* Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2014, Intel Corporation.
+ * Copyright (c) 2012, 2015, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
/* Clients typically hold 2x their max_rpcs_in_flight of grant space */
#define OFD_GRANT_SHRINK_LIMIT(exp) (2ULL * 8 * exp_max_brw_size(exp))
-static inline obd_size ofd_grant_from_cli(struct obd_export *exp,
- struct ofd_device *ofd, obd_size val)
+static inline u64 ofd_grant_from_cli(struct obd_export *exp,
+ struct ofd_device *ofd, u64 val)
{
if (ofd_grant_compat(exp, ofd))
/* clients not supporting OBD_CONNECT_GRANT_PARAM actually
return val;
}
-static inline obd_size ofd_grant_to_cli(struct obd_export *exp,
- struct ofd_device *ofd, obd_size val)
+static inline u64 ofd_grant_to_cli(struct obd_export *exp,
+ struct ofd_device *ofd, u64 val)
{
if (ofd_grant_compat(exp, ofd))
return val >> (ofd->ofd_blockbits - COMPAT_BSIZE_SHIFT);
return val;
}
-static inline obd_size ofd_grant_chunk(struct obd_export *exp,
- struct ofd_device *ofd)
+static inline u64 ofd_grant_chunk(struct obd_export *exp,
+ struct ofd_device *ofd)
{
if (ofd_obd(ofd)->obd_self_export == exp)
/* Grant enough space to handle a big precreate request */
*/
void ofd_grant_sanity_check(struct obd_device *obd, const char *func)
{
- struct ofd_device *ofd = ofd_dev(obd->obd_lu_dev);
- struct obd_export *exp;
- obd_size maxsize;
- obd_size tot_dirty = 0;
- obd_size tot_pending = 0;
- obd_size tot_granted = 0;
- obd_size fo_tot_granted;
- obd_size fo_tot_pending;
- obd_size fo_tot_dirty;
-
- if (cfs_list_empty(&obd->obd_exports))
+ struct ofd_device *ofd = ofd_dev(obd->obd_lu_dev);
+ struct obd_export *exp;
+ u64 maxsize;
+ u64 tot_dirty = 0;
+ u64 tot_pending = 0;
+ u64 tot_granted = 0;
+ u64 fo_tot_granted;
+ u64 fo_tot_pending;
+ u64 fo_tot_dirty;
+
+ if (list_empty(&obd->obd_exports))
return;
/* We don't want to do this for large machines that do lots of
spin_lock(&obd->obd_dev_lock);
spin_lock(&ofd->ofd_grant_lock);
- cfs_list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
+ list_for_each_entry(exp, &obd->obd_exports, exp_obd_chain) {
struct filter_export_data *fed;
int error = 0;
tot_pending += fed->fed_pending;
tot_dirty += fed->fed_dirty;
}
+
+ /* exports about to be unlinked should also be taken into account since
+ * they might still hold pending grant space to be released at
+ * commit time */
+ list_for_each_entry(exp, &obd->obd_unlinked_exports, exp_obd_chain) {
+ struct filter_export_data *fed;
+ int error = 0;
+
+ fed = &exp->exp_filter_data;
+
+ if (fed->fed_grant < 0 || fed->fed_pending < 0 ||
+ fed->fed_dirty < 0)
+ error = 1;
+ if (fed->fed_grant + fed->fed_pending > maxsize) {
+ CERROR("%s: cli %s/%p fed_grant(%ld) + fed_pending(%ld)"
+ " > maxsize("LPU64")\n", obd->obd_name,
+ exp->exp_client_uuid.uuid, exp, fed->fed_grant,
+ fed->fed_pending, maxsize);
+ spin_unlock(&obd->obd_dev_lock);
+ spin_unlock(&ofd->ofd_grant_lock);
+ LBUG();
+ }
+ if (fed->fed_dirty > maxsize) {
+ CERROR("%s: cli %s/%p fed_dirty(%ld) > maxsize("LPU64
+ ")\n", obd->obd_name, exp->exp_client_uuid.uuid,
+ exp, fed->fed_dirty, maxsize);
+ spin_unlock(&obd->obd_dev_lock);
+ spin_unlock(&ofd->ofd_grant_lock);
+ LBUG();
+ }
+ CDEBUG_LIMIT(error ? D_ERROR : D_CACHE, "%s: cli %s/%p dirty "
+ "%ld pend %ld grant %ld\n", obd->obd_name,
+ exp->exp_client_uuid.uuid, exp, fed->fed_dirty,
+ fed->fed_pending, fed->fed_grant);
+ tot_granted += fed->fed_grant + fed->fed_pending;
+ tot_pending += fed->fed_pending;
+ tot_dirty += fed->fed_dirty;
+ }
+
spin_unlock(&obd->obd_dev_lock);
fo_tot_granted = ofd->ofd_tot_granted;
fo_tot_pending = ofd->ofd_tot_pending;
* of available space is requested
* \retval amount of non-allocated space, in bytes
*/
-static obd_size ofd_grant_space_left(struct obd_export *exp)
+static u64 ofd_grant_space_left(struct obd_export *exp)
{
- struct obd_device *obd = exp->exp_obd;
- struct ofd_device *ofd = ofd_exp(exp);
- obd_size tot_granted;
- obd_size left;
- obd_size avail;
- obd_size unstable;
+ struct obd_device *obd = exp->exp_obd;
+ struct ofd_device *ofd = ofd_exp(exp);
+ u64 tot_granted;
+ u64 left;
+ u64 avail;
+ u64 unstable;
ENTRY;
assert_spin_locked(&ofd->ofd_grant_lock);
* overhead estimate made by the OSD layer. If we grant all the free
* space, we have no way (grant space cannot be revoked yet) to
* adjust if the write overhead has been underestimated. */
- left -= min_t(obd_size, left, ofd_grant_reserved(ofd, avail));
+ left -= min_t(u64, left, ofd_grant_reserved(ofd, avail));
/* Align left on block size */
left &= ~((1ULL << ofd->ofd_blockbits) - 1);
oa->o_dirty = 0;
dirty = ofd_grant_from_cli(exp, ofd, oa->o_dirty);
- dropped = ofd_grant_from_cli(exp, ofd, (obd_size)oa->o_dropped);
+ dropped = ofd_grant_from_cli(exp, ofd, (u64)oa->o_dropped);
grant_chunk = ofd_grant_chunk(exp, ofd);
/* Update our accounting now so that statfs takes it into account.
* taken out
*/
static void ofd_grant_shrink(struct obd_export *exp, struct obdo *oa,
- obd_size left_space)
+ u64 left_space)
{
struct filter_export_data *fed;
struct ofd_device *ofd = ofd_exp(exp);
* \retval space (in bytes) that will be consumed to write the
* network buffer
*/
-static inline int ofd_grant_rnb_size(struct obd_export *exp,
+static inline u64 ofd_grant_rnb_size(struct obd_export *exp,
struct ofd_device *ofd,
struct niobuf_remote *rnb)
{
- obd_size blocksize;
- obd_size bytes;
- obd_size end;
+ u64 blocksize;
+ u64 bytes;
+ u64 end;
if (exp && ofd_grant_compat(exp, ofd))
blocksize = 1ULL << COMPAT_BSIZE_SHIFT;
bytes += blocksize - end;
if (exp)
/* Apply per-export pecularities if one is given */
- bytes = ofd_grant_from_cli(exp, ofd, (obd_size)bytes);
+ bytes = ofd_grant_from_cli(exp, ofd, bytes);
return bytes;
}
* \param[in] oa incoming obdo in which we should return the pack the
* additional grant
* \param[in,out] rnb the list of network buffers
- * \param[in] niocont the number of network buffers in the list
+ * \param[in] niocount the number of network buffers in the list
* \param[in] left the remaining free space with space already granted
* taken out
*/
static void ofd_grant_check(const struct lu_env *env, struct obd_export *exp,
struct obdo *oa, struct niobuf_remote *rnb,
- int niocount, obd_size *left)
+ int niocount, u64 *left)
{
struct filter_export_data *fed = &exp->exp_filter_data;
struct obd_device *obd = exp->exp_obd;
exp->exp_client_uuid.uuid, exp, i, bytes);
}
- /* record space used for the I/O, will be used in ofd_grant_commmit() */
+ /* record in o_grant_used the actual space reserved for the I/O, will be
+ * used later in ofd_grant_commmit() */
+ oa->o_grant_used = granted + ungranted;
+
/* Now substract what the clients has used already. We don't subtract
* this from the tot_granted yet, so that other client's can't grab
* that space before we have actually allocated our blocks. That
info->fti_used = granted + ungranted;
*left -= ungranted;
fed->fed_grant -= granted;
- fed->fed_pending += info->fti_used;
+ fed->fed_pending += oa->o_grant_used;
ofd->ofd_tot_granted += ungranted;
- ofd->ofd_tot_pending += info->fti_used;
+ ofd->ofd_tot_pending += oa->o_grant_used;
CDEBUG(D_CACHE,
"%s: cli %s/%p granted: %lu ungranted: %lu grant: %lu dirty: %lu"
*
* \retval amount of grant space allocated
*/
-static long ofd_grant_alloc(struct obd_export *exp, obd_size curgrant,
- obd_size want, obd_size left, bool conservative)
+static long ofd_grant_alloc(struct obd_export *exp, u64 curgrant,
+ u64 want, u64 left, bool conservative)
{
struct obd_device *obd = exp->exp_obd;
struct ofd_device *ofd = ofd_exp(exp);
struct filter_export_data *fed = &exp->exp_filter_data;
long grant_chunk;
- obd_size grant;
+ u64 grant;
ENTRY;
/* don't grant more than 1/8th of the remaining free space in
* one chunk */
left >>= 3;
- grant = min(want, left);
+ grant = min(want - curgrant, left);
/* round grant upt to the next block size */
grant = (grant + (1 << ofd->ofd_blockbits) - 1) &
~((1ULL << ofd->ofd_blockbits) - 1);
* \retval amount of grant space currently owned by the client
*/
long ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
- obd_size want, bool new_conn)
+ u64 want, bool new_conn)
{
struct ofd_device *ofd = ofd_exp(exp);
struct filter_export_data *fed = &exp->exp_filter_data;
- obd_size left = 0;
+ u64 left = 0;
long grant;
int from_cache;
int force = 0; /* can use cached data */
}
ofd_grant_alloc(exp,
- ofd_grant_to_cli(exp, ofd, (obd_size)fed->fed_grant),
+ ofd_grant_to_cli(exp, ofd, (u64)fed->fed_grant),
want, left, new_conn);
/* return to client its current grant */
- grant = ofd_grant_to_cli(exp, ofd, (obd_size)fed->fed_grant);
+ grant = ofd_grant_to_cli(exp, ofd, (u64)fed->fed_grant);
ofd->ofd_tot_granted_clients++;
spin_unlock(&ofd->ofd_grant_lock);
obd->obd_name, ofd->ofd_tot_pending,
exp->exp_client_uuid.uuid, exp, fed->fed_pending);
/* ofd_tot_pending is handled in ofd_grant_commit as bulk
- * finishes */
+ * commmits */
LASSERTF(ofd->ofd_tot_dirty >= fed->fed_dirty,
"%s: tot_dirty "LPU64" cli %s/%p fed_dirty %ld\n",
obd->obd_name, ofd->ofd_tot_dirty,
*
* \param[in] env is the lu environment provided by the caller
* \param[in] exp is the export of the client which sent the request
- * \paral[in,out] oa is the incoming obdo sent by the client
+ * \param[in,out] oa is the incoming obdo sent by the client
*/
void ofd_grant_prepare_read(const struct lu_env *env,
struct obd_export *exp, struct obdo *oa)
{
struct ofd_device *ofd = ofd_exp(exp);
int do_shrink;
- obd_size left = 0;
+ u64 left = 0;
if (!oa)
return;
* \param[in] exp export of the client which sent the request
* \param[in] oa incoming obdo sent by the client
* \param[in] rnb list of network buffers
- * \param[in] niocont number of network buffers in the list
+ * \param[in] niocount number of network buffers in the list
*/
void ofd_grant_prepare_write(const struct lu_env *env,
struct obd_export *exp, struct obdo *oa,
{
struct obd_device *obd = exp->exp_obd;
struct ofd_device *ofd = ofd_exp(exp);
- obd_size left;
+ u64 left;
int from_cache;
int force = 0; /* can use cached data intially */
int rc;
* export currently)
* \param[in] nr number of objects to be created
*
- * \retval 0 for success
+ * \retval >= 0 amount of grant space allocated to the precreate request
* \retval -ENOSPC on failure
*/
-int ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr)
+long ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr)
{
- struct ofd_thread_info *info = ofd_info(env);
struct ofd_device *ofd = ofd_exp(exp);
struct filter_export_data *fed = &exp->exp_filter_data;
- obd_size left = 0;
+ u64 left = 0;
unsigned long wanted;
+ unsigned long granted;
ENTRY;
- info->fti_used = 0;
-
if (exp->exp_obd->obd_recovering ||
ofd->ofd_dt_conf.ddp_inodespace == 0)
/* don't enforce grant during recovery */
left -= wanted - fed->fed_grant;
fed->fed_grant = 0;
}
- info->fti_used = wanted;
- fed->fed_pending += info->fti_used;
- ofd->ofd_tot_pending += info->fti_used;
+ granted = wanted;
+ fed->fed_pending += granted;
+ ofd->ofd_tot_pending += granted;
/* grant more space for precreate purpose if possible. */
wanted = OST_MAX_PRECREATE * ofd->ofd_dt_conf.ddp_inodespace / 2;
ofd_grant_alloc(exp, fed->fed_grant, wanted, left, false);
}
spin_unlock(&ofd->ofd_grant_lock);
- RETURN(0);
+ RETURN(granted);
}
/**
*
* Update pending grant counter once buffers have been written to the disk.
*
- * \param[in] env LU environment provided by the caller
* \param[in] exp export of the client which sent the request
+ * \param[in] pending amount of reserved space to be released
+ * \param[in] rc return code of pre-commit operations
*/
-void ofd_grant_commit(const struct lu_env *env, struct obd_export *exp,
+void ofd_grant_commit(struct obd_export *exp, unsigned long pending,
int rc)
{
struct ofd_device *ofd = ofd_exp(exp);
- struct ofd_thread_info *info = ofd_info(env);
- unsigned long pending;
-
ENTRY;
/* get space accounted in tot_pending for the I/O, set in
* ofd_grant_check() */
- pending = info->fti_used;
if (pending == 0)
RETURN_EXIT;
if (rc == 0) {
spin_lock(&ofd->ofd_osfs_lock);
/* Take pending out of cached statfs data */
- ofd->ofd_osfs.os_bavail -= min_t(obd_size,
+ ofd->ofd_osfs.os_bavail -= min_t(u64,
ofd->ofd_osfs.os_bavail,
pending >> ofd->ofd_blockbits);
if (ofd->ofd_statfs_inflight)
spin_unlock(&ofd->ofd_grant_lock);
EXIT;
}
+
+struct ofd_grant_cb {
+ /* commit callback structure */
+ struct dt_txn_commit_cb ogc_cb;
+ /* export associated with the bulk write */
+ struct obd_export *ogc_exp;
+ /* pending grant to be released */
+ unsigned long ogc_granted;
+};
+
+/**
+ * Callback function for grant releasing
+ *
+ * Release grant space reserved by the client node.
+ *
+ * \param[in] env execution environment
+ * \param[in] th transaction handle
+ * \param[in] cb callback data
+ * \param[in] err error code
+ */
+static void ofd_grant_commit_cb(struct lu_env *env, struct thandle *th,
+ struct dt_txn_commit_cb *cb, int err)
+{
+ struct ofd_grant_cb *ogc;
+
+ ogc = container_of(cb, struct ofd_grant_cb, ogc_cb);
+
+ ofd_grant_commit(ogc->ogc_exp, ogc->ogc_granted, err);
+ class_export_cb_put(ogc->ogc_exp);
+ OBD_FREE_PTR(ogc);
+}
+
+/**
+ * Add callback for grant releasing
+ *
+ * Register a commit callback to release grant space.
+ *
+ * \param[in] th transaction handle
+ * \param[in] exp OBD export of client
+ * \param[in] granted amount of grant space to be released upon commit
+ *
+ * \retval 0 on successful callback adding
+ * \retval negative value on error
+ */
+int ofd_grant_commit_cb_add(struct thandle *th, struct obd_export *exp,
+ unsigned long granted)
+{
+ struct ofd_grant_cb *ogc;
+ struct dt_txn_commit_cb *dcb;
+ int rc;
+ ENTRY;
+
+ OBD_ALLOC_PTR(ogc);
+ if (ogc == NULL)
+ RETURN(-ENOMEM);
+
+ ogc->ogc_exp = class_export_cb_get(exp);
+ ogc->ogc_granted = granted;
+
+ dcb = &ogc->ogc_cb;
+ dcb->dcb_func = ofd_grant_commit_cb;
+ INIT_LIST_HEAD(&dcb->dcb_linkage);
+ strlcpy(dcb->dcb_name, "ofd_grant_commit_cb", sizeof(dcb->dcb_name));
+
+ rc = dt_trans_cb_add(th, dcb);
+ if (rc) {
+ class_export_cb_put(ogc->ogc_exp);
+ OBD_FREE_PTR(ogc);
+ }
+
+ RETURN(rc);
+}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff