*
*/
-/*
- **
+/**
* Check whether statfs data is expired
*
* OSP device caches statfs data for the target, the function checks
struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer);
LASSERT(d);
- if (d->opd_pre != NULL && osp_precreate_running(d))
+ if (osp_precreate_running(d))
wake_up(&d->opd_pre_waitq);
}
+static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs);
+
+/*
+ * The function updates current precreation status if broken, and
+ * updates that cached statfs state if functional, then wakes up waiters.
+ * We don't clear opd_pre_status directly here, but rather leave this
+ * to osp_pre_update_msfs() to do if everything is OK so that we don't
+ * have a race to clear opd_pre_status and then set it to -ENOSPC again.
+ *
+ * \param[in] d OSP device
+ * \param[in] msfs statfs data
+ * \param[in] rc new precreate status for device \a d
+ */
+static void osp_pre_update_status_msfs(struct osp_device *d,
+ struct obd_statfs *msfs, int rc)
+{
+ if (rc)
+ d->opd_pre_status = rc;
+ else
+ osp_pre_update_msfs(d, msfs);
+
+ wake_up(&d->opd_pre_user_waitq);
+}
+
+/* Pass in the old statfs data in case the limits have changed */
+void osp_pre_update_status(struct osp_device *d, int rc)
+{
+ osp_pre_update_status_msfs(d, &d->opd_statfs, rc);
+}
+
+
/**
* RPC interpret callback for OST_STATFS RPC
*
* \retval negative negated errno on error
*/
static int osp_statfs_interpret(const struct lu_env *env,
- struct ptlrpc_request *req,
- union ptlrpc_async_args *aa, int rc)
+ struct ptlrpc_request *req, void *args, int rc)
{
+ union ptlrpc_async_args *aa = args;
struct obd_import *imp = req->rq_import;
struct obd_statfs *msfs;
struct osp_device *d;
ENTRY;
- aa = ptlrpc_req_async_args(req);
+ aa = ptlrpc_req_async_args(aa, req);
d = aa->pointer_arg[0];
LASSERT(d);
if (msfs == NULL)
GOTO(out, rc = -EPROTO);
- d->opd_statfs = *msfs;
-
- osp_pre_update_status(d, rc);
+ if (d->opd_pre)
+ osp_pre_update_status_msfs(d, msfs, 0);
+ else
+ d->opd_statfs = *msfs;
/* schedule next update */
maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC;
imp = d->opd_obd->u.cli.cl_import;
LASSERT(imp);
- req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
+ req = ptlrpc_request_alloc(imp,
+ d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS);
if (req == NULL)
RETURN(-ENOMEM);
- rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
+ rc = ptlrpc_request_pack(req,
+ d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION,
+ d->opd_pre ? OST_STATFS : MDS_STATFS);
if (rc) {
ptlrpc_request_free(req);
RETURN(rc);
}
ptlrpc_request_set_replen(req);
- req->rq_request_portal = OST_CREATE_PORTAL;
+ if (d->opd_pre)
+ req->rq_request_portal = OST_CREATE_PORTAL;
ptlrpc_at_set_req_timeout(req);
- req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
- aa = ptlrpc_req_async_args(req);
+ req->rq_interpret_reply = osp_statfs_interpret;
+ aa = ptlrpc_req_async_args(aa, req);
aa->pointer_arg[0] = d;
/*
{
int rc;
+ if (d->opd_pre == NULL)
+ return 0;
+
/* XXX: do we really need locking here? */
spin_lock(&d->opd_pre_lock);
rc = osp_precreate_near_empty_nolock(env, d);
struct lu_buf *lb_oid = &oti->osi_lb;
struct lu_buf *lb_oseq = &oti->osi_lb2;
loff_t oid_off;
+ u64 oid;
loff_t oseq_off;
struct thandle *th;
int rc;
/* Note: through f_oid is only 32 bits, it will also write 64 bits
* for oid to keep compatibility with the previous version. */
- lb_oid->lb_buf = &fid->f_oid;
- lb_oid->lb_len = sizeof(u64);
- oid_off = sizeof(u64) * osp->opd_index;
+ oid = fid->f_oid;
+ osp_objid_buf_prep(lb_oid, &oid_off,
+ &oid, osp->opd_index);
- lb_oseq->lb_buf = &fid->f_seq;
- lb_oseq->lb_len = sizeof(u64);
- oseq_off = sizeof(u64) * osp->opd_index;
+ osp_objseq_buf_prep(lb_oseq, &oseq_off,
+ &fid->f_seq, osp->opd_index);
th = dt_trans_create(env, osp->opd_storage);
if (IS_ERR(th))
/* Update last_xxx to the new seq */
spin_lock(&osp->opd_pre_lock);
osp->opd_last_used_fid = *fid;
+ osp_fid_to_obdid(fid, &osp->opd_last_id);
osp->opd_gap_start_fid = *fid;
osp->opd_pre_used_fid = *fid;
osp->opd_pre_last_created_fid = *fid;
*
* The function finds how many objects should be precreated. Then allocates,
* prepares and schedules precreate RPC synchronously. Upon reply the function
- * wake ups the threads waiting for the new objects on this target. If the
+ * wakes up the threads waiting for the new objects on this target. If the
* target wasn't able to create all the objects requested, then the next
- * precreate will be asking less objects (i.e. slow precreate down).
+ * precreate will be asking for fewer objects (i.e. slow precreate down).
*
* \param[in] env LU environment provided by the caller
* \param[in] d OSP device
*fid = d->opd_pre_last_created_fid;
rc = osp_precreate_fids(env, d, fid, &grow);
- if (rc == 1) {
+ if (rc == 1)
/* Current seq has been used up*/
- if (!osp_is_fid_client(d)) {
- osp_pre_update_status(d, -ENOSPC);
- rc = -ENOSPC;
- }
- wake_up(&d->opd_pre_waitq);
- GOTO(out_req, rc);
- }
+ GOTO(out_req, rc = -ENOSPC);
if (!osp_is_fid_client(d)) {
/* Non-FID client will always send seq 0 because of
ready:
if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
- CERROR("%s: precreate fid "DFID" < local used fid "DFID
+ CERROR("%s: precreate fid "DFID" <= local used fid "DFID
": rc = %d\n", d->opd_obd->obd_name,
PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
GOTO(out_req, rc = -ESTALE);
* data is used to make this decision. If the latest result of statfs
* request (rc argument) is not success, then just mark OSP unavailable
* right away.
-
- * Add a bit of hysteresis so this flag isn't continually flapping,
- * and ensure that new files don't get extremely fragmented due to
- * only a small amount of available space in the filesystem.
- * We want to set the ENOSPC when there is less than reserved size
- * free and clear it when there is at least 2*reserved size free space.
- * the function updates current precreation status used: functional or not
*
- * \param[in] d OSP device
- * \param[in] rc new precreate status for device \a d
+ * The new statfs data is passed in \a msfs and needs to be stored into
+ * opd_statfs, but only after the various flags in os_state are set, so
+ * that the new statfs data is not visible without appropriate flags set.
+ * As such, there is no need to clear the flags here, since this is called
+ * with new statfs data, and they should not be cleared if sent from OST.
*
- * \retval 0 on success
- * \retval negative negated errno on error
+ * Add a bit of hysteresis so this flag isn't continually flapping, and
+ * ensure that new files don't get extremely fragmented due to only a
+ * small amount of available space in the filesystem. We want to set
+ * the ENOSPC/ENOINO flags unconditionally when there is less than the
+ * reserved size free, and still copy them from the old state when there
+ * is less than 2*reserved size free space or inodes.
+ *
+ * \param[in] d OSP device
+ * \param[in] msfs statfs data
*/
-void osp_pre_update_status(struct osp_device *d, int rc)
+static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs)
{
- struct obd_statfs *msfs = &d->opd_statfs;
- int old = d->opd_pre_status;
- __u64 available;
+ u32 old_state = d->opd_statfs.os_state;
+ u32 reserved_ino_low = 32; /* could be tunable in the future */
+ u32 reserved_ino_high = reserved_ino_low * 2;
+ u64 available_mb;
+
+ /* statfs structure not initialized yet */
+ if (unlikely(!msfs->os_type))
+ return;
+
+ /* if the low and high watermarks have not been initialized yet */
+ if (unlikely(d->opd_reserved_mb_high == 0 &&
+ d->opd_reserved_mb_low == 0)) {
+ /* Use ~0.1% by default to disable object allocation,
+ * and ~0.2% to enable, size in MB, set both watermark
+ */
+ spin_lock(&d->opd_pre_lock);
+ if (d->opd_reserved_mb_high == 0 &&
+ d->opd_reserved_mb_low == 0) {
+ d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) *
+ msfs->os_blocks) >> 20;
+ if (d->opd_reserved_mb_low == 0)
+ d->opd_reserved_mb_low = 1;
+ d->opd_reserved_mb_high =
+ (d->opd_reserved_mb_low << 1) + 1;
+ }
+ spin_unlock(&d->opd_pre_lock);
+ }
- d->opd_pre_status = rc;
- if (rc)
- goto out;
+ available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
+ if (msfs->os_ffree < reserved_ino_low)
+ msfs->os_state |= OS_STATE_ENOINO;
+ else if (msfs->os_ffree <= reserved_ino_high)
+ msfs->os_state |= old_state & OS_STATE_ENOINO;
+ /* else don't clear flags in new msfs->os_state sent from OST */
+
+ CDEBUG(D_INFO,
+ "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n",
+ d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree,
+ msfs->os_bavail, available_mb, d->opd_reserved_mb_high,
+ msfs->os_files, msfs->os_ffree, msfs->os_state,
+ d->opd_pre_status);
+ if (available_mb < d->opd_reserved_mb_low)
+ msfs->os_state |= OS_STATE_ENOSPC;
+ else if (available_mb <= d->opd_reserved_mb_high)
+ msfs->os_state |= old_state & OS_STATE_ENOSPC;
+ /* else don't clear flags in new msfs->os_state sent from OST */
+
+ if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
+ d->opd_pre_status = -ENOSPC;
+ if (!(old_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)))
+ CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n",
+ d->opd_obd->obd_name, msfs->os_state,
+ d->opd_pre_status);
+ CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n",
+ atomic_read(&d->opd_sync_changes),
+ atomic_read(&d->opd_sync_rpcs_in_progress));
+ } else if (old_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
+ d->opd_pre_status = 0;
+ spin_lock(&d->opd_pre_lock);
+ d->opd_pre_create_slow = 0;
+ d->opd_pre_create_count = OST_MIN_PRECREATE;
+ spin_unlock(&d->opd_pre_lock);
+ wake_up(&d->opd_pre_waitq);
- if (likely(msfs->os_type)) {
- if (unlikely(d->opd_reserved_mb_high == 0 &&
- d->opd_reserved_mb_low == 0)) {
- /* Use ~0.1% by default to disable object allocation,
- * and ~0.2% to enable, size in MB, set both watermark
- */
- spin_lock(&d->opd_pre_lock);
- if (d->opd_reserved_mb_high == 0 &&
- d->opd_reserved_mb_low == 0) {
- d->opd_reserved_mb_low =
- ((msfs->os_bsize >> 10) *
- msfs->os_blocks) >> 20;
- if (d->opd_reserved_mb_low == 0)
- d->opd_reserved_mb_low = 1;
- d->opd_reserved_mb_high =
- (d->opd_reserved_mb_low << 1) + 1;
- }
- spin_unlock(&d->opd_pre_lock);
- }
- /* in MB */
- available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
- if (msfs->os_ffree < 32)
- msfs->os_state |= OS_STATE_ENOINO;
- else if (msfs->os_ffree > 64)
- msfs->os_state &= ~OS_STATE_ENOINO;
-
- if (available < d->opd_reserved_mb_low)
- msfs->os_state |= OS_STATE_ENOSPC;
- else if (available > d->opd_reserved_mb_high)
- msfs->os_state &= ~OS_STATE_ENOSPC;
- if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
- d->opd_pre_status = -ENOSPC;
- if (old != -ENOSPC)
- CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
- "free, %llu avail, %llu MB avail, %u "
- "hwm -> %d: rc = %d\n",
- d->opd_obd->obd_name, msfs->os_blocks,
- msfs->os_bfree, msfs->os_bavail,
- available, d->opd_reserved_mb_high,
- d->opd_pre_status, rc);
- CDEBUG(D_INFO,
- "non-committed changes: %u, in progress: %u\n",
- atomic_read(&d->opd_sync_changes),
- atomic_read(&d->opd_sync_rpcs_in_progress));
- } else if (unlikely(old == -ENOSPC)) {
- d->opd_pre_status = 0;
- spin_lock(&d->opd_pre_lock);
- d->opd_pre_create_slow = 0;
- d->opd_pre_create_count = OST_MIN_PRECREATE;
- spin_unlock(&d->opd_pre_lock);
- wake_up(&d->opd_pre_waitq);
-
- CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
- " free, %llu avail, %lluMB avail, %u lwm"
- " -> %d: rc = %d\n", d->opd_obd->obd_name,
- msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
- available, d->opd_reserved_mb_low,
- d->opd_pre_status, rc);
- }
+ CDEBUG(D_INFO,
+ "%s: available: state=%x: rc = %d\n",
+ d->opd_obd->obd_name, msfs->os_state,
+ d->opd_pre_status);
+ } else {
+ /* we only get here if rc == 0 in the caller */
+ d->opd_pre_status = 0;
}
-out:
- wake_up(&d->opd_pre_user_waitq);
+
+ /* Object precreation skipped on OST if manually disabled */
+ if (d->opd_pre_max_create_count == 0)
+ msfs->os_state |= OS_STATE_NOPRECREATE;
+ /* else don't clear flags in new msfs->os_state sent from OST */
+
+ /* copy only new statfs state to make it visible to MDS threads */
+ if (&d->opd_statfs != msfs)
+ d->opd_statfs = *msfs;
}
/**
* need to be connected to OST
*/
while (osp_precreate_running(d)) {
- if (d->opd_pre_recovering &&
+ if ((d->opd_pre == NULL || d->opd_pre_recovering) &&
d->opd_imp_connected &&
!d->opd_got_disconnected)
break;
if (!osp_precreate_running(d))
break;
- LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
- /* Sigh, fid client is not ready yet */
- if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
- continue;
+ if (d->opd_pre) {
+ LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
+ /* Sigh, fid client is not ready yet */
+ if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
+ continue;
- /* Init fid for osp_precreate if necessary */
- rc = osp_init_pre_fid(d);
- if (rc != 0) {
- class_export_put(d->opd_exp);
- d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
- CERROR("%s: init pre fid error: rc = %d\n",
- d->opd_obd->obd_name, rc);
- continue;
+ /* Init fid for osp_precreate if necessary */
+ rc = osp_init_pre_fid(d);
+ if (rc != 0) {
+ class_export_put(d->opd_exp);
+ d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
+ CERROR("%s: init pre fid error: rc = %d\n",
+ d->opd_obd->obd_name, rc);
+ continue;
+ }
}
if (osp_statfs_update(&env, d)) {
continue;
}
- /*
- * Clean up orphans or recreate missing objects.
- */
- rc = osp_precreate_cleanup_orphans(&env, d);
- if (rc != 0) {
- schedule_timeout_interruptible(cfs_time_seconds(1));
- continue;
+ if (d->opd_pre) {
+ /*
+ * Clean up orphans or recreate missing objects.
+ */
+ rc = osp_precreate_cleanup_orphans(&env, d);
+ if (rc != 0) {
+ schedule_timeout_interruptible(cfs_time_seconds(1));
+ continue;
+ }
}
+
/*
* connected, can handle precreates now
*/
if (osp_statfs_update(&env, d))
break;
+ if (d->opd_pre == NULL)
+ continue;
+
/* To avoid handling different seq in precreate/orphan
* cleanup, it will hold precreate until current seq is
* used up. */
if (atomic_read(&d->opd_sync_changes) && synced == 0) {
/* force local commit to release space */
dt_commit_async(env, d->opd_storage);
- osp_sync_force(env, d);
+ osp_sync_check_for_work(d);
synced = 1;
}
if (atomic_read(&d->opd_sync_rpcs_in_progress)) {
if (req == NULL)
RETURN(-ENOMEM);
- /* XXX: capa support? */
- /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
if (rc) {
ptlrpc_request_free(req);
*/
int osp_init_precreate(struct osp_device *d)
{
- struct l_wait_info lwi = { 0 };
- struct task_struct *task;
-
ENTRY;
OBD_ALLOC_PTR(d->opd_pre);
RETURN(-ENOMEM);
/* initially precreation isn't ready */
+ init_waitqueue_head(&d->opd_pre_user_waitq);
d->opd_pre_status = -EAGAIN;
fid_zero(&d->opd_pre_used_fid);
d->opd_pre_used_fid.f_oid = 1;
fid_zero(&d->opd_pre_last_created_fid);
d->opd_pre_last_created_fid.f_oid = 1;
+ d->opd_last_id = 0;
d->opd_pre_reserved = 0;
d->opd_got_disconnected = 1;
d->opd_pre_create_slow = 0;
d->opd_reserved_mb_high = 0;
d->opd_reserved_mb_low = 0;
+ RETURN(0);
+}
+
+/**
+ * Finish precreate functionality of OSP
+ *
+ *
+ * Asks all the activity (the thread, update timer) to stop, then
+ * wait till that is done.
+ *
+ * \param[in] d OSP device
+ */
+void osp_precreate_fini(struct osp_device *d)
+{
+ ENTRY;
+
+ if (d->opd_pre == NULL)
+ RETURN_EXIT;
+
+ OBD_FREE_PTR(d->opd_pre);
+ d->opd_pre = NULL;
+
+ EXIT;
+}
+
+int osp_init_statfs(struct osp_device *d)
+{
+ struct l_wait_info lwi = { 0 };
+ struct task_struct *task;
+
+ ENTRY;
+
spin_lock_init(&d->opd_pre_lock);
init_waitqueue_head(&d->opd_pre_waitq);
- init_waitqueue_head(&d->opd_pre_user_waitq);
thread_set_flags(&d->opd_pre_thread, SVC_INIT);
init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
RETURN(0);
}
-/**
- * Finish precreate functionality of OSP
- *
- *
- * Asks all the activity (the thread, update timer) to stop, then
- * wait till that is done.
- *
- * \param[in] d OSP device
- */
-void osp_precreate_fini(struct osp_device *d)
+void osp_statfs_fini(struct osp_device *d)
{
struct ptlrpc_thread *thread = &d->opd_pre_thread;
ENTRY;
del_timer(&d->opd_statfs_timer);
- if (d->opd_pre == NULL)
- RETURN_EXIT;
-
if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
thread->t_flags = SVC_STOPPING;
wake_up(&d->opd_pre_waitq);
wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
}
- OBD_FREE_PTR(d->opd_pre);
- d->opd_pre = NULL;
-
EXIT;
}
-
git://git.whamcloud.com / fs / lustre-release.git / blobdiff