#endif
#define DEBUG_SUBSYSTEM S_MDS
+#include <linux/version.h>
#include <linux/module.h>
#include <linux/init.h>
+#include <linux/version.h>
#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
# include <linux/buffer_head.h>
# include <linux/workqueue.h>
* mds_mfd_close - for force closing files when a client dies
*/
-/*
+/*
* MDS file data handling: file data holds a handle for a file opened
* by a client.
*/
}
-/*
- * Write access to a file: executors cause a negative count,
- * writers a positive count. The spin lock is needed to perform
- * a check for the sign and then increment or decrement atomically.
- */
+/* Caller must hold mds->mds_epoch_sem */
+static int mds_alloc_filterdata(struct inode *inode)
+{
+ LASSERT(inode->i_filterdata == NULL);
+ OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
+ if (inode->i_filterdata == NULL)
+ return -ENOMEM;
+ LASSERT(igrab(inode) == inode);
+ return 0;
+}
-static spinlock_t mds_exec_lock = SPIN_LOCK_UNLOCKED;
-static int mds_get_write_access(struct inode * inode)
+/* Caller must hold mds->mds_epoch_sem */
+static void mds_free_filterdata(struct inode *inode)
{
- ENTRY;
- spin_lock(&mds_exec_lock);
+ LASSERT(inode->i_filterdata != NULL);
+ OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
+ inode->i_filterdata = NULL;
+ iput(inode);
+}
+
+/* Write access to a file: executors cause a negative count,
+ * writers a positive count. The semaphore is needed to perform
+ * a check for the sign and then increment or decrement atomically.
+ *
+ * This code is closely tied to the allocation of the d_fsdata and the
+ * MDS epoch, so we use the same semaphore for the whole lot.
+ *
+ * We could use a different semaphore for each file, if it ever shows
+ * up in a profile, which it won't.
+ *
+ * epoch argument is nonzero during recovery */
+static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
+ __u64 epoch)
+{
+ int rc = 0;
+
+ down(&mds->mds_epoch_sem);
+
if (atomic_read(&inode->i_writecount) < 0) {
- spin_unlock(&mds_exec_lock);
+ up(&mds->mds_epoch_sem);
RETURN(-ETXTBSY);
}
- atomic_inc(&inode->i_writecount);
- spin_unlock(&mds_exec_lock);
- RETURN(0);
+
+
+ if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
+ CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
+ MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
+ inode->i_generation);
+ goto out;
+ }
+
+ if (inode->i_filterdata == NULL)
+ mds_alloc_filterdata(inode);
+ if (inode->i_filterdata == NULL) {
+ rc = -ENOMEM;
+ goto out;
+ }
+ if (epoch > mds->mds_io_epoch)
+ mds->mds_io_epoch = epoch;
+ else
+ mds->mds_io_epoch++;
+ MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
+ CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
+ mds->mds_io_epoch, inode->i_ino, inode->i_generation);
+ out:
+ if (rc == 0)
+ atomic_inc(&inode->i_writecount);
+ up(&mds->mds_epoch_sem);
+ return rc;
}
-static int mds_deny_write_access(struct inode *inode)
+/* Returns EAGAIN if the client needs to get size and/or cookies and close
+ * again -- which is never true if the file is about to be unlinked. Otherwise
+ * returns the number of remaining writers. */
+static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
+ struct mds_body *body, int unlinking)
{
+ int rc = 0;
ENTRY;
- spin_lock(&mds_exec_lock);
- if (atomic_read(&inode->i_writecount) > 0) {
- spin_unlock(&mds_exec_lock);
- RETURN(-ETXTBSY);
- }
+
+ down(&mds->mds_epoch_sem);
atomic_dec(&inode->i_writecount);
- spin_unlock(&mds_exec_lock);
- RETURN(0);
+ rc = atomic_read(&inode->i_writecount);
+ if (rc > 0)
+ GOTO(out, rc);
+#if 0
+ if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
+ GOTO(out, rc = EAGAIN);
+#endif
+ mds_free_filterdata(inode);
+ out:
+ up(&mds->mds_epoch_sem);
+ return rc;
}
-static void mds_put_write_access(struct inode * inode)
+static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
{
ENTRY;
+ down(&mds->mds_epoch_sem);
+ if (atomic_read(&inode->i_writecount) > 0) {
+ up(&mds->mds_epoch_sem);
+ RETURN(-ETXTBSY);
+ }
atomic_dec(&inode->i_writecount);
+ up(&mds->mds_epoch_sem);
+ RETURN(0);
}
static void mds_allow_write_access(struct inode *inode)
struct ptlrpc_request *req)
{
struct mds_export_data *med = &req->rq_export->exp_mds_data;
+ struct mds_obd *mds = mds_req2mds(req);
struct mds_file_data *mfd;
+ struct mds_body *body;
int error;
ENTRY;
-
+
mfd = mds_mfd_new();
if (mfd == NULL) {
CERROR("mds: out of memory\n");
GOTO(cleanup_dentry, error = -ENOMEM);
}
+ body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
+
if (flags & FMODE_WRITE) {
- error = mds_get_write_access(dentry->d_inode);
+ /* FIXME: in recovery, need to pass old epoch here */
+ error = mds_get_write_access(mds, dentry->d_inode, 0);
if (error)
GOTO(cleanup_mfd, error);
+ body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
} else if (flags & FMODE_EXEC) {
- error = mds_deny_write_access(dentry->d_inode);
+ error = mds_deny_write_access(mds, dentry->d_inode);
if (error)
GOTO(cleanup_mfd, error);
}
list_add(&mfd->mfd_list, &med->med_open_head);
spin_unlock(&med->med_open_lock);
mds_mfd_put(mfd);
+
+ body->handle.cookie = mfd->mfd_handle.h_cookie;
+
RETURN(mfd);
cleanup_mfd:
}
/* replay case */
- if (rec->ur_fid2->id) {
+ if(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
+ LASSERT (rec->ur_fid2->id);
body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
lmm_size = rec->ur_eadatalen;
lmm = rec->ur_eadata;
RETURN(0);
}
+
+ if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
+ GOTO(out_ids, rc = -ENOMEM);
+
oa = obdo_alloc();
if (oa == NULL)
GOTO(out_ids, rc = -ENOMEM);
obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
OBD_MD_FLCTIME);
- /* check if things like lstripe/lfs stripe are sending us the ea */
- if (rec->ur_flags & MDS_OPEN_HAS_EA) {
- rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, mds->mds_osc_exp,
+ if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
+ /* check if things like lstripe/lfs stripe are sending us the ea */
+ if (rec->ur_flags & MDS_OPEN_HAS_EA) {
+ rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
+ mds->mds_osc_exp,
+ 0, &lsm, rec->ur_eadata);
+ if (rc)
+ GOTO(out_oa, rc);
+ }
+ rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
+ if (rc) {
+ int level = D_ERROR;
+ if (rc == -ENOSPC)
+ level = D_INODE;
+ CDEBUG(level, "error creating objects for "
+ "inode %lu: rc = %d\n",
+ inode->i_ino, rc);
+ if (rc > 0) {
+ CERROR("obd_create returned invalid "
+ "rc %d\n", rc);
+ rc = -EIO;
+ }
+ GOTO(out_oa, rc);
+ }
+ } else {
+ rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
0, &lsm, rec->ur_eadata);
- if (rc)
+ if (rc) {
GOTO(out_oa, rc);
- }
-
- rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
- if (rc) {
- int level = D_ERROR;
- if (rc == -ENOSPC)
- level = D_INODE;
- CDEBUG(level, "error creating objects for inode %lu: rc = %d\n",
- inode->i_ino, rc);
- if (rc > 0) {
- CERROR("obd_create returned invalid rc %d\n", rc);
- rc = -EIO;
}
- GOTO(out_oa, rc);
+ lsm->lsm_object_id = oa->o_id;
}
-
if (inode->i_size) {
oa->o_size = inode->i_size;
obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
RETURN(rc);
}
-/* Caller must hold mds->mds_epoch_sem */
-static int mds_alloc_filterdata(struct inode *inode)
-{
- LASSERT(inode->i_filterdata == NULL);
- OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
- if (inode->i_filterdata == NULL)
- return -ENOMEM;
- LASSERT(igrab(inode) == inode);
- return 0;
-}
-
-/* Caller must hold mds->mds_epoch_sem */
-static void mds_free_filterdata(struct inode *inode)
-{
- LASSERT(inode->i_filterdata != NULL);
- OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
- inode->i_filterdata = NULL;
- iput(inode);
-}
-
-/* epoch argument is nonzero during recovery */
-static int mds_open_io_epoch(struct mds_obd *mds, struct inode *inode,
- __u64 epoch)
-{
- int rc = 0;
-
- down(&mds->mds_epoch_sem);
- if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
- CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
- MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
- inode->i_generation);
- goto out;
- }
-
- if (inode->i_filterdata == NULL)
- mds_alloc_filterdata(inode);
- if (inode->i_filterdata == NULL) {
- rc = -ENOMEM;
- goto out;
- }
- if (epoch > mds->mds_io_epoch)
- mds->mds_io_epoch = epoch;
- else
- mds->mds_io_epoch++;
- MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
- CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
- mds->mds_io_epoch, inode->i_ino, inode->i_generation);
- out:
- up(&mds->mds_epoch_sem);
- return rc;
-}
-
-/* Returns EAGAIN if the client needs to get size and/or cookies and close
- * again -- which is never true if the file is about to be unlinked. */
-static int mds_close_io_epoch(struct mds_obd *mds, struct inode *inode,
- struct mds_body *body, int unlinking)
-{
- int rc = 0;
- ENTRY;
-
- down(&mds->mds_epoch_sem);
- if (mds_query_write_access(inode) > 0)
- GOTO(out, rc);
-#if 0
- if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
- GOTO(out, rc = EAGAIN);
-#endif
- mds_free_filterdata(inode);
- out:
- up(&mds->mds_epoch_sem);
- return rc;
-}
-
static void reconstruct_open(struct mds_update_record *rec, int offset,
struct ptlrpc_request *req,
struct lustre_handle *child_lockh)
{
- struct ptlrpc_request *oldreq = req->rq_export->exp_outstanding_reply;
struct mds_export_data *med = &req->rq_export->exp_mds_data;
struct mds_client_data *mcd = med->med_mcd;
struct mds_obd *mds = mds_req2mds(req);
mfd = NULL;
}
- if (oldreq != NULL) {
- /* if we're not recovering, it had better be found */
- LASSERT(mfd != NULL);
- } else if (mfd == NULL) {
+#warning "XXX fixme"
+ /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
+ * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
+ * to detect a re-open */
+ if (mfd == NULL) {
mntget(mds->mds_vfsmnt);
CERROR("Re-opened file \n");
mfd = mds_dentry_open(child, mds->mds_vfsmnt,
put_child = 0;
}
- if ((rec->ur_flags & FMODE_WRITE) &&
- mds_open_io_epoch(mds, child->d_inode, 0) == 0)
- body->io_epoch = MDS_FILTERDATA(child->d_inode)->io_epoch;
- body->handle.cookie = mfd->mfd_handle.h_cookie;
-
out_dput:
if (put_child)
l_dput(child);
struct obd_device *obd = req->rq_export->exp_obd;
struct mds_file_data *mfd = NULL;
obd_id *ids = NULL; /* object IDs created */
- int rc;
+ int rc = 0;
ENTRY;
/* atomically create objects if necessary */
if (IS_ERR(mfd))
RETURN(PTR_ERR(mfd));
- if ((rec->ur_flags & FMODE_WRITE) &&
- mds_open_io_epoch(mds, dchild->d_inode, 0) == 0)
- body->io_epoch = MDS_FILTERDATA(dchild->d_inode)->io_epoch;
- body->handle.cookie = mfd->mfd_handle.h_cookie;
-
CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
mfd->mfd_handle.h_cookie);
if (ids != NULL) {
mds_lov_update_objids(obd, ids);
OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
}
- if (rc)
- mds_mfd_destroy(mfd);
+ //if (rc)
+ // mds_mfd_destroy(mfd);
RETURN(rc);
}
int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
struct lustre_handle *child_lockh)
{
- struct ldlm_res_id child_res_id = { .name = { inode->i_ino } };
+ struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
struct lustre_handle lockh;
int lock_flags = 0;
int rc;
if (child_lockh == NULL)
child_lockh = &lockh;
- rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
- child_res_id, LDLM_PLAIN, NULL, 0,
- LCK_EX, &lock_flags, ldlm_completion_ast,
- mds_blocking_ast, NULL, child_lockh);
+ rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
+ LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
+ mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
+ NULL, 0, NULL, child_lockh);
if (rc != ELDLM_OK)
CERROR("ldlm_cli_enqueue: %d\n", rc);
else if (child_lockh == &lockh)
ldlm_lock_decref(child_lockh, LCK_EX);
- return rc;
+ RETURN(rc);
}
int mds_open(struct mds_update_record *rec, int offset,
/* Step 0: If we are passed a fid, then we assume the client already
* opened this file and is only replaying the RPC, so we open the
* inode by fid (at some large expense in security). */
- if (rec->ur_fid2->id) {
+ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
+ DEBUG_REQ(D_HA, req, "open replay, disp: "LPX64"\n",
+ rep->lock_policy_res1);
+
+ LASSERT(rec->ur_fid2->id);
+
rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
rec, rep);
- if (rc != -ENOENT)
+ if (rc != -ENOENT)
RETURN(rc);
/* We didn't find the correct inode on disk either, so we
* need to re-create it via a regular replay. */
LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
+ } else {
+ LASSERT(!rec->ur_fid2->id);
}
+
LASSERT(offset == 2); /* If we got here, we must be called via intent */
med = &req->rq_export->exp_mds_data;
/* Step 2: Lookup the child */
dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
- if (IS_ERR(dchild))
- GOTO(cleanup, rc = PTR_ERR(dchild));
+ if (IS_ERR(dchild)) {
+ rc = PTR_ERR(dchild);
+ dchild = NULL; /* don't confuse mds_finish_transno() below */
+ GOTO(cleanup, rc);
+ }
cleanup_phase = 2; /* child dentry */
if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
!S_ISDIR(dchild->d_inode->i_mode))
GOTO(cleanup, rc = -ENOTDIR);
+
+ if (S_ISDIR(dchild->d_inode->i_mode)) {
+ if (rec->ur_flags & MDS_OPEN_CREAT || rec->ur_flags & FMODE_WRITE) {
+ /*we are tryying to create or write a exist dir*/
+ GOTO(cleanup, rc = -EISDIR);
+ }
+ if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
+ GOTO(cleanup, rc = -EACCES);
+ }
+ }
/* Step 5: mds_open it */
rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
cleanup:
rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
req, rc, rep ? rep->lock_policy_res1 : 0);
- /* XXX what do we do here if mds_finish_transno itself failed? */
- if (created)
- mds_lock_new_child(obd, dchild->d_inode, NULL);
switch (cleanup_phase) {
case 2:
dchild->d_name.len, dchild->d_name.name,
err);
}
+ } else if (created) {
+ mds_lock_new_child(obd, dchild->d_inode, NULL);
}
l_dput(dchild);
case 1:
if (rc)
ldlm_lock_decref(&parent_lockh, parent_mode);
else
- ldlm_put_lock_into_req(req, &parent_lockh, parent_mode);
+ ptlrpc_save_lock (req, &parent_lockh, parent_mode);
}
+ if (rc == 0)
+ atomic_inc(&mds->mds_open_count);
RETURN(rc);
}
struct mds_obd *mds = &obd->u.mds;
struct inode *pending_dir = mds->mds_pending_dir->d_inode;
void *handle = NULL;
- struct mds_body *request_body, *reply_body;
+ struct mds_body *request_body = NULL, *reply_body = NULL;
struct dentry_params dp;
ENTRY;
last_orphan = mds_open_orphan_dec_test(inode) &&
mds_inode_is_orphan(inode);
- /* this is the actual "close" */
+ /* this is half of the actual "close" */
if (mfd->mfd_mode & FMODE_WRITE) {
- mds_put_write_access(inode);
+ rc = mds_put_write_access(mds, inode, request_body,
+ last_orphan && unlink_orphan);
} else if (mfd->mfd_mode & FMODE_EXEC) {
mds_allow_write_access(inode);
}
if (last_orphan && unlink_orphan) {
- LASSERT(mds_query_write_access(inode) == 0);
- if (mfd->mfd_mode & FMODE_WRITE)
- rc = mds_close_io_epoch(mds, inode, request_body, 1);
+ LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
- CWARN("destroying orphan object %s\n", fidname);
+ CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
/* Sadly, there is no easy way to save pending_child from
* mds_reint_unlink() into mfd, so we need to re-lookup,
GOTO(cleanup, rc);
}
-#ifdef ENABLE_ORPHANS
if (req != NULL &&
(reply_body->valid & OBD_MD_FLEASIZE) &&
mds_log_op_unlink(obd, pending_child->d_inode,
- req->rq_repmsg, 1) > 0) {
+ lustre_msg_buf(req->rq_repmsg, 1, 0),
+ req->rq_repmsg->buflens[1],
+ lustre_msg_buf(req->rq_repmsg, 2, 0),
+ req->rq_repmsg->buflens[2]) > 0) {
reply_body->valid |= OBD_MD_FLCOOKIE;
}
-#endif
+
pending_child->d_fsdata = (void *) &dp;
dp.p_inum = 0;
dp.p_ptr = req;
rc = vfs_unlink(pending_dir, pending_child);
if (rc)
CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
- } else if (mfd->mfd_mode & FMODE_WRITE &&
- mds_query_write_access(inode) == 0) {
- // XXX this should probably be abstracted with mds_reint_setattr
+ } else if (mfd->mfd_mode & FMODE_WRITE && rc == 0) {
+ /* Update the on-disk attributes if this was the last write
+ * close, and all information was provided (i.e., rc == 0)
+ *
+ * XXX this should probably be abstracted with mds_reint_setattr
+ */
+#if 0
struct iattr iattr;
- rc = mds_close_io_epoch(mds, inode, request_body, 0);
- if (rc == EAGAIN)
- goto dput;
-
-#if 0
/* XXX can't set block count with fsfilt_setattr (!) */
iattr.ia_valid = ATTR_CTIME | ATTR_ATIME |
ATTR_MTIME | ATTR_SIZE;
CERROR("error in setattr(%s): rc %d\n", fidname, rc);
#endif
}
- dput:
+ /* If other clients have this file open for write, rc will be > 0 */
+ if (rc > 0)
+ rc = 0;
l_dput(mfd->mfd_dentry);
mds_mfd_destroy(mfd);
cleanup:
+ atomic_dec(&mds->mds_open_count);
if (req) {
rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
} else if (handle) {
obd->u.mds.mds_max_cookiesize};
ENTRY;
- MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
+ rc = lustre_pack_reply(req, 3, repsize, NULL);
+ if (rc) {
+ CERROR("lustre_pack_reply: rc = %d\n", rc);
+ req->rq_status = rc;
+ } else {
+ MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
+ }
body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
if (body == NULL) {
RETURN(-ESTALE);
}
- rc = lustre_pack_reply(req, 3, repsize, NULL);
- if (rc) {
- CERROR("lustre_pack_reply: rc = %d\n", rc);
- req->rq_status = rc;
- }
-
inode = mfd->mfd_dentry->d_inode;
if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));