* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2013, Intel Corporation.
+ * Copyright (c) 2012, 2014, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#define DEBUG_SUBSYSTEM S_MDS
#include <lustre_log.h>
+#include <lustre_update.h>
#include "osp_internal.h"
static int osp_sync_id_traction_init(struct osp_device *d);
#define OSP_JOB_MAGIC 0x26112005
+struct osp_job_req_args {
+ /** bytes reserved for ptlrpc_replay_req() */
+ struct ptlrpc_replay_async_args jra_raa;
+ struct list_head jra_link;
+ __u32 jra_magic;
+};
+
static inline int osp_sync_running(struct osp_device *d)
{
return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
}
+/**
+ * Check status: whether OSP thread has stopped
+ *
+ * \param[in] d OSP device
+ *
+ * \retval 0 still running
+ * \retval 1 stopped
+ */
static inline int osp_sync_stopped(struct osp_device *d)
{
return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
}
+/*
+ ** Check for new changes to sync
+ *
+ * \param[in] d OSP device
+ *
+ * \retval 1 there are changes
+ * \retval 0 there are no changes
+ */
static inline int osp_sync_has_new_job(struct osp_device *d)
{
return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
}
+/**
+ * Check for room in the network pipe to OST
+ *
+ * \param[in] d OSP device
+ *
+ * \retval 1 there is room
+ * \retval 0 no room, the pipe is full
+ */
static inline int osp_sync_low_in_flight(struct osp_device *d)
{
return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
}
+/**
+ * Wake up check for the main sync thread
+ *
+ * \param[in] d OSP device
+ *
+ * \retval 1 time to wake up
+ * \retval 0 no need to wake up
+ */
static inline int osp_sync_has_work(struct osp_device *d)
{
/* has new/old changes and low in-progress? */
osp_sync_check_for_work(d);
}
+/**
+ * Check and return ready-for-new status.
+ *
+ * The thread processing llog record uses this function to check whether
+ * it's time to take another record and process it. The number of conditions
+ * must be met: the connection should be ready, RPCs in flight not exceeding
+ * the limit, the record is committed locally, etc (see the lines below).
+ *
+ * \param[in] d OSP device
+ * \param[in] rec next llog record to process
+ *
+ * \retval 0 not ready
+ * \retval 1 ready
+ */
static inline int osp_sync_can_process_new(struct osp_device *d,
struct llog_rec_hdr *rec)
{
LASSERT(d);
+ if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
+ return 0;
if (!osp_sync_low_in_progress(d))
return 0;
if (!osp_sync_low_in_flight(d))
return 0;
}
+/**
+ * Declare intention to add a new change.
+ *
+ * With regard to OSD API, we have to declare any changes ahead. In this
+ * case we declare an intention to add a llog record representing the
+ * change on the local storage.
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] o OSP object
+ * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
+ * \param[in] th transaction handle (local)
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
+ */
int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
llog_op_type type, struct thandle *th)
{
struct osp_thread_info *osi = osp_env_info(env);
struct osp_device *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
struct llog_ctxt *ctxt;
+ struct thandle *storage_th;
int rc;
ENTRY;
/* it's a layering violation, to access internals of th,
* but we can do this as a sanity check, for a while */
- LASSERT(th->th_dev == d->opd_storage);
+ LASSERT(th->th_top != NULL);
+ storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
+ if (IS_ERR(storage_th))
+ RETURN(PTR_ERR(storage_th));
switch (type) {
case MDS_UNLINK64_REC:
}
/* we want ->dt_trans_start() to allocate per-thandle structure */
- th->th_tags |= LCT_OSP_THREAD;
+ storage_th->th_tags |= LCT_OSP_THREAD;
ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
LASSERT(ctxt);
- rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr, th);
+ rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr,
+ storage_th);
llog_ctxt_put(ctxt);
RETURN(rc);
}
+/**
+ * Generate a llog record for a given change.
+ *
+ * Generates a llog record for the change passed. The change can be of two
+ * types: unlink and setattr. The record gets an ID which later will be
+ * used to track commit status of the change. For unlink changes, the caller
+ * can supply a starting FID and the count of the objects to destroy. For
+ * setattr the caller should apply attributes to apply.
+ *
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] d OSP device
+ * \param[in] fid fid of the object the change should be applied to
+ * \param[in] type type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
+ * \param[in] count count of objects to destroy
+ * \param[in] th transaction handle (local)
+ * \param[in] attr attributes for setattr
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
+ */
static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
const struct lu_fid *fid, llog_op_type type,
int count, struct thandle *th,
struct osp_thread_info *osi = osp_env_info(env);
struct llog_ctxt *ctxt;
struct osp_txn_info *txn;
+ struct thandle *storage_th;
int rc;
ENTRY;
/* it's a layering violation, to access internals of th,
* but we can do this as a sanity check, for a while */
- LASSERT(th->th_dev == d->opd_storage);
+ LASSERT(th->th_top != NULL);
+ storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
+ if (IS_ERR(storage_th))
+ RETURN(PTR_ERR(storage_th));
switch (type) {
case MDS_UNLINK64_REC:
LASSERT(attr);
osi->osi_setattr.lsr_uid = attr->la_uid;
osi->osi_setattr.lsr_gid = attr->la_gid;
- osi->osi_setattr.lsr_valid = attr->la_valid;
+ osi->osi_setattr.lsr_valid =
+ ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
+ ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0);
break;
default:
LBUG();
}
- txn = osp_txn_info(&th->th_ctx);
+ txn = osp_txn_info(&storage_th->th_ctx);
LASSERT(txn);
txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
if (ctxt == NULL)
RETURN(-ENOMEM);
+
rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
- th);
+ storage_th);
llog_ctxt_put(ctxt);
- CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
- d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
- (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
- (unsigned long) osi->osi_cookie.lgc_index, rc);
-
- if (rc > 0)
- rc = 0;
-
- if (likely(rc == 0)) {
+ if (likely(rc >= 0)) {
+ CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
+ d->opd_obd->obd_name,
+ POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
+ (unsigned long)osi->osi_cookie.lgc_lgl.lgl_ogen,
+ (unsigned long)osi->osi_cookie.lgc_index, rc);
spin_lock(&d->opd_syn_lock);
d->opd_syn_changes++;
spin_unlock(&d->opd_syn_lock);
}
-
- RETURN(rc);
+ /* return 0 always here, error case just cause no llog record */
+ RETURN(0);
}
int osp_sync_add(const struct lu_env *env, struct osp_object *o,
}
int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
- struct lu_fid *fid, int lost, struct thandle *th)
+ struct lu_fid *fid, int lost, struct thandle *th)
{
return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
}
* subsequent commit callback (at the most)
*/
-/*
- * called for each atomic on-disk change (not once per transaction batch)
- * and goes over the list
- * XXX: should be optimized?
- */
-
/**
- * called for each RPC reported committed
+ * ptlrpc commit callback.
+ *
+ * The callback is called by PTLRPC when a RPC is reported committed by the
+ * target (OST). We register the callback for the every RPC applying a change
+ * from the llog. This way we know then the llog records can be cancelled.
+ * Notice the callback can be called when OSP is finishing. We can detect this
+ * checking that actual transno in the request is less or equal of known
+ * committed transno (see osp_sync_process_committed() for the details).
+ * XXX: this is pretty expensive and can be improved later using batching.
+ *
+ * \param[in] req request
*/
static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
{
struct osp_device *d = req->rq_cb_data;
+ struct osp_job_req_args *jra;
CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
/* do not do any opd_dyn_rpc_* accounting here
* it's done in osp_sync_interpret sooner or later */
-
LASSERT(d);
- LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
- LASSERT(list_empty(&req->rq_exp_list));
+
+ jra = ptlrpc_req_async_args(req);
+ LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
+ LASSERT(list_empty(&jra->jra_link));
ptlrpc_request_addref(req);
spin_lock(&d->opd_syn_lock);
- list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
+ list_add(&jra->jra_link, &d->opd_syn_committed_there);
spin_unlock(&d->opd_syn_lock);
/* XXX: some batching wouldn't hurt */
wake_up(&d->opd_syn_waitq);
}
+/**
+ * RPC interpretation callback.
+ *
+ * The callback is called by ptlrpc when RPC is replied. Now we have to decide
+ * whether we should:
+ * - put request on a special list to wait until it's committed by the target,
+ * if the request is successful
+ * - schedule llog record cancel if no target object is found
+ * - try later (essentially after reboot) in case of unexpected error
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] req request replied
+ * \param[in] aa callback data
+ * \param[in] rc result of RPC
+ *
+ * \retval 0 always
+ */
static int osp_sync_interpret(const struct lu_env *env,
struct ptlrpc_request *req, void *aa, int rc)
{
struct osp_device *d = req->rq_cb_data;
+ struct osp_job_req_args *jra = aa;
- if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
- DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
- LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
+ if (jra->jra_magic != OSP_JOB_MAGIC) {
+ DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
+ LBUG();
+ }
LASSERT(d);
CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
* but object doesn't exist anymore - cancell llog record
*/
LASSERT(req->rq_transno == 0);
- LASSERT(list_empty(&req->rq_exp_list));
+ LASSERT(list_empty(&jra->jra_link));
ptlrpc_request_addref(req);
spin_lock(&d->opd_syn_lock);
- list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
+ list_add(&jra->jra_link, &d->opd_syn_committed_there);
spin_unlock(&d->opd_syn_lock);
wake_up(&d->opd_syn_waitq);
spin_lock(&d->opd_syn_lock);
d->opd_syn_rpc_in_flight--;
spin_unlock(&d->opd_syn_lock);
+ if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
+ wake_up(&d->opd_syn_barrier_waitq);
CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
d->opd_syn_rpc_in_progress);
}
/*
- * the function walks through list of committed locally changes
- * and send them to RPC until the pipe is full
+ ** Add request to ptlrpc queue.
+ *
+ * This is just a tiny helper function to put the request on the sending list
+ *
+ * \param[in] d OSP device
+ * \param[in] req request
*/
static void osp_sync_send_new_rpc(struct osp_device *d,
struct ptlrpc_request *req)
{
+ struct osp_job_req_args *jra;
+
LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
- LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
+
+ jra = ptlrpc_req_async_args(req);
+ jra->jra_magic = OSP_JOB_MAGIC;
+ INIT_LIST_HEAD(&jra->jra_link);
ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
}
+
+/**
+ * Allocate and prepare RPC for a new change.
+ *
+ * The function allocates and initializes an RPC which will be sent soon to
+ * apply the change to the target OST. The request is initialized from the
+ * llog record passed. Notice only the fields common to all type of changes
+ * are initialized.
+ *
+ * \param[in] d OSP device
+ * \param[in] llh llog handle where the record is stored
+ * \param[in] h llog record
+ * \param[in] op type of the change
+ * \param[in] format request format to be used
+ *
+ * \retval pointer new request on success
+ * \retval ERR_PTR(errno) on error
+ */
static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
struct llog_handle *llh,
struct llog_rec_hdr *h,
body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
body->oa.o_lcookie.lgc_index = h->lrh_index;
- INIT_LIST_HEAD(&req->rq_exp_list);
- req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
req->rq_interpret_reply = osp_sync_interpret;
req->rq_commit_cb = osp_sync_request_commit_cb;
return req;
}
+/**
+ * Generate a request for setattr change.
+ *
+ * The function prepares a new RPC, initializes it with setattr specific
+ * bits and send the RPC.
+ *
+ * \param[in] d OSP device
+ * \param[in] llh llog handle where the record is stored
+ * \param[in] h llog record
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
+ */
static int osp_sync_new_setattr_job(struct osp_device *d,
struct llog_handle *llh,
struct llog_rec_hdr *h)
ENTRY;
LASSERT(h->lrh_type == MDS_SETATTR64_REC);
- /* lsr_valid can only be 0 or LA_UID/GID set */
- if (!rec->lsr_valid && !(rec->lsr_valid & ~(LA_UID | LA_GID))) {
+ /* lsr_valid can only be 0 or have OBD_MD_{FLUID,FLGID} set,
+ * so no bits other than these should be set. */
+ if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID)) != 0) {
CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
d->opd_obd->obd_name, rec->lsr_valid);
- RETURN(-EINVAL);
+ /* return 0 so that sync thread can continue processing
+ * other records. */
+ RETURN(0);
}
req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
/* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
* we assume that both UID and GID are valid in that case. */
- if (rec->lsr_valid == 0) {
+ if (rec->lsr_valid == 0)
body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
- } else {
- if (rec->lsr_valid & LA_UID)
- body->oa.o_valid |= OBD_MD_FLUID;
- if (rec->lsr_valid & LA_GID)
- body->oa.o_valid |= OBD_MD_FLGID;
- }
+ else
+ body->oa.o_valid |= rec->lsr_valid;
osp_sync_send_new_rpc(d, req);
- RETURN(0);
+ RETURN(1);
}
-/* Old records may be in old format, so we handle that too */
+/**
+ * Generate a request for unlink change.
+ *
+ * The function prepares a new RPC, initializes it with unlink(destroy)
+ * specific bits and sends the RPC. The function is used to handle
+ * llog_unlink_rec which were used in the older versions of Lustre.
+ * Current version uses llog_unlink_rec64.
+ *
+ * \param[in] d OSP device
+ * \param[in] llh llog handle where the record is stored
+ * \param[in] h llog record
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
+ */
static int osp_sync_new_unlink_job(struct osp_device *d,
struct llog_handle *llh,
struct llog_rec_hdr *h)
body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
osp_sync_send_new_rpc(d, req);
- RETURN(0);
-}
-
-static int osp_prep_unlink_update_req(const struct lu_env *env,
- struct osp_device *osp,
- struct llog_handle *llh,
- struct llog_rec_hdr *h,
- struct ptlrpc_request **reqp)
-{
- struct llog_unlink64_rec *rec = (struct llog_unlink64_rec *)h;
- struct dt_update_request *update = NULL;
- struct ptlrpc_request *req;
- const char *buf;
- struct llog_cookie lcookie;
- int size;
- int rc;
- ENTRY;
-
- update = out_create_update_req(&osp->opd_dt_dev);
- if (IS_ERR(update))
- RETURN(PTR_ERR(update));
-
- /* This can only happens for unlink slave directory, so decrease
- * ref for ".." and "." */
- rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
- NULL, NULL);
- if (rc != 0)
- GOTO(out, rc);
-
- rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
- NULL, NULL);
- if (rc != 0)
- GOTO(out, rc);
-
- lcookie.lgc_lgl = llh->lgh_id;
- lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
- lcookie.lgc_index = h->lrh_index;
- size = sizeof(lcookie);
- buf = (const char *)&lcookie;
-
- rc = out_insert_update(env, update, OUT_DESTROY, &rec->lur_fid, 1,
- &size, &buf);
- if (rc != 0)
- GOTO(out, rc);
-
- rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
- update->dur_req, &req);
- if (rc != 0)
- GOTO(out, rc);
-
- INIT_LIST_HEAD(&req->rq_exp_list);
- req->rq_svc_thread = (void *)OSP_JOB_MAGIC;
-
- req->rq_interpret_reply = osp_sync_interpret;
- req->rq_commit_cb = osp_sync_request_commit_cb;
- req->rq_cb_data = osp;
-
- ptlrpc_request_set_replen(req);
- *reqp = req;
-out:
- if (update != NULL)
- out_destroy_update_req(update);
-
- RETURN(rc);
+ RETURN(1);
}
+/**
+ * Generate a request for unlink change.
+ *
+ * The function prepares a new RPC, initializes it with unlink(destroy)
+ * specific bits and sends the RPC. Depending on the target (MDT or OST)
+ * two different protocols are used. For MDT we use OUT (basically OSD API
+ * updates transferred via a network). For OST we still use the old
+ * protocol (OBD?), originally for compatibility. Later we can start to
+ * use OUT for OST as well, this will allow batching and better code
+ * unification.
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] d OSP device
+ * \param[in] llh llog handle where the record is stored
+ * \param[in] h llog record
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
+ */
static int osp_sync_new_unlink64_job(const struct lu_env *env,
struct osp_device *d,
struct llog_handle *llh,
ENTRY;
LASSERT(h->lrh_type == MDS_UNLINK64_REC);
+ req = osp_sync_new_job(d, llh, h, OST_DESTROY,
+ &RQF_OST_DESTROY);
+ if (IS_ERR(req))
+ RETURN(PTR_ERR(req));
- if (d->opd_connect_mdt) {
- rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
- if (rc != 0)
- RETURN(rc);
- } else {
- req = osp_sync_new_job(d, llh, h, OST_DESTROY,
- &RQF_OST_DESTROY);
- if (IS_ERR(req))
- RETURN(PTR_ERR(req));
-
- body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
- if (body == NULL)
- RETURN(-EFAULT);
- rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
- if (rc < 0)
- RETURN(rc);
- body->oa.o_misc = rec->lur_count;
- body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
- OBD_MD_FLOBJCOUNT;
- }
+ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
+ if (body == NULL)
+ RETURN(-EFAULT);
+ rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
+ if (rc < 0)
+ RETURN(rc);
+ body->oa.o_misc = rec->lur_count;
+ body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
+ OBD_MD_FLOBJCOUNT;
osp_sync_send_new_rpc(d, req);
- RETURN(0);
+ RETURN(1);
}
+/**
+ * Process llog records.
+ *
+ * This function is called to process the llog records committed locally.
+ * In the recovery model used by OSP we can apply a change to a remote
+ * target once corresponding transaction (like posix unlink) is committed
+ * locally so can't revert.
+ * Depending on the llog record type, a given handler is called that is
+ * responsible for preparing and sending the RPC to apply the change.
+ * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] d OSP device
+ * \param[in] llh llog handle where the record is stored
+ * \param[in] rec llog record
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
+ */
static int osp_sync_process_record(const struct lu_env *env,
struct osp_device *d,
struct llog_handle *llh,
CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
rec->lrh_type);
/* we should continue processing */
- return 0;
}
- if (likely(rc == 0)) {
+ /* rc > 0 means sync RPC being added to the queue */
+ if (likely(rc > 0)) {
spin_lock(&d->opd_syn_lock);
if (d->opd_syn_prev_done) {
LASSERT(d->opd_syn_changes > 0);
* NOTE: it's possible to meet same id if
* OST stores few stripes of same file
*/
- if (rec->lrh_id > d->opd_syn_last_processed_id)
+ if (rec->lrh_id > d->opd_syn_last_processed_id) {
d->opd_syn_last_processed_id = rec->lrh_id;
+ wake_up(&d->opd_syn_barrier_waitq);
+ }
d->opd_syn_changes--;
}
d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
d->opd_syn_rpc_in_progress);
spin_unlock(&d->opd_syn_lock);
+ rc = 0;
} else {
spin_lock(&d->opd_syn_lock);
d->opd_syn_rpc_in_flight--;
return rc;
}
+/**
+ * Cancel llog records for the committed changes.
+ *
+ * The function walks through the list of the committed RPCs and cancels
+ * corresponding llog records. see osp_sync_request_commit_cb() for the
+ * details.
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] d OSP device
+ */
static void osp_sync_process_committed(const struct lu_env *env,
struct osp_device *d)
{
struct obd_device *obd = d->opd_obd;
struct obd_import *imp = obd->u.cli.cl_import;
struct ost_body *body;
- struct ptlrpc_request *req, *tmp;
+ struct ptlrpc_request *req;
struct llog_ctxt *ctxt;
struct llog_handle *llh;
struct list_head list;
INIT_LIST_HEAD(&d->opd_syn_committed_there);
spin_unlock(&d->opd_syn_lock);
- list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
+ while (!list_empty(&list)) {
struct llog_cookie *lcookie = NULL;
-
- LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
- list_del_init(&req->rq_exp_list);
-
- if (d->opd_connect_mdt) {
- struct object_update_request *ureq;
- struct object_update *update;
- ureq = req_capsule_client_get(&req->rq_pill,
- &RMF_OUT_UPDATE);
- LASSERT(ureq != NULL &&
- ureq->ourq_magic == UPDATE_REQUEST_MAGIC);
-
- /* 1st/2nd is for decref . and .., 3rd one is for
- * destroy, where the log cookie is stored.
- * See osp_prep_unlink_update_req */
- update = object_update_request_get(ureq, 2, NULL);
- LASSERT(update != NULL);
- lcookie = object_update_param_get(update, 0, NULL);
- LASSERT(lcookie != NULL);
- } else {
- body = req_capsule_client_get(&req->rq_pill,
- &RMF_OST_BODY);
- LASSERT(body);
- lcookie = &body->oa.o_lcookie;
- }
+ struct osp_job_req_args *jra;
+
+ jra = list_entry(list.next, struct osp_job_req_args, jra_link);
+ LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
+ list_del_init(&jra->jra_link);
+
+ req = container_of((void *)jra, struct ptlrpc_request,
+ rq_async_args);
+ body = req_capsule_client_get(&req->rq_pill,
+ &RMF_OST_BODY);
+ LASSERT(body);
+ lcookie = &body->oa.o_lcookie;
/* import can be closing, thus all commit cb's are
* called we can check committness directly */
if (req->rq_transno <= imp->imp_peer_committed_transno) {
EXIT;
}
-/*
- * this is where most of queues processing happens
+/**
+ * The core of the syncing mechanism.
+ *
+ * This is a callback called by the llog processing function. Essentially it
+ * suspends llog processing until there is a record to process (it's supposed
+ * to be committed locally). The function handles RPCs committed by the target
+ * and cancels corresponding llog records.
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] llh llog handle we're processing
+ * \param[in] rec current llog record
+ * \param[in] data callback data containing a pointer to the device
+ *
+ * \retval 0 to ask the caller (llog_process()) to continue
+ * \retval LLOG_PROC_BREAK to ask the caller to break
*/
static int osp_sync_process_queues(const struct lu_env *env,
struct llog_handle *llh,
if (osp_sync_can_process_new(d, rec)) {
if (llh == NULL) {
/* ask llog for another record */
- CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
+ CDEBUG(D_HA, "%lu changes, %u in progress,"
+ " %u in flight\n",
d->opd_syn_changes,
d->opd_syn_rpc_in_progress,
d->opd_syn_rpc_in_flight);
} while (1);
}
-/*
- * this thread runs llog_cat_process() scanner calling our callback
+/**
+ * OSP sync thread.
+ *
+ * This thread runs llog_cat_process() scanner calling our callback
* to process llog records. in the callback we implement tricky
* state machine as we don't want to start scanning of the llog again
* and again, also we don't want to process too many records and send
* being synced to OST) the callback can suspend awaiting for some
* new conditions, like syncs completed.
*
- * in order to process llog records left by previous boots and to allow
+ * In order to process llog records left by previous boots and to allow
* llog_process_thread() to find something (otherwise it'd just exit
* immediately) we add a special GENERATATION record on each boot.
+ *
+ * \param[in] _arg a pointer to thread's arguments
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
*/
static int osp_sync_thread(void *_arg)
{
RETURN(0);
}
+/**
+ * Initialize llog.
+ *
+ * Initializes the llog. Specific llog to be used depends on the type of the
+ * target OSP represents (OST or MDT). The function adds appends a new llog
+ * record to mark the place where the records associated with this boot
+ * start.
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] d OSP device
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
+ */
static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
{
struct osp_thread_info *osi = osp_env_info(env);
OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
obd->obd_lvfs_ctxt.dt = d->opd_storage;
- if (d->opd_connect_mdt)
- lu_local_obj_fid(fid, SLAVE_LLOG_CATALOGS_OID);
- else
- lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
+ lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
&osi->osi_cid, fid);
- if (rc) {
- CERROR("%s: can't get id from catalogs: rc = %d\n",
- obd->obd_name, rc);
- RETURN(rc);
+ if (rc < 0) {
+ if (rc != -EFAULT) {
+ CERROR("%s: can't get id from catalogs: rc = %d\n",
+ obd->obd_name, rc);
+ RETURN(rc);
+ }
+
+ /* After sparse OST indices is supported, the CATALOG file
+ * may become a sparse file that results in failure on
+ * reading. Skip this error as the llog will be created
+ * later */
+ memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
+ rc = 0;
}
CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
RETURN(rc);
}
+/**
+ * Cleanup llog used for syncing.
+ *
+ * Closes and cleanups the llog. The function is called when the device is
+ * shutting down.
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] d OSP device
+ */
static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
{
struct llog_ctxt *ctxt;
llog_cleanup(env, ctxt);
}
-/*
- * initializes sync component of OSP
+/**
+ * Initialization of the sync component of OSP.
+ *
+ * Initializes the llog and starts a new thread to handle the changes to
+ * the remote target (OST or MDT).
+ *
+ * \param[in] env LU environment provided by the caller
+ * \param[in] d OSP device
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
*/
int osp_sync_init(const struct lu_env *env, struct osp_device *d)
{
d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
spin_lock_init(&d->opd_syn_lock);
init_waitqueue_head(&d->opd_syn_waitq);
+ init_waitqueue_head(&d->opd_syn_barrier_waitq);
init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
INIT_LIST_HEAD(&d->opd_syn_committed_there);
return rc;
}
+/**
+ * Stop the syncing thread.
+ *
+ * Asks the syncing thread to stop and wait until it's stopped.
+ *
+ * \param[in] d OSP device
+ *
+ * \retval 0
+ */
int osp_sync_fini(struct osp_device *d)
{
struct ptlrpc_thread *thread = &d->opd_syn_thread;
static struct list_head osp_id_tracker_list =
LIST_HEAD_INIT(osp_id_tracker_list);
+/**
+ * OSD commit callback.
+ *
+ * The function is used as a local OSD commit callback to track the highest
+ * committed llog record id. see osp_sync_id_traction_init() for the details.
+ *
+ * \param[in] th local transaction handle committed
+ * \param[in] cookie commit callback data (our private structure)
+ */
static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
{
struct osp_id_tracker *tr = cookie;
spin_unlock(&tr->otr_lock);
}
+/**
+ * Initialize commit tracking mechanism.
+ *
+ * Some setups may have thousands of OSTs and each will be represented by OSP.
+ * Meaning order of magnitute many more changes to apply every second. In order
+ * to keep the number of commit callbacks low this mechanism was introduced.
+ * The mechanism is very similar to transno used by MDT service: it's an single
+ * ID stream which can be assigned by any OSP to its llog records. The tricky
+ * part is that ID is stored in per-transaction data and re-used by all the OSPs
+ * involved in that transaction. Then all these OSPs are woken up utilizing a single OSD commit callback.
+ *
+ * The function initializes the data used by the tracker described above.
+ * A singler tracker per OSD device is created.
+ *
+ * \param[in] d OSP device
+ *
+ * \retval 0 on success
+ * \retval negative negated errno on error
+ */
static int osp_sync_id_traction_init(struct osp_device *d)
{
struct osp_id_tracker *tr, *found = NULL;
return rc;
}
+/**
+ * Release commit tracker.
+ *
+ * Decrease a refcounter on the tracker used by the given OSP device \a d.
+ * If no more users left, then the tracker is released.
+ *
+ * \param[in] d OSP device
+ */
static void osp_sync_id_traction_fini(struct osp_device *d)
{
struct osp_id_tracker *tr;
EXIT;
}
-/*
- * generates id for the tracker
+/**
+ * Generate a new ID on a tracker.
+ *
+ * Generates a new ID using the tracker associated with the given OSP device
+ * \a d, if the given ID \a id is non-zero. Unconditially adds OSP device to
+ * the wakeup list, so OSP won't miss when a transaction using the ID is
+ * committed. Notice ID is 32bit, but llog doesn't support >2^32 records anyway.
+ *
+ * \param[in] d OSP device
+ * \param[in] id 0 or ID generated previously
+ *
+ * \retval ID the caller should use
*/
static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
{
return id;
}
+/**
+ * Stop to propagate commit status to OSP.
+ *
+ * If the OSP does not have any llog records she's waiting to commit, then
+ * it is possible to unsubscribe from wakeups from the tracking using this
+ * method.
+ *
+ * \param[in] d OSP device not willing to wakeup
+ */
static void osp_sync_remove_from_tracker(struct osp_device *d)
{
struct osp_id_tracker *tr;
git://git.whamcloud.com / fs / lustre-release.git / blobdiff