*
* You should have received a copy of the GNU General Public License
* version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
*
* GPL HEADER END
*/
* Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2013, Intel Corporation.
+ * Copyright (c) 2011, 2016, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#include "ptlrpc_internal.h"
+const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
+ .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
+ .release_frags = ptlrpc_release_bulk_page_pin,
+};
+EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
+
+const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
+ .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
+ .release_frags = ptlrpc_release_bulk_noop,
+};
+EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
+
+const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
+ .add_iov_frag = ptlrpc_prep_bulk_frag,
+};
+EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
+
static int ptlrpc_send_new_req(struct ptlrpc_request *req);
static int ptlrpcd_check_work(struct ptlrpc_request *req);
+static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
/**
* Initialize passed in client structure \a cl.
/**
* Return PortalRPC connection for remore uud \a uuid
*/
-struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
+struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
+ lnet_nid_t nid4refnet)
{
- struct ptlrpc_connection *c;
- lnet_nid_t self;
- lnet_process_id_t peer;
- int err;
+ struct ptlrpc_connection *c;
+ lnet_nid_t self;
+ struct lnet_process_id peer;
+ int err;
/* ptlrpc_uuid_to_peer() initializes its 2nd parameter
* before accessing its values. */
/* coverity[uninit_use_in_call] */
- err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
- if (err != 0) {
- CNETERR("cannot find peer %s!\n", uuid->uuid);
- return NULL;
- }
+ peer.nid = nid4refnet;
+ err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
+ if (err != 0) {
+ CNETERR("cannot find peer %s!\n", uuid->uuid);
+ return NULL;
+ }
- c = ptlrpc_connection_get(peer, self, uuid);
- if (c) {
- memcpy(c->c_remote_uuid.uuid,
- uuid->uuid, sizeof(c->c_remote_uuid.uuid));
- }
+ c = ptlrpc_connection_get(peer, self, uuid);
+ if (c) {
+ memcpy(c->c_remote_uuid.uuid,
+ uuid->uuid, sizeof(c->c_remote_uuid.uuid));
+ }
- CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
+ CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
- return c;
+ return c;
}
-EXPORT_SYMBOL(ptlrpc_uuid_to_connection);
/**
* Allocate and initialize new bulk descriptor on the sender.
* Returns pointer to the descriptor or NULL on error.
*/
-struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned npages, unsigned max_brw,
- unsigned type, unsigned portal)
+struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
+ enum ptlrpc_bulk_op_type type,
+ unsigned portal,
+ const struct ptlrpc_bulk_frag_ops *ops)
{
struct ptlrpc_bulk_desc *desc;
int i;
- OBD_ALLOC(desc, offsetof(struct ptlrpc_bulk_desc, bd_iov[npages]));
- if (!desc)
+ /* ensure that only one of KIOV or IOVEC is set but not both */
+ LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
+ ops->add_kiov_frag != NULL) ||
+ (ptlrpc_is_bulk_desc_kvec(type) &&
+ ops->add_iov_frag != NULL));
+
+ OBD_ALLOC_PTR(desc);
+ if (desc == NULL)
return NULL;
+ if (type & PTLRPC_BULK_BUF_KIOV) {
+ OBD_ALLOC_LARGE(GET_KIOV(desc),
+ nfrags * sizeof(*GET_KIOV(desc)));
+ if (GET_KIOV(desc) == NULL)
+ goto out;
+ } else {
+ OBD_ALLOC_LARGE(GET_KVEC(desc),
+ nfrags * sizeof(*GET_KVEC(desc)));
+ if (GET_KVEC(desc) == NULL)
+ goto out;
+ }
spin_lock_init(&desc->bd_lock);
init_waitqueue_head(&desc->bd_waitq);
- desc->bd_max_iov = npages;
+ desc->bd_max_iov = nfrags;
desc->bd_iov_count = 0;
desc->bd_portal = portal;
desc->bd_type = type;
desc->bd_md_count = 0;
+ desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
LASSERT(max_brw > 0);
desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
/* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
* node. Negotiated ocd_brw_size will always be <= this number. */
for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
- LNetInvalidateHandle(&desc->bd_mds[i]);
+ LNetInvalidateMDHandle(&desc->bd_mds[i]);
return desc;
+out:
+ OBD_FREE_PTR(desc);
+ return NULL;
}
/**
* Prepare bulk descriptor for specified outgoing request \a req that
- * can fit \a npages * pages. \a type is bulk type. \a portal is where
+ * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
* the bulk to be sent. Used on client-side.
* Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
* error.
*/
struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
- unsigned npages, unsigned max_brw,
- unsigned type, unsigned portal)
+ unsigned nfrags, unsigned max_brw,
+ unsigned int type,
+ unsigned portal,
+ const struct ptlrpc_bulk_frag_ops
+ *ops)
{
struct obd_import *imp = req->rq_import;
struct ptlrpc_bulk_desc *desc;
ENTRY;
- LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
- desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
+ LASSERT(ptlrpc_is_bulk_op_passive(type));
+
+ desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
if (desc == NULL)
RETURN(NULL);
}
EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
-/*
- * Add a page \a page to the bulk descriptor \a desc.
- * Data to transfer in the page starts at offset \a pageoffset and
- * amount of data to transfer from the page is \a len
- */
void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
- struct page *page, int pageoffset, int len, int pin)
+ struct page *page, int pageoffset, int len,
+ int pin)
{
+ lnet_kiov_t *kiov;
+
LASSERT(desc->bd_iov_count < desc->bd_max_iov);
LASSERT(page != NULL);
LASSERT(pageoffset >= 0);
LASSERT(len > 0);
- LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
+ LASSERT(pageoffset + len <= PAGE_SIZE);
+ LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
+
+ kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
desc->bd_nob += len;
if (pin)
- page_cache_get(page);
+ get_page(page);
- ptlrpc_add_bulk_page(desc, page, pageoffset, len);
+ kiov->kiov_page = page;
+ kiov->kiov_offset = pageoffset;
+ kiov->kiov_len = len;
+
+ desc->bd_iov_count++;
}
EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
-/**
- * Uninitialize and free bulk descriptor \a desc.
- * Works on bulk descriptors both from server and client side.
- */
-void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc, int unpin)
+int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
+ void *frag, int len)
+{
+ struct kvec *iovec;
+ ENTRY;
+
+ LASSERT(desc->bd_iov_count < desc->bd_max_iov);
+ LASSERT(frag != NULL);
+ LASSERT(len > 0);
+ LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
+
+ iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
+
+ desc->bd_nob += len;
+
+ iovec->iov_base = frag;
+ iovec->iov_len = len;
+
+ desc->bd_iov_count++;
+
+ RETURN(desc->bd_nob);
+}
+EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
+
+void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
{
- int i;
ENTRY;
LASSERT(desc != NULL);
LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
LASSERT(desc->bd_md_count == 0); /* network hands off */
LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
+ LASSERT(desc->bd_frag_ops != NULL);
- sptlrpc_enc_pool_put_pages(desc);
+ if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
+ sptlrpc_enc_pool_put_pages(desc);
if (desc->bd_export)
class_export_put(desc->bd_export);
else
class_import_put(desc->bd_import);
- if (unpin) {
- for (i = 0; i < desc->bd_iov_count ; i++)
- page_cache_release(desc->bd_iov[i].kiov_page);
- }
+ if (desc->bd_frag_ops->release_frags != NULL)
+ desc->bd_frag_ops->release_frags(desc);
- OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
- bd_iov[desc->bd_max_iov]));
+ if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
+ OBD_FREE_LARGE(GET_KIOV(desc),
+ desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
+ else
+ OBD_FREE_LARGE(GET_KVEC(desc),
+ desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
+ OBD_FREE_PTR(desc);
EXIT;
}
-EXPORT_SYMBOL(__ptlrpc_free_bulk);
+EXPORT_SYMBOL(ptlrpc_free_bulk);
/**
* Set server timelimit for this req, i.e. how long are we willing to wait
{
unsigned int nl, oldnl;
struct imp_at *at;
- time_t now = cfs_time_current_sec();
+ time64_t now = ktime_get_real_seconds();
LASSERT(req->rq_import);
*/
CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
D_ADAPTTO : D_WARNING,
- "Reported service time %u > total measured time "
- CFS_DURATION_T"\n", service_time,
- cfs_time_sub(now, req->rq_sent));
+ "Reported service time %u > total measured time %lld\n",
+ service_time, now - req->rq_sent);
return;
}
static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
__must_hold(&req->rq_lock)
{
- struct ptlrpc_request *early_req;
- time_t olddl;
- int rc;
- ENTRY;
+ struct ptlrpc_request *early_req;
+ time64_t olddl;
+ int rc;
+ ENTRY;
req->rq_early = 0;
spin_unlock(&req->rq_lock);
rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
if (rc) {
spin_lock(&req->rq_lock);
- RETURN(rc);
- }
-
- rc = unpack_reply(early_req);
- if (rc == 0) {
- /* Expecting to increase the service time estimate here */
- ptlrpc_at_adj_service(req,
- lustre_msg_get_timeout(early_req->rq_repmsg));
- ptlrpc_at_adj_net_latency(req,
- lustre_msg_get_service_time(early_req->rq_repmsg));
- }
-
- sptlrpc_cli_finish_early_reply(early_req);
+ RETURN(rc);
+ }
+ rc = unpack_reply(early_req);
if (rc != 0) {
+ sptlrpc_cli_finish_early_reply(early_req);
spin_lock(&req->rq_lock);
RETURN(rc);
}
- /* Adjust the local timeout for this req */
- ptlrpc_at_set_req_timeout(req);
+ /* Use new timeout value just to adjust the local value for this
+ * request, don't include it into at_history. It is unclear yet why
+ * service time increased and should it be counted or skipped, e.g.
+ * that can be recovery case or some error or server, the real reply
+ * will add all new data if it is worth to add. */
+ req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
+ lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
+
+ /* Network latency can be adjusted, it is pure network delays */
+ ptlrpc_at_adj_net_latency(req,
+ lustre_msg_get_service_time(early_req->rq_repmsg));
+
+ sptlrpc_cli_finish_early_reply(early_req);
spin_lock(&req->rq_lock);
olddl = req->rq_deadline;
ptlrpc_at_get_net_latency(req);
DEBUG_REQ(D_ADAPTTO, req,
- "Early reply #%d, new deadline in "CFS_DURATION_T"s "
- "("CFS_DURATION_T"s)", req->rq_early_count,
- cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
- cfs_time_sub(req->rq_deadline, olddl));
+ "Early reply #%d, new deadline in %llds (%llds)",
+ req->rq_early_count,
+ req->rq_deadline - ktime_get_real_seconds(),
+ req->rq_deadline - olddl);
RETURN(rc);
}
-struct kmem_cache *request_cache;
+static struct kmem_cache *request_cache;
int ptlrpc_request_cache_init(void)
{
/**
* Allocates, initializes and adds \a num_rq requests to the pool \a pool
*/
-void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
+int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
{
int i;
int size = 1;
spin_unlock(&pool->prp_lock);
req = ptlrpc_request_cache_alloc(GFP_NOFS);
if (!req)
- return;
+ return i;
OBD_ALLOC_LARGE(msg, size);
if (!msg) {
ptlrpc_request_cache_free(req);
- return;
- }
- req->rq_reqbuf = msg;
- req->rq_reqbuf_len = size;
- req->rq_pool = pool;
+ return i;
+ }
+ req->rq_reqbuf = msg;
+ req->rq_reqbuf_len = size;
+ req->rq_pool = pool;
spin_lock(&pool->prp_lock);
list_add_tail(&req->rq_list, &pool->prp_req_list);
}
spin_unlock(&pool->prp_lock);
- return;
+ return num_rq;
}
EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
*/
struct ptlrpc_request_pool *
ptlrpc_init_rq_pool(int num_rq, int msgsize,
- void (*populate_pool)(struct ptlrpc_request_pool *, int))
+ int (*populate_pool)(struct ptlrpc_request_pool *, int))
{
struct ptlrpc_request_pool *pool;
populate_pool(pool, num_rq);
- if (list_empty(&pool->prp_req_list)) {
- /* have not allocated a single request for the pool */
- OBD_FREE(pool, sizeof(struct ptlrpc_request_pool));
- pool = NULL;
- }
return pool;
}
EXPORT_SYMBOL(ptlrpc_init_rq_pool);
spin_unlock(&pool->prp_lock);
}
-static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
- __u32 version, int opcode,
- int count, __u32 *lengths, char **bufs,
- struct ptlrpc_cli_ctx *ctx)
+void ptlrpc_add_unreplied(struct ptlrpc_request *req)
{
- struct obd_import *imp = request->rq_import;
- int rc;
- ENTRY;
+ struct obd_import *imp = req->rq_import;
+ struct list_head *tmp;
+ struct ptlrpc_request *iter;
- if (unlikely(ctx))
- request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
- else {
- rc = sptlrpc_req_get_ctx(request);
- if (rc)
- GOTO(out_free, rc);
- }
+ assert_spin_locked(&imp->imp_lock);
+ LASSERT(list_empty(&req->rq_unreplied_list));
- sptlrpc_req_set_flavor(request, opcode);
+ /* unreplied list is sorted by xid in ascending order */
+ list_for_each_prev(tmp, &imp->imp_unreplied_list) {
+ iter = list_entry(tmp, struct ptlrpc_request,
+ rq_unreplied_list);
- rc = lustre_pack_request(request, imp->imp_msg_magic, count,
- lengths, bufs);
- if (rc) {
- LASSERT(!request->rq_pool);
- GOTO(out_ctx, rc);
- }
+ LASSERT(req->rq_xid != iter->rq_xid);
+ if (req->rq_xid < iter->rq_xid)
+ continue;
+ list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
+ return;
+ }
+ list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
+}
- lustre_msg_add_version(request->rq_reqmsg, version);
- request->rq_send_state = LUSTRE_IMP_FULL;
- request->rq_type = PTL_RPC_MSG_REQUEST;
+void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
+{
+ req->rq_xid = ptlrpc_next_xid();
+ ptlrpc_add_unreplied(req);
+}
- request->rq_req_cbid.cbid_fn = request_out_callback;
- request->rq_req_cbid.cbid_arg = request;
+static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
+{
+ spin_lock(&req->rq_import->imp_lock);
+ ptlrpc_assign_next_xid_nolock(req);
+ spin_unlock(&req->rq_import->imp_lock);
+}
+
+int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
+ __u32 version, int opcode, char **bufs,
+ struct ptlrpc_cli_ctx *ctx)
+{
+ int count;
+ struct obd_import *imp;
+ __u32 *lengths;
+ int rc;
+
+ ENTRY;
+
+ count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
+ imp = request->rq_import;
+ lengths = request->rq_pill.rc_area[RCL_CLIENT];
+
+ if (ctx != NULL) {
+ request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
+ } else {
+ rc = sptlrpc_req_get_ctx(request);
+ if (rc)
+ GOTO(out_free, rc);
+ }
+ sptlrpc_req_set_flavor(request, opcode);
+
+ rc = lustre_pack_request(request, imp->imp_msg_magic, count,
+ lengths, bufs);
+ if (rc)
+ GOTO(out_ctx, rc);
+
+ lustre_msg_add_version(request->rq_reqmsg, version);
+ request->rq_send_state = LUSTRE_IMP_FULL;
+ request->rq_type = PTL_RPC_MSG_REQUEST;
- request->rq_reply_cbid.cbid_fn = reply_in_callback;
- request->rq_reply_cbid.cbid_arg = request;
+ request->rq_req_cbid.cbid_fn = request_out_callback;
+ request->rq_req_cbid.cbid_arg = request;
- request->rq_reply_deadline = 0;
- request->rq_phase = RQ_PHASE_NEW;
- request->rq_next_phase = RQ_PHASE_UNDEFINED;
+ request->rq_reply_cbid.cbid_fn = reply_in_callback;
+ request->rq_reply_cbid.cbid_arg = request;
- request->rq_request_portal = imp->imp_client->cli_request_portal;
- request->rq_reply_portal = imp->imp_client->cli_reply_portal;
+ request->rq_reply_deadline = 0;
+ request->rq_bulk_deadline = 0;
+ request->rq_req_deadline = 0;
+ request->rq_phase = RQ_PHASE_NEW;
+ request->rq_next_phase = RQ_PHASE_UNDEFINED;
- ptlrpc_at_set_req_timeout(request);
+ request->rq_request_portal = imp->imp_client->cli_request_portal;
+ request->rq_reply_portal = imp->imp_client->cli_reply_portal;
+
+ ptlrpc_at_set_req_timeout(request);
- request->rq_xid = ptlrpc_next_xid();
lustre_msg_set_opc(request->rq_reqmsg, opcode);
+ ptlrpc_assign_next_xid(request);
+
+ /* Let's setup deadline for req/reply/bulk unlink for opcode. */
+ if (cfs_fail_val == opcode) {
+ time64_t *fail_t = NULL, *fail2_t = NULL;
+
+ if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
+ fail_t = &request->rq_bulk_deadline;
+ else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
+ fail_t = &request->rq_reply_deadline;
+ else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
+ fail_t = &request->rq_req_deadline;
+ else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
+ fail_t = &request->rq_reply_deadline;
+ fail2_t = &request->rq_bulk_deadline;
+ }
+
+ if (fail_t) {
+ *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
+
+ if (fail2_t)
+ *fail2_t = ktime_get_real_seconds() +
+ LONG_UNLINK;
+
+ /*
+ * The RPC is infected, let the test to change the
+ * fail_loc
+ */
+ msleep(4 * MSEC_PER_SEC);
+ }
+ }
RETURN(0);
+
out_ctx:
+ LASSERT(!request->rq_pool);
sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
out_free:
class_import_put(imp);
- return rc;
-}
-int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
- __u32 version, int opcode, char **bufs,
- struct ptlrpc_cli_ctx *ctx)
-{
- int count;
+ return rc;
- count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
- return __ptlrpc_request_bufs_pack(request, version, opcode, count,
- request->rq_pill.rc_area[RCL_CLIENT],
- bufs, ctx);
}
EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
{
struct ptlrpc_request *request = NULL;
- if (pool)
- request = ptlrpc_prep_req_from_pool(pool);
+ request = ptlrpc_request_cache_alloc(GFP_NOFS);
- if (!request)
- request = ptlrpc_request_cache_alloc(GFP_NOFS);
+ if (!request && pool)
+ request = ptlrpc_prep_req_from_pool(pool);
if (request) {
ptlrpc_cli_req_init(request);
EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
/**
- * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
- * for operation \a opcode. Request would contain \a count buffers.
- * Sizes of buffers are described in array \a lengths and buffers themselves
- * are provided by a pointer \a bufs.
- * Returns prepared request structure pointer or NULL on error.
- */
-struct ptlrpc_request *
-ptlrpc_prep_req_pool(struct obd_import *imp,
- __u32 version, int opcode,
- int count, __u32 *lengths, char **bufs,
- struct ptlrpc_request_pool *pool)
-{
- struct ptlrpc_request *request;
- int rc;
-
- request = __ptlrpc_request_alloc(imp, pool);
- if (!request)
- return NULL;
-
- rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
- lengths, bufs, NULL);
- if (rc) {
- ptlrpc_request_free(request);
- request = NULL;
- }
- return request;
-}
-EXPORT_SYMBOL(ptlrpc_prep_req_pool);
-
-/**
- * Same as ptlrpc_prep_req_pool, but without pool
- */
-struct ptlrpc_request *
-ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
- __u32 *lengths, char **bufs)
-{
- return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
- NULL);
-}
-EXPORT_SYMBOL(ptlrpc_prep_req);
-
-/**
- * Allocate and initialize new request set structure.
+ * Allocate and initialize new request set structure on the current CPT.
* Returns a pointer to the newly allocated set structure or NULL on error.
*/
struct ptlrpc_request_set *ptlrpc_prep_set(void)
{
- struct ptlrpc_request_set *set;
+ struct ptlrpc_request_set *set;
+ int cpt;
ENTRY;
- OBD_ALLOC(set, sizeof *set);
+ cpt = cfs_cpt_current(cfs_cpt_table, 0);
+ OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
if (!set)
RETURN(NULL);
atomic_set(&set->set_refcount, 1);
RETURN(set);
}
-EXPORT_SYMBOL(ptlrpc_prep_fcset);
/**
* Wind down and free request set structure previously allocated with
RETURN(0);
}
-EXPORT_SYMBOL(ptlrpc_set_add_cb);
/**
* Add a new request to the general purpose request set.
{
LASSERT(list_empty(&req->rq_set_chain));
+ if (req->rq_allow_intr)
+ set->set_allow_intr = 1;
+
/* The set takes over the caller's request reference */
list_add_tail(&req->rq_set_chain, &set->set_requests);
req->rq_set = set;
wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
}
}
-EXPORT_SYMBOL(ptlrpc_set_add_new_req);
/**
* Based on the current state of the import, determine if the request
if (atomic_read(&imp->imp_inval_count) != 0) {
DEBUG_REQ(D_ERROR, req, "invalidate in flight");
*status = -EIO;
- } else if (imp->imp_dlm_fake || req->rq_no_delay) {
+ } else if (req->rq_no_delay) {
*status = -EWOULDBLOCK;
} else if (req->rq_allow_replay &&
(imp->imp_state == LUSTRE_IMP_REPLAY ||
}
/**
- * Decide if the eror message regarding provided request \a req
- * should be printed to the console or not.
- * Makes it's decision on request status and other properties.
- * Returns 1 to print error on the system console or 0 if not.
+ * Decide if the error message should be printed to the console or not.
+ * Makes its decision based on request type, status, and failure frequency.
+ *
+ * \param[in] req request that failed and may need a console message
+ *
+ * \retval false if no message should be printed
+ * \retval true if console message should be printed
*/
-static int ptlrpc_console_allow(struct ptlrpc_request *req)
+static bool ptlrpc_console_allow(struct ptlrpc_request *req)
{
- __u32 opc;
- int err;
+ __u32 opc;
- LASSERT(req->rq_reqmsg != NULL);
- opc = lustre_msg_get_opc(req->rq_reqmsg);
+ LASSERT(req->rq_reqmsg != NULL);
+ opc = lustre_msg_get_opc(req->rq_reqmsg);
- /* Suppress particular reconnect errors which are to be expected. No
- * errors are suppressed for the initial connection on an import */
- if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
- (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
+ /* Suppress particular reconnect errors which are to be expected. */
+ if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
+ int err;
- /* Suppress timed out reconnect requests */
- if (req->rq_timedout)
- return 0;
+ /* Suppress timed out reconnect requests */
+ if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
+ req->rq_timedout)
+ return false;
- /* Suppress unavailable/again reconnect requests */
- err = lustre_msg_get_status(req->rq_repmsg);
- if (err == -ENODEV || err == -EAGAIN)
- return 0;
- }
+ /* Suppress most unavailable/again reconnect requests, but
+ * print occasionally so it is clear client is trying to
+ * connect to a server where no target is running. */
+ err = lustre_msg_get_status(req->rq_repmsg);
+ if ((err == -ENODEV || err == -EAGAIN) &&
+ req->rq_import->imp_conn_cnt % 30 != 20)
+ return false;
+ }
- return 1;
+ return true;
}
/**
err = lustre_msg_get_status(req->rq_repmsg);
if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
struct obd_import *imp = req->rq_import;
+ lnet_nid_t nid = imp->imp_connection->c_peer.nid;
__u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
- if (ptlrpc_console_allow(req))
- LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s,"
- " operation %s failed with %d.\n",
- imp->imp_obd->obd_name,
- libcfs_nid2str(
- imp->imp_connection->c_peer.nid),
- ll_opcode2str(opc), err);
+
+ /* -EAGAIN is normal when using POSIX flocks */
+ if (ptlrpc_console_allow(req) &&
+ !(opc == LDLM_ENQUEUE && err == -EAGAIN))
+ LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
+ "failed: rc = %d\n",
+ imp->imp_obd->obd_name,
+ ll_opcode2str(opc),
+ libcfs_nid2str(nid), err);
RETURN(err < 0 ? err : -EINVAL);
}
LASSERT(versions);
lustre_msg_set_versions(reqmsg, versions);
- CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
+ CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
versions[0], versions[1]);
EXIT;
}
+__u64 ptlrpc_known_replied_xid(struct obd_import *imp)
+{
+ struct ptlrpc_request *req;
+
+ assert_spin_locked(&imp->imp_lock);
+ if (list_empty(&imp->imp_unreplied_list))
+ return 0;
+
+ req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
+ rq_unreplied_list);
+ LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
+
+ if (imp->imp_known_replied_xid < req->rq_xid - 1)
+ imp->imp_known_replied_xid = req->rq_xid - 1;
+
+ return req->rq_xid - 1;
+}
+
/**
* Callback function called when client receives RPC reply for \a req.
* Returns 0 on success or error code.
*/
static int after_reply(struct ptlrpc_request *req)
{
- struct obd_import *imp = req->rq_import;
- struct obd_device *obd = req->rq_import->imp_obd;
- int rc;
- struct timeval work_start;
- long timediff;
- ENTRY;
+ struct obd_import *imp = req->rq_import;
+ struct obd_device *obd = req->rq_import->imp_obd;
+ ktime_t work_start;
+ u64 committed;
+ s64 timediff;
+ int rc;
+ ENTRY;
LASSERT(obd != NULL);
/* repbuf must be unlinked */
- LASSERT(!req->rq_receiving_reply && !req->rq_reply_unlink);
+ LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
- if (req->rq_reply_truncate) {
+ if (req->rq_reply_truncated) {
if (ptlrpc_no_resend(req)) {
DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
" expected: %d, actual size: %d",
RETURN(0);
}
+ work_start = ktime_get_real();
+ timediff = ktime_us_delta(work_start, req->rq_sent_ns);
+
/*
* NB Until this point, the whole of the incoming message,
* including buflens, status etc is in the sender's byte order.
/* retry indefinitely on EINPROGRESS */
if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
- time_t now = cfs_time_current_sec();
+ time64_t now = ktime_get_real_seconds();
DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
spin_lock(&req->rq_lock);
spin_unlock(&req->rq_lock);
req->rq_nr_resend++;
- /* allocate new xid to avoid reply reconstruction */
- if (!req->rq_bulk) {
- /* new xid is already allocated for bulk in
- * ptlrpc_check_set() */
- req->rq_xid = ptlrpc_next_xid();
- DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
- "resend on EINPROGRESS");
- }
-
/* Readjust the timeout for current conditions */
ptlrpc_at_set_req_timeout(req);
/* delay resend to give a chance to the server to get ready.
else
req->rq_sent = now + req->rq_nr_resend;
+ /* Resend for EINPROGRESS will use a new XID */
+ spin_lock(&imp->imp_lock);
+ list_del_init(&req->rq_unreplied_list);
+ spin_unlock(&imp->imp_lock);
+
RETURN(0);
}
- do_gettimeofday(&work_start);
- timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
if (obd->obd_svc_stats != NULL) {
lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
timediff);
rc = ptlrpc_check_status(req);
imp->imp_connect_error = rc;
- if (rc) {
- /*
- * Either we've been evicted, or the server has failed for
- * some reason. Try to reconnect, and if that fails, punt to
- * the upcall.
- */
- if (ll_rpc_recoverable_error(rc)) {
- if (req->rq_send_state != LUSTRE_IMP_FULL ||
- imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
- RETURN(rc);
- }
- ptlrpc_request_handle_notconn(req);
- RETURN(rc);
- }
+ if (rc) {
+ /*
+ * Either we've been evicted, or the server has failed for
+ * some reason. Try to reconnect, and if that fails, punt to
+ * the upcall.
+ */
+ if (ptlrpc_recoverable_error(rc)) {
+ if (req->rq_send_state != LUSTRE_IMP_FULL ||
+ imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
+ RETURN(rc);
+ }
+ ptlrpc_request_handle_notconn(req);
+ RETURN(rc);
+ }
} else {
/*
* Let's look if server sent slv. Do it only for RPC with
/*
* Replay-enabled imports return commit-status information.
*/
- if (lustre_msg_get_last_committed(req->rq_repmsg)) {
- imp->imp_peer_committed_transno =
- lustre_msg_get_last_committed(req->rq_repmsg);
- }
+ committed = lustre_msg_get_last_committed(req->rq_repmsg);
+ if (likely(committed > imp->imp_peer_committed_transno))
+ imp->imp_peer_committed_transno = committed;
ptlrpc_free_committed(imp);
static int ptlrpc_send_new_req(struct ptlrpc_request *req)
{
struct obd_import *imp = req->rq_import;
+ __u64 min_xid = 0;
int rc;
ENTRY;
LASSERT(req->rq_phase == RQ_PHASE_NEW);
- if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
+
+ /* do not try to go further if there is not enough memory in enc_pool */
+ if (req->rq_sent && req->rq_bulk != NULL)
+ if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
+ pool_is_at_full_capacity())
+ RETURN(-ENOMEM);
+
+ if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
(!req->rq_generation_set ||
req->rq_import_generation == imp->imp_generation))
RETURN (0);
spin_lock(&imp->imp_lock);
+ LASSERT(req->rq_xid != 0);
+ LASSERT(!list_empty(&req->rq_unreplied_list));
+
if (!req->rq_generation_set)
req->rq_import_generation = imp->imp_generation;
req->rq_waiting = 1;
spin_unlock(&req->rq_lock);
- DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
- "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
+ DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
ptlrpc_import_state_name(req->rq_send_state),
ptlrpc_import_state_name(imp->imp_state));
LASSERT(list_empty(&req->rq_list));
LASSERT(list_empty(&req->rq_list));
list_add_tail(&req->rq_list, &imp->imp_sending_list);
atomic_inc(&req->rq_import->imp_inflight);
+
+ /* find the known replied XID from the unreplied list, CONNECT
+ * and DISCONNECT requests are skipped to make the sanity check
+ * on server side happy. see process_req_last_xid().
+ *
+ * For CONNECT: Because replay requests have lower XID, it'll
+ * break the sanity check if CONNECT bump the exp_last_xid on
+ * server.
+ *
+ * For DISCONNECT: Since client will abort inflight RPC before
+ * sending DISCONNECT, DISCONNECT may carry an XID which higher
+ * than the inflight RPC.
+ */
+ if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
+ min_xid = ptlrpc_known_replied_xid(imp);
spin_unlock(&imp->imp_lock);
+ lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
+
lustre_msg_set_status(req->rq_reqmsg, current_pid());
rc = sptlrpc_req_refresh_ctx(req, -1);
}
CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
- " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
+ " %s:%s:%d:%llu:%s:%d\n", current_comm(),
imp->imp_obd->obd_uuid.uuid,
lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
libcfs_nid2str(imp->imp_connection->c_peer.nid),
lustre_msg_get_opc(req->rq_reqmsg));
rc = ptl_send_rpc(req, 0);
+ if (rc == -ENOMEM) {
+ spin_lock(&imp->imp_lock);
+ if (!list_empty(&req->rq_list)) {
+ list_del_init(&req->rq_list);
+ atomic_dec(&req->rq_import->imp_inflight);
+ }
+ spin_unlock(&imp->imp_lock);
+ ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
+ RETURN(rc);
+ }
if (rc) {
DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
spin_lock(&req->rq_lock);
rq_set_chain);
struct obd_import *imp = req->rq_import;
int unregistered = 0;
+ int async = 1;
int rc = 0;
+ if (req->rq_phase == RQ_PHASE_COMPLETE) {
+ list_move_tail(&req->rq_set_chain, &comp_reqs);
+ continue;
+ }
+
/* This schedule point is mainly for the ptlrpcd caller of this
* function. Most ptlrpc sets are not long-lived and unbounded
* in length, but at the least the set used by the ptlrpcd is.
*/
cond_resched();
- if (req->rq_phase == RQ_PHASE_NEW &&
- ptlrpc_send_new_req(req)) {
- force_timer_recalc = 1;
- }
+ /* If the caller requires to allow to be interpreted by force
+ * and it has really been interpreted, then move the request
+ * to RQ_PHASE_INTERPRET phase in spite of what the current
+ * phase is. */
+ if (unlikely(req->rq_allow_intr && req->rq_intr)) {
+ req->rq_status = -EINTR;
+ ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
+
+ /* Since it is interpreted and we have to wait for
+ * the reply to be unlinked, then use sync mode. */
+ async = 0;
- /* delayed send - skip */
- if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
+ GOTO(interpret, req->rq_status);
+ }
+
+ if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
+ force_timer_recalc = 1;
+
+ /* delayed send - skip */
+ if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
continue;
/* delayed resend - skip */
if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
- req->rq_sent > cfs_time_current_sec())
+ req->rq_sent > ktime_get_real_seconds())
continue;
- if (!(req->rq_phase == RQ_PHASE_RPC ||
- req->rq_phase == RQ_PHASE_BULK ||
- req->rq_phase == RQ_PHASE_INTERPRET ||
- req->rq_phase == RQ_PHASE_UNREGISTERING ||
- req->rq_phase == RQ_PHASE_COMPLETE)) {
- DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
- LBUG();
- }
+ if (!(req->rq_phase == RQ_PHASE_RPC ||
+ req->rq_phase == RQ_PHASE_BULK ||
+ req->rq_phase == RQ_PHASE_INTERPRET ||
+ req->rq_phase == RQ_PHASE_UNREG_RPC ||
+ req->rq_phase == RQ_PHASE_UNREG_BULK)) {
+ DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
+ LBUG();
+ }
- if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
- LASSERT(req->rq_next_phase != req->rq_phase);
- LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
+ if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
+ req->rq_phase == RQ_PHASE_UNREG_BULK) {
+ LASSERT(req->rq_next_phase != req->rq_phase);
+ LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
+
+ if (req->rq_req_deadline &&
+ !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
+ req->rq_req_deadline = 0;
+ if (req->rq_reply_deadline &&
+ !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
+ req->rq_reply_deadline = 0;
+ if (req->rq_bulk_deadline &&
+ !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
+ req->rq_bulk_deadline = 0;
- /*
- * Skip processing until reply is unlinked. We
- * can't return to pool before that and we can't
- * call interpret before that. We need to make
- * sure that all rdma transfers finished and will
- * not corrupt any data.
- */
- if (ptlrpc_client_recv_or_unlink(req) ||
- ptlrpc_client_bulk_active(req))
- continue;
+ /*
+ * Skip processing until reply is unlinked. We
+ * can't return to pool before that and we can't
+ * call interpret before that. We need to make
+ * sure that all rdma transfers finished and will
+ * not corrupt any data.
+ */
+ if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
+ ptlrpc_client_recv_or_unlink(req))
+ continue;
+ if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
+ ptlrpc_client_bulk_active(req))
+ continue;
/*
* Turn fail_loc off to prevent it from looping
ptlrpc_rqphase_move(req, req->rq_next_phase);
}
- if (req->rq_phase == RQ_PHASE_COMPLETE) {
- list_move_tail(&req->rq_set_chain, &comp_reqs);
- continue;
- }
-
if (req->rq_phase == RQ_PHASE_INTERPRET)
GOTO(interpret, req->rq_status);
/* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
* so it sets rq_intr regardless of individual rpc
- * timeouts. The synchronous IO waiting path sets
+ * timeouts. The synchronous IO waiting path sets
* rq_intr irrespective of whether ptlrpcd
* has seen a timeout. Our policy is to only interpret
* interrupted rpcs after they have timed out, so we
spin_lock(&req->rq_lock);
req->rq_resend = 1;
spin_unlock(&req->rq_lock);
- if (req->rq_bulk) {
- __u64 old_xid;
-
- if (!ptlrpc_unregister_bulk(req, 1))
- continue;
-
- /* ensure previous bulk fails */
- old_xid = req->rq_xid;
- req->rq_xid = ptlrpc_next_xid();
- CDEBUG(D_HA, "resend bulk "
- "old x"LPU64
- " new x"LPU64"\n",
- old_xid, req->rq_xid);
- }
+
+ if (req->rq_bulk != NULL &&
+ !ptlrpc_unregister_bulk(req, 1))
+ continue;
}
/*
* rq_wait_ctx is only touched by ptlrpcd,
}
rc = ptl_send_rpc(req, 0);
+ if (rc == -ENOMEM) {
+ spin_lock(&imp->imp_lock);
+ if (!list_empty(&req->rq_list))
+ list_del_init(&req->rq_list);
+ spin_unlock(&imp->imp_lock);
+ ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
+ continue;
+ }
if (rc) {
DEBUG_REQ(D_HA, req,
"send failed: rc = %d", rc);
req->rq_status = -EIO;
}
- ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
+ ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
- interpret:
- LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
+ interpret:
+ LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
- /* This moves to "unregistering" phase we need to wait for
- * reply unlink. */
- if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
- /* start async bulk unlink too */
- ptlrpc_unregister_bulk(req, 1);
- continue;
- }
+ /* This moves to "unregistering" phase we need to wait for
+ * reply unlink. */
+ if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
+ /* start async bulk unlink too */
+ ptlrpc_unregister_bulk(req, 1);
+ continue;
+ }
- if (!ptlrpc_unregister_bulk(req, 1))
- continue;
+ if (!ptlrpc_unregister_bulk(req, async))
+ continue;
- /* When calling interpret receiving already should be
- * finished. */
- LASSERT(!req->rq_receiving_reply);
+ /* When calling interpret receiving already should be
+ * finished. */
+ LASSERT(!req->rq_receiving_reply);
- ptlrpc_req_interpret(env, req, req->rq_status);
+ ptlrpc_req_interpret(env, req, req->rq_status);
if (ptlrpcd_check_work(req)) {
atomic_dec(&set->set_remaining);
}
ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
- CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
- "Completed RPC pname:cluuid:pid:xid:nid:"
- "opc %s:%s:%d:"LPU64":%s:%d\n",
- current_comm(), imp->imp_obd->obd_uuid.uuid,
- lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
- libcfs_nid2str(imp->imp_connection->c_peer.nid),
- lustre_msg_get_opc(req->rq_reqmsg));
+ if (req->rq_reqmsg != NULL)
+ CDEBUG(D_RPCTRACE,
+ "Completed RPC pname:cluuid:pid:xid:nid:"
+ "opc %s:%s:%d:%llu:%s:%d\n", current_comm(),
+ imp->imp_obd->obd_uuid.uuid,
+ lustre_msg_get_status(req->rq_reqmsg),
+ req->rq_xid,
+ libcfs_nid2str(imp->imp_connection->c_peer.nid),
+ lustre_msg_get_opc(req->rq_reqmsg));
spin_lock(&imp->imp_lock);
/* Request already may be not on sending or delaying list. This
list_del_init(&req->rq_list);
atomic_dec(&imp->imp_inflight);
}
+ list_del_init(&req->rq_unreplied_list);
spin_unlock(&imp->imp_lock);
atomic_dec(&set->set_remaining);
req->rq_timedout = 1;
spin_unlock(&req->rq_lock);
- DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
- "/real "CFS_DURATION_T"]",
+ DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent %lld/real %lld]",
req->rq_net_err ? "failed due to network error" :
((req->rq_real_sent == 0 ||
- cfs_time_before(req->rq_real_sent, req->rq_sent) ||
- cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
+ req->rq_real_sent < req->rq_sent ||
+ req->rq_real_sent >= req->rq_deadline) ?
"timed out for sent delay" : "timed out for slow reply"),
- req->rq_sent, req->rq_real_sent);
+ (s64)req->rq_sent, (s64)req->rq_real_sent);
- if (imp != NULL && obd_debug_peer_on_timeout)
- LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
+ if (imp != NULL && obd_debug_peer_on_timeout)
+ LNetDebugPeer(imp->imp_connection->c_peer);
ptlrpc_unregister_reply(req, async_unlink);
ptlrpc_unregister_bulk(req, async_unlink);
*/
int ptlrpc_expired_set(void *data)
{
- struct ptlrpc_request_set *set = data;
- struct list_head *tmp;
- time_t now = cfs_time_current_sec();
- ENTRY;
+ struct ptlrpc_request_set *set = data;
+ struct list_head *tmp;
+ time64_t now = ktime_get_real_seconds();
+ ENTRY;
LASSERT(set != NULL);
/*
*/
RETURN(1);
}
-EXPORT_SYMBOL(ptlrpc_expired_set);
/**
* Sets rq_intr flag in \a req under spinlock.
* Interrupts (sets interrupted flag) all uncompleted requests in
* a set \a data. Callback for l_wait_event for interruptible waits.
*/
-void ptlrpc_interrupted_set(void *data)
+static void ptlrpc_interrupted_set(void *data)
{
struct ptlrpc_request_set *set = data;
struct list_head *tmp;
struct ptlrpc_request *req =
list_entry(tmp, struct ptlrpc_request, rq_set_chain);
+ if (req->rq_intr)
+ continue;
+
if (req->rq_phase != RQ_PHASE_RPC &&
- req->rq_phase != RQ_PHASE_UNREGISTERING)
+ req->rq_phase != RQ_PHASE_UNREG_RPC &&
+ !req->rq_allow_intr)
continue;
ptlrpc_mark_interrupted(req);
}
}
-EXPORT_SYMBOL(ptlrpc_interrupted_set);
/**
* Get the smallest timeout in the set; this does NOT set a timeout.
*/
int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
{
- struct list_head *tmp;
- time_t now = cfs_time_current_sec();
- int timeout = 0;
- struct ptlrpc_request *req;
- int deadline;
- ENTRY;
+ struct list_head *tmp;
+ time64_t now = ktime_get_real_seconds();
+ int timeout = 0;
+ struct ptlrpc_request *req;
+ time64_t deadline;
+ ENTRY;
list_for_each(tmp, &set->set_requests) {
req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
}
RETURN(timeout);
}
-EXPORT_SYMBOL(ptlrpc_set_next_timeout);
/**
* Send all unset request from the set and then wait untill all
CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
set, timeout);
- if (timeout == 0 && !cfs_signal_pending())
- /*
- * No requests are in-flight (ether timed out
- * or delayed), so we can allow interrupts.
- * We still want to block for a limited time,
- * so we allow interrupts during the timeout.
- */
- lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
- ptlrpc_expired_set,
- ptlrpc_interrupted_set, set);
+ if ((timeout == 0 && !signal_pending(current)) ||
+ set->set_allow_intr)
+ /* No requests are in-flight (ether timed out
+ * or delayed), so we can allow interrupts.
+ * We still want to block for a limited time,
+ * so we allow interrupts during the timeout. */
+ lwi = LWI_TIMEOUT_INTR_ALL(
+ cfs_time_seconds(timeout ? timeout : 1),
+ ptlrpc_expired_set,
+ ptlrpc_interrupted_set, set);
else
/*
* At least one request is in flight, so no
* interrupts are allowed. Wait until all
- * complete, or an in-flight req times out.
+ * complete, or an in-flight req times out.
*/
lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
ptlrpc_expired_set, set);
/* LU-769 - if we ignored the signal because it was already
* pending when we started, we need to handle it now or we risk
* it being ignored forever */
- if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
- cfs_signal_pending()) {
+ if (rc == -ETIMEDOUT &&
+ (!lwi.lwi_allow_intr || set->set_allow_intr) &&
+ signal_pending(current)) {
sigset_t blocked_sigs =
cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
* like SIGINT or SIGKILL. We still ignore less
* important signals since ptlrpc set is not easily
* reentrant from userspace again */
- if (cfs_signal_pending())
+ if (signal_pending(current))
ptlrpc_interrupted_set(set);
cfs_restore_sigs(blocked_sigs);
}
if (!locked)
spin_lock(&request->rq_import->imp_lock);
list_del_init(&request->rq_replay_list);
+ list_del_init(&request->rq_unreplied_list);
if (!locked)
spin_unlock(&request->rq_import->imp_lock);
}
request->rq_import = NULL;
}
if (request->rq_bulk != NULL)
- ptlrpc_free_bulk_pin(request->rq_bulk);
+ ptlrpc_free_bulk(request->rq_bulk);
if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
sptlrpc_cli_free_reqbuf(request);
static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
/**
* Drop one request reference. Must be called with import imp_lock held.
- * When reference count drops to zero, reuqest is freed.
+ * When reference count drops to zero, request is freed.
*/
void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
{
assert_spin_locked(&request->rq_import->imp_lock);
(void)__ptlrpc_req_finished(request, 1);
}
-EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
/**
* Helper function
* Drops one reference count for request \a request.
* \a locked set indicates that caller holds import imp_lock.
* Frees the request whe reference count reaches zero.
+ *
+ * \retval 1 the request is freed
+ * \retval 0 some others still hold references on the request
*/
static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
{
- ENTRY;
- if (request == NULL)
- RETURN(1);
+ int count;
+ ENTRY;
- if (request == LP_POISON ||
- request->rq_reqmsg == LP_POISON) {
- CERROR("dereferencing freed request (bug 575)\n");
- LBUG();
- RETURN(1);
- }
+ if (!request)
+ RETURN(1);
- DEBUG_REQ(D_INFO, request, "refcount now %u",
+ LASSERT(request != LP_POISON);
+ LASSERT(request->rq_reqmsg != LP_POISON);
+
+ DEBUG_REQ(D_INFO, request, "refcount now %u",
atomic_read(&request->rq_refcount) - 1);
- if (atomic_dec_and_test(&request->rq_refcount)) {
- __ptlrpc_free_req(request, locked);
- RETURN(1);
- }
+ spin_lock(&request->rq_lock);
+ count = atomic_dec_return(&request->rq_refcount);
+ LASSERTF(count >= 0, "Invalid ref count %d\n", count);
- RETURN(0);
+ /* For open RPC, the client does not know the EA size (LOV, ACL, and
+ * so on) before replied, then the client has to reserve very large
+ * reply buffer. Such buffer will not be released until the RPC freed.
+ * Since The open RPC is replayable, we need to keep it in the replay
+ * list until close. If there are a lot of files opened concurrently,
+ * then the client may be OOM.
+ *
+ * If fact, it is unnecessary to keep reply buffer for open replay,
+ * related EAs have already been saved via mdc_save_lovea() before
+ * coming here. So it is safe to free the reply buffer some earlier
+ * before releasing the RPC to avoid client OOM. LU-9514 */
+ if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
+ spin_lock(&request->rq_early_free_lock);
+ sptlrpc_cli_free_repbuf(request);
+ request->rq_repbuf = NULL;
+ request->rq_repbuf_len = 0;
+ request->rq_repdata = NULL;
+ request->rq_reqdata_len = 0;
+ spin_unlock(&request->rq_early_free_lock);
+ }
+ spin_unlock(&request->rq_lock);
+
+ if (!count)
+ __ptlrpc_free_req(request, locked);
+
+ RETURN(!count);
}
/**
* The request owner (i.e. the thread doing the I/O) must call...
* Returns 0 on success or 1 if unregistering cannot be made.
*/
-int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
+static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
{
int rc;
struct l_wait_info lwi;
*/
LASSERT(!in_interrupt());
- /*
- * Let's setup deadline for reply unlink.
- */
- if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
- async && request->rq_reply_deadline == 0)
- request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
+ /* Let's setup deadline for reply unlink. */
+ if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
+ async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
+ request->rq_reply_deadline = ktime_get_real_seconds() +
+ LONG_UNLINK;
/*
* Nothing left to do.
if (!ptlrpc_client_recv_or_unlink(request))
RETURN(1);
- /*
- * Move to "Unregistering" phase as reply was not unlinked yet.
- */
- ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
+ /* Move to "Unregistering" phase as reply was not unlinked yet. */
+ ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
/*
* Do not wait for unlink to finish.
}
LASSERT(rc == -ETIMEDOUT);
- DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
- "rvcng=%d unlnk=%d/%d", request->rq_receiving_reply,
- request->rq_req_unlink, request->rq_reply_unlink);
+ DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
+ "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
+ request->rq_receiving_reply,
+ request->rq_req_unlinked,
+ request->rq_reply_unlinked);
}
RETURN(0);
}
-EXPORT_SYMBOL(ptlrpc_unregister_reply);
static void ptlrpc_free_request(struct ptlrpc_request *req)
{
if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
imp->imp_generation == imp->imp_last_generation_checked) {
- CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
+ CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
RETURN_EXIT;
}
- CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
+ CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
imp->imp_generation);
- if (imp->imp_generation != imp->imp_last_generation_checked)
+ if (imp->imp_generation != imp->imp_last_generation_checked ||
+ imp->imp_last_transno_checked == 0)
skip_committed_list = false;
imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
continue;
}
- DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
+ DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
imp->imp_peer_committed_transno);
free_req:
ptlrpc_free_request(req);
GOTO(out, 0);
list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
- rq_replay_list) {
+ rq_replay_list) {
LASSERT(req->rq_transno != 0);
- if (req->rq_import_generation < imp->imp_generation) {
- DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
+ if (req->rq_import_generation < imp->imp_generation ||
+ !req->rq_replay) {
+ DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
+ req->rq_import_generation <
+ imp->imp_generation ? "stale" : "closed");
+
+ if (imp->imp_replay_cursor == &req->rq_replay_list)
+ imp->imp_replay_cursor =
+ req->rq_replay_list.next;
+
ptlrpc_free_request(req);
}
}
ENTRY;
EXIT;
}
-EXPORT_SYMBOL(ptlrpc_cleanup_client);
/**
* Schedule previously sent request for resend.
req->rq_resend = 1;
req->rq_net_err = 0;
req->rq_timedout = 0;
- if (req->rq_bulk) {
- __u64 old_xid = req->rq_xid;
- /* ensure previous bulk fails */
- req->rq_xid = ptlrpc_next_xid();
- CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
- old_xid, req->rq_xid);
- }
ptlrpc_client_wake_req(req);
spin_unlock(&req->rq_lock);
}
-EXPORT_SYMBOL(ptlrpc_resend_req);
/* XXX: this function and rq_status are currently unused */
void ptlrpc_restart_req(struct ptlrpc_request *req)
ptlrpc_client_wake_req(req);
spin_unlock(&req->rq_lock);
}
-EXPORT_SYMBOL(ptlrpc_restart_req);
/**
* Grab additional reference on a request \a req
lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
+ spin_lock(&req->rq_lock);
+ req->rq_resend = 0;
+ spin_unlock(&req->rq_lock);
+
LASSERT(imp->imp_replayable);
/* Balanced in ptlrpc_free_committed, usually. */
ptlrpc_request_addref(req);
list_add(&req->rq_replay_list, &imp->imp_replay_list);
}
-EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
/**
* Send request and wait until it completes.
/**
* Callback used for replayed requests reply processing.
- * In case of succesful reply calls registeresd request replay callback.
+ * In case of successful reply calls registered request replay callback.
* In case of error restart replay process.
*/
static int ptlrpc_replay_interpret(const struct lu_env *env,
ENTRY;
atomic_dec(&imp->imp_replay_inflight);
- if (!ptlrpc_client_replied(req)) {
- CERROR("request replay timed out, restarting recovery\n");
- GOTO(out, rc = -ETIMEDOUT);
- }
+ /* Note: if it is bulk replay (MDS-MDS replay), then even if
+ * server got the request, but bulk transfer timeout, let's
+ * replay the bulk req again */
+ if (!ptlrpc_client_replied(req) ||
+ (req->rq_bulk != NULL &&
+ lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
+ DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
+ GOTO(out, rc = -ETIMEDOUT);
+ }
if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
(lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
lustre_msg_get_transno(req->rq_repmsg) ||
lustre_msg_get_transno(req->rq_repmsg) == 0,
- LPX64"/"LPX64"\n",
+ "%#llx/%#llx\n",
lustre_msg_get_transno(req->rq_reqmsg),
lustre_msg_get_transno(req->rq_repmsg));
}
/* transaction number shouldn't be bigger than the latest replayed */
if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
DEBUG_REQ(D_ERROR, req,
- "Reported transno "LPU64" is bigger than the "
- "replayed one: "LPU64, req->rq_transno,
+ "Reported transno %llu is bigger than the "
+ "replayed one: %llu", req->rq_transno,
lustre_msg_get_transno(req->rq_reqmsg));
GOTO(out, rc = -EINVAL);
}
DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
lustre_msg_get_status(req->rq_repmsg),
aa->praa_old_status);
+
+ /* Note: If the replay fails for MDT-MDT recovery, let's
+ * abort all of the following requests in the replay
+ * and sending list, because MDT-MDT update requests
+ * are dependent on each other, see LU-7039 */
+ if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
+ struct ptlrpc_request *free_req;
+ struct ptlrpc_request *tmp;
+
+ spin_lock(&imp->imp_lock);
+ list_for_each_entry_safe(free_req, tmp,
+ &imp->imp_replay_list,
+ rq_replay_list) {
+ ptlrpc_free_request(free_req);
+ }
+
+ list_for_each_entry_safe(free_req, tmp,
+ &imp->imp_committed_list,
+ rq_replay_list) {
+ ptlrpc_free_request(free_req);
+ }
+
+ list_for_each_entry_safe(free_req, tmp,
+ &imp->imp_delayed_list,
+ rq_list) {
+ spin_lock(&free_req->rq_lock);
+ free_req->rq_err = 1;
+ free_req->rq_status = -EIO;
+ ptlrpc_client_wake_req(free_req);
+ spin_unlock(&free_req->rq_lock);
+ }
+
+ list_for_each_entry_safe(free_req, tmp,
+ &imp->imp_sending_list,
+ rq_list) {
+ spin_lock(&free_req->rq_lock);
+ free_req->rq_err = 1;
+ free_req->rq_status = -EIO;
+ ptlrpc_client_wake_req(free_req);
+ spin_unlock(&free_req->rq_lock);
+ }
+ spin_unlock(&imp->imp_lock);
+ }
} else {
/* Put it back for re-replay. */
lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
DEBUG_REQ(D_HA, req, "REPLAY");
atomic_inc(&req->rq_import->imp_replay_inflight);
+ spin_lock(&req->rq_lock);
+ req->rq_early_free_repbuf = 0;
+ spin_unlock(&req->rq_lock);
ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
- ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
+ ptlrpcd_add_req(req);
RETURN(0);
}
-EXPORT_SYMBOL(ptlrpc_replay_req);
/**
* Aborts all in-flight request on import \a imp sending and delayed lists
EXIT;
}
-EXPORT_SYMBOL(ptlrpc_abort_inflight);
/**
* Abort all uncompleted requests in request set \a set
#define YEAR_2004 (1ULL << 30)
void ptlrpc_init_xid(void)
{
- time_t now = cfs_time_current_sec();
+ time64_t now = ktime_get_real_seconds();
spin_lock_init(&ptlrpc_last_xid_lock);
if (now < YEAR_2004) {
return next;
}
-EXPORT_SYMBOL(ptlrpc_next_xid);
+
+/**
+ * If request has a new allocated XID (new request or EINPROGRESS resend),
+ * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
+ * request to ensure previous bulk fails and avoid problems with lost replies
+ * and therefore several transfers landing into the same buffer from different
+ * sending attempts.
+ */
+void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
+{
+ struct ptlrpc_bulk_desc *bd = req->rq_bulk;
+
+ LASSERT(bd != NULL);
+
+ /* Generate new matchbits for all resend requests, including
+ * resend replay. */
+ if (req->rq_resend) {
+ __u64 old_mbits = req->rq_mbits;
+
+ /* First time resend on -EINPROGRESS will generate new xid,
+ * so we can actually use the rq_xid as rq_mbits in such case,
+ * however, it's bit hard to distinguish such resend with a
+ * 'resend for the -EINPROGRESS resend'. To make it simple,
+ * we opt to generate mbits for all resend cases. */
+ if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){
+ req->rq_mbits = ptlrpc_next_xid();
+ } else {
+ /* Old version transfers rq_xid to peer as
+ * matchbits. */
+ spin_lock(&req->rq_import->imp_lock);
+ list_del_init(&req->rq_unreplied_list);
+ ptlrpc_assign_next_xid_nolock(req);
+ spin_unlock(&req->rq_import->imp_lock);
+ req->rq_mbits = req->rq_xid;
+ }
+ CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
+ old_mbits, req->rq_mbits);
+ } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
+ /* Request being sent first time, use xid as matchbits. */
+ req->rq_mbits = req->rq_xid;
+ } else {
+ /* Replay request, xid and matchbits have already been
+ * correctly assigned. */
+ return;
+ }
+
+ /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
+ * that server can infer the number of bulks that were prepared,
+ * see LU-1431 */
+ req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
+ LNET_MAX_IOV) - 1;
+
+ /* Set rq_xid as rq_mbits to indicate the final bulk for the old
+ * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
+ *
+ * It's ok to directly set the rq_xid here, since this xid bump
+ * won't affect the request position in unreplied list. */
+ if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
+ req->rq_xid = req->rq_mbits;
+}
/**
* Get a glimpse at what next xid value might have been.
{
/* re-initialize the req */
req->rq_timeout = obd_timeout;
- req->rq_sent = cfs_time_current_sec();
+ req->rq_sent = ktime_get_real_seconds();
req->rq_deadline = req->rq_sent + req->rq_timeout;
- req->rq_reply_deadline = req->rq_deadline;
req->rq_phase = RQ_PHASE_INTERPRET;
req->rq_next_phase = RQ_PHASE_COMPLETE;
req->rq_xid = ptlrpc_next_xid();
req->rq_import_generation = req->rq_import->imp_generation;
- ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
+ ptlrpcd_add_req(req);
}
static int work_interpreter(const struct lu_env *env,
req->rq_import = class_import_get(imp);
req->rq_interpret_reply = work_interpreter;
/* don't want reply */
- req->rq_receiving_reply = 0;
- req->rq_req_unlink = req->rq_reply_unlink = 0;
req->rq_no_delay = req->rq_no_resend = 1;
req->rq_pill.rc_fmt = (void *)&worker_format;