X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fllite%2Fdcache.c;h=dcf56ef38dd2c618123af3ce32f41165b28be06b;hb=2dc9c16e770415d56839e1996015fec5fab93f29;hp=41c68d94408032d3b829641cd916fd92f49870c9;hpb=040033cef24c5aca2967daf2da7a862abcd074cf;p=fs%2Flustre-release.git

diff --git a/lustre/llite/dcache.c b/lustre/llite/dcache.c
index 41c68d9..dcf56ef 100644
--- a/lustre/llite/dcache.c
+++ b/lustre/llite/dcache.c
@@ -30,195 +30,434 @@
 #include <linux/lustre_lite.h>
 #include <linux/lustre_idl.h>
 #include <linux/lustre_dlm.h>
+#include <linux/lustre_version.h>
+
+#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;
+        LASSERT(de != NULL);
+        lld = ll_d2d(de);
+        LASSERT(lld != NULL);
+        LASSERT(lld->lld_cwd_count == 0);
+        LASSERT(lld->lld_mnt_count == 0);
         OBD_FREE(de->d_fsdata, sizeof(struct ll_dentry_data));
+
         EXIT;
 }
 
+/* should NOT be called with the dcache lock, see fs/dcache.c */
+static int ll_ddelete(struct dentry *de)
+{
+        ENTRY;
+        LASSERT(de);
+        CDEBUG(D_DENTRY, "%s dentry %*s (%p, parent %p, inode %p) %s%s\n",
+               (de->d_flags & DCACHE_LUSTRE_INVALID ? "keeping" : "deleting"),
+               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");
+        RETURN(0);
+}
+
 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;
-
-        LASSERT(ll_d2d(de) != NULL);
 
-        if (it->it_lock_mode) {
-                handle = (struct lustre_handle *)it->it_lock_handle;
-                ldlm_lock_decref(handle, it->it_lock_mode);
+        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->d.lustre.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;
+                /* 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->it_op == IT_RELEASED_MAGIC) {
-                EXIT;
+void ll_intent_release(struct lookup_intent *it)
+{
+        ENTRY;
+
+        ll_intent_drop_lock(it);
+        it->it_magic = 0;
+        it->it_op_release = 0;
+        it->d.lustre.it_disposition = 0;
+        it->d.lustre.it_data = NULL;
+        EXIT;
+}
+
+void ll_unhash_aliases(struct inode *inode)
+{
+        struct list_head *tmp, *head;
+        struct ll_sb_info *sbi;
+        ENTRY;
+
+        if (inode == NULL) {
+                CERROR("unexpected NULL inode, tell phil\n");
                 return;
         }
 
-        if (de->d_it == it)
-                LL_GET_INTENT(de, it);
-        else 
-                CERROR("STRANGE intent release: %p %p\n", de->d_it, it);
+        CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
+               inode->i_ino, inode->i_generation, inode);
 
+        sbi = ll_i2sbi(inode);
+        head = &inode->i_dentry;
+restart:
+        spin_lock(&dcache_lock);
+        tmp = head;
+        while ((tmp = tmp->next) != head) {
+                struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
+                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);
+                        spin_unlock(&dcache_lock);
+                        dput(dentry);
+                        goto restart;
+                } else if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
+                        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));
+                        hlist_del_init(&dentry->d_hash);
+                        dentry->d_flags |= DCACHE_LUSTRE_INVALID;
+                        hlist_add_head(&dentry->d_hash,
+                                       &sbi->ll_orphan_dentry_list);
+                }
+        }
+        spin_unlock(&dcache_lock);
         EXIT;
 }
 
 extern struct dentry *ll_find_alias(struct inode *, struct dentry *);
 
-static int revalidate2_finish(int flag, struct ptlrpc_request *request,
-                              struct dentry **de, struct lookup_intent *it,
-                              int offset, obd_id ino)
+static int revalidate_it_finish(struct ptlrpc_request *request, 
+                                int offset, struct lookup_intent *it,
+                                struct dentry *de)
 {
-        struct mds_body *body;
-        struct lov_mds_md *lmm = NULL;
-        int rc = 0; 
+        struct ll_sb_info *sbi;
+        int rc = 0;
         ENTRY;
 
-        if (!(flag & LL_LOOKUP_NEGATIVE)) {
-                body = lustre_msg_buf(request->rq_repmsg, offset);
-                if (body->valid & OBD_MD_FLEASIZE)
-                        lmm = lustre_msg_buf(request->rq_repmsg, offset + 1);
-                ll_update_inode((*de)->d_inode, body, lmm);
-                mdc_lock_set_inode((struct lustre_handle *)it->it_lock_handle,
-                                   (*de)->d_inode);
-        } else 
-                rc = -ENOENT;
-
-        ptlrpc_req_finished(request);
-        RETURN(rc);
-}
+        if (!request)
+                RETURN(0);
 
-int ll_have_md_lock(struct dentry *de)
-{
-        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;
-        ENTRY;
+        if (it_disposition(it, DISP_LOOKUP_NEG))
+                RETURN(-ENOENT);
 
-        if (!de->d_inode)
-               RETURN(0);
+        sbi = ll_i2sbi(de->d_inode);
+        rc = ll_prep_inode(sbi->ll_osc_exp, sbi->ll_mdc_exp,
+                           &de->d_inode, request, offset, NULL);
 
-        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;
+        RETURN(rc);
+}
 
-        CDEBUG(D_INFO, "trying to match res "LPU64"\n", res_id.name[0]);
+void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
+{
+        LASSERT(it != NULL);
+        LASSERT(dentry != NULL);
+
+        if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
+                struct inode *inode = dentry->d_inode;
+                CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
+                       inode, inode->i_ino, inode->i_generation);
+                mdc_set_lock_data(NULL, &it->d.lustre.it_lock_handle, inode);
+        }
 
-        if (ldlm_lock_match(obddev->obd_namespace, LDLM_FL_BLOCK_GRANTED,
-                            &res_id, LDLM_PLAIN, NULL, 0, LCK_PR, &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, LDLM_FL_BLOCK_GRANTED,
-                            &res_id, LDLM_PLAIN, NULL, 0, LCK_PW, &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 (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
+        if (it && it->it_magic != INTENT_MAGIC) {
+                CERROR("WARNING: uninitialized intent\n");
+                LBUG();
         }
-        RETURN(0);
+        if (it && (it->it_op == IT_GETATTR || it->it_op == 0))
+                it->it_op = IT_LOOKUP;
+#endif
+
+        if (!it || it->it_op == IT_GETXATTR)
+                it = *itp = deft;
+
+        it->it_op_release = ll_intent_release;
 }
 
-int ll_revalidate2(struct dentry *de, int flags, struct lookup_intent *it)
+int ll_revalidate_it(struct dentry *de, int flags, struct lookup_intent *it)
 {
         int rc;
+        struct ll_fid pfid, cfid;
+        struct it_cb_data icbd;
+        struct ll_uctxt ctxt;
+        struct ptlrpc_request *req = NULL;
+        struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
+        struct obd_export *exp;
+
         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");
+        /* Cached negative dentries are unsafe for now - look them up again */
+        if (de->d_inode == NULL)
                 RETURN(0);
-        }
 
-        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. */
-                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;
-                rc = ldlm_lock_match(obddev->obd_namespace,
-                                     LDLM_FL_BLOCK_GRANTED, &res_id,
-                                     LDLM_PLAIN, NULL, 0, LCK_PR, &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);
+        CDEBUG(D_INODE, "revalidate 0x%p: %*s -> %lu/%lu\n",
+                        de, de->d_name.len, de->d_name.name,
+                        (unsigned long) de->d_inode->i_ino,
+                        (unsigned long) de->d_inode->i_generation);
+
+        exp = ll_i2mdcexp(de->d_inode);
+        ll_inode2fid(&pfid, de->d_parent->d_inode);
+        ll_inode2fid(&cfid, de->d_inode);
+        icbd.icbd_parent = de->d_parent->d_inode;
+        icbd.icbd_childp = &de;
+
+        /* Never execute intents for mount points.
+         * Attributes will be fixed up in ll_inode_revalidate_it */
+        if (d_mountpoint(de))
+                RETURN(1);
+
+        ll_frob_intent(&it, &lookup_it);
+        LASSERT(it);
+
+        ll_i2uctxt(&ctxt, de->d_parent->d_inode, de->d_inode);
+
+        if (it->it_op == IT_GETATTR) { /* We need to check for LOOKUP lock
+                                          as well */
+                rc = md_intent_lock(exp, &ctxt, &pfid, de->d_name.name,
+                                    de->d_name.len, NULL, 0, &cfid, &lookup_it,
+                                    flags, &req, ll_mdc_blocking_ast);
+                /* If there was no lookup lock, no point in even checking for
+                   UPDATE lock */
+                if (!rc) {
+                        it = &lookup_it;
+                        GOTO(out, rc);
                 }
-                rc = ldlm_lock_match(obddev->obd_namespace,
-                                     LDLM_FL_BLOCK_GRANTED, &res_id,
-                                     LDLM_PLAIN, NULL, 0, LCK_PW, &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);
-                        }
-                        RETURN(1);
+                if (it_disposition(&lookup_it, DISP_LOOKUP_NEG)) {
+                        ll_intent_release(&lookup_it);
+                        it = &lookup_it;
+                        GOTO(out, rc = 0);
                 }
-                if (S_ISDIR(de->d_inode->i_mode))
-                        ll_invalidate_inode_pages(de->d_inode);
-                d_unhash_aliases(de->d_inode);
-                RETURN(0);
+
+                if (req)
+                        ptlrpc_req_finished(req);
+                req = NULL;
+                ll_lookup_finish_locks(&lookup_it, de);
         }
 
-        rc = ll_intent_lock(de->d_parent->d_inode, &de, it, revalidate2_finish);
+        rc = md_intent_lock(exp, &ctxt, &pfid, de->d_name.name, de->d_name.len,
+                            NULL, 0, &cfid, it, flags, &req,
+                            ll_mdc_blocking_ast);
+        /* If req is NULL, then mdc_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)
+                GOTO(out, rc);
+
         if (rc < 0) {
-                CERROR("ll_intent_lock: rc %d : it->it_status %d\n", rc,
-                       it->it_status);
-                RETURN(0);
+                if (rc != -ESTALE) {
+                        CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
+                               "%d\n", rc, it->d.lustre.it_status);
+                }
+                GOTO(out, rc = 0);
+        }
+
+        rc = revalidate_it_finish(req, 1, it, de);
+        if (rc != 0) {
+                ll_intent_release(it);
+                GOTO(out, rc = 0);
         }
+        rc = 1;
+
         /* unfortunately ll_intent_lock may cause a callback and revoke our
            dentry */
         spin_lock(&dcache_lock);
-        list_del_init(&de->d_hash);
+        hlist_del_init(&de->d_hash);
+        __d_rehash(de, 0);
         spin_unlock(&dcache_lock);
-        d_rehash(de);
 
-        RETURN(1);
+ out:
+        if (req != NULL && rc == 1)
+                ptlrpc_req_finished(req);
+        if (rc == 0) {
+                ll_unhash_aliases(de->d_inode);
+                /* done in ll_unhash_aliases()
+                de->d_flags |= DCACHE_LUSTRE_INVALID; */
+        } 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);
+                de->d_flags &= ~DCACHE_LUSTRE_INVALID;
+        }
+        RETURN(rc);
+}
+
+/*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;
+        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;
+        rc = obd_pin(sbi->ll_mdc_exp, inode->i_ino, inode->i_generation,
+                     inode->i_mode & S_IFMT, handle, flag);
+
+        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_mdc_exp, &handle, flag);
+        EXIT;
+        return;
+}
+
+#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
+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);
+}
+#endif
+
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
+static void ll_dentry_iput(struct dentry *dentry, struct inode *inode)
+{
+        struct ll_sb_info *sbi = ll_i2sbi(inode);
+        struct ll_fid parent, child;
+
+        LASSERT(dentry->d_parent && dentry->d_parent->d_inode);
+        ll_inode2fid(&parent, dentry->d_parent->d_inode);
+        ll_inode2fid(&child, inode);
+        md_change_cbdata_name(sbi->ll_mdc_exp, &parent,
+                              (char *)dentry->d_name.name, 
+                              dentry->d_name.len, &child, 
+                              null_if_equal, inode);
+        iput(inode);
 }
+#else
+#error "implement ->d_iput() for 2.6"
+#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_iput = ll_dentry_iput,
+        .d_delete = ll_ddelete,
+#if 0
+        .d_pin = ll_pin,
+        .d_unpin = ll_unpin,
+#endif
 };