode); + + /* see if we got same inode, if not - return error */ + if (lu_fid_eq(&fid, &mdt_body->fid1)) { + ll_finish_md_op_data(op_data); + op_data = NULL; + goto revalidate_finish; + } + ll_intent_release(it); + } + ll_finish_md_op_data(op_data); + GOTO(out, rc = 0); +} + +/*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag) +{ + struct inode *inode= de->d_inode; + struct ll_sb_info *sbi = ll_i2sbi(inode); + struct ll_dentry_data *ldd = ll_d2d(de); + struct obd_client_handle *handle; + struct obd_capa *oc; + int rc = 0; + ENTRY; + LASSERT(ldd); + + lock_kernel(); + /* Strictly speaking this introduces an additional race: the + * increments should wait until the rpc has returned. + * However, given that at present the function is void, this + * issue is moot. */ + if (flag == 1 && (++ldd->lld_mnt_count) > 1) { + unlock_kernel(); + EXIT; + return; + } + + if (flag == 0 && (++ldd->lld_cwd_count) > 1) { + unlock_kernel(); + EXIT; + return; + } + unlock_kernel(); + + handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och; + oc = ll_mdscapa_get(inode); + rc = obd_pin(sbi->ll_md_exp, ll_inode2fid(inode), oc, handle, flag); + capa_put(oc); + if (rc) { + lock_kernel(); + memset(handle, 0, sizeof(*handle)); + if (flag == 0) + ldd->lld_cwd_count--; + else + ldd->lld_mnt_count--; + unlock_kernel(); + } + + EXIT; + return; +} + +/*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag) +{ + struct ll_sb_info *sbi = ll_i2sbi(de->d_inode); + struct ll_dentry_data *ldd = ll_d2d(de); + struct obd_client_handle handle; + int count, rc = 0; + ENTRY; + LASSERT(ldd); + + lock_kernel(); + /* Strictly speaking this introduces an additional race: the + * increments should wait until the rpc has returned. + * However, given that at present the function is void, this + * issue is moot. */ + handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och; + if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) { + /* the "pin" failed */ + unlock_kernel(); + EXIT; + return; + } + + if (flag) + count = --ldd->lld_mnt_count; + else + count = --ldd->lld_cwd_count; + unlock_kernel(); + + if (count != 0) { + EXIT; + return; + } + + rc = obd_unpin(sbi->ll_md_exp, &handle, flag); + EXIT; + return; +} + +#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0)) +#ifdef LUSTRE_KERNEL_VERSION +static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd) +{ + int rc; + ENTRY; + + if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST)) + rc = ll_revalidate_it(dentry, nd->flags, &nd->intent); + else + rc = ll_revalidate_it(dentry, 0, NULL); + + RETURN(rc); +} +#else +int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd) +{ + int rc; + ENTRY; + + if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) { + struct lookup_intent *it; + it = ll_convert_intent(&nd->intent.open, nd->flags); + if (IS_ERR(it)) + RETURN(0); + if (it->it_op == (IT_OPEN|IT_CREAT)) + if (nd->intent.open.flags & O_EXCL) { + CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n"); + rc = 0; + goto out_it; + } + + rc = ll_revalidate_it(dentry, nd->flags, it); + + if (rc && (nd->flags & LOOKUP_OPEN) && + it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/ +#ifdef HAVE_FILE_IN_STRUCT_INTENT +// XXX Code duplication with ll_lookup_nd + if (S_ISFIFO(dentry->d_inode->i_mode)) { + // We cannot call open here as it would + // deadlock. + ptlrpc_req_finished( + (struct ptlrpc_request *) + it->d.lustre.it_data); + } else { + struct file *filp; + + nd->intent.open.file->private_data = it; + filp = lookup_instantiate_filp(nd, dentry,NULL); +#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,17)) +/* 2.6.1[456] have a bug in open_namei() that forgets to check + * nd->intent.open.file for error, so we need to return it as lookup's result + * instead */ + if (IS_ERR(filp)) + rc = 0; +#endif + } +#else + ll_release_openhandle(dentry, it); +#endif /* HAVE_FILE_IN_STRUCT_INTENT */ + } + if (!rc && (nd->flags & LOOKUP_CREATE) && + it_disposition(it, DISP_OPEN_CREATE)) { + /* We created something but we may only return + * negative dentry here, so save request in dentry, + * if lookup will be called later on, it will + * pick the request, otherwise it would be freed + * with dentry */ + ll_d2d(dentry)->lld_it = it; + it = NULL; /* avoid freeing */ + } + +out_it: + if (it) { + ll_intent_release(it); + OBD_FREE(it, sizeof(*it)); + } + } else { + rc = ll_revalidate_it(dentry, 0, NULL); + } + + RETURN(rc); } +#endif +#endif struct dentry_operations ll_d_ops = { - .d_revalidate2 = ll_revalidate2, - .d_intent_release = ll_intent_release, +#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0)) + .d_revalidate = ll_revalidate_nd, +#else + .d_revalidate_it = ll_revalidate_it, +#endif .d_release = ll_release, - .d_delete = ll_delete, + .d_delete = ll_ddelete, +#ifdef LUSTRE_KERNEL_VERSION + .d_compare = ll_dcompare, +#endif +#if 0 + .d_pin = ll_pin, + .d_unpin = ll_unpin, +#endif };