X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Fclient.c;h=6b502073de6f11837945e9f925ae40012da083b1;hp=7e6d90a52b724accaf036ffe819e107a3944bf4a;hb=51206e8cd42134400fa0b6259a92d7138f3dc984;hpb=3fb3492697fd4356934675870cd1dad07485b0dc diff --git a/lustre/ptlrpc/client.c b/lustre/ptlrpc/client.c index 7e6d90a..6b50207 100644 --- a/lustre/ptlrpc/client.c +++ b/lustre/ptlrpc/client.c @@ -26,14 +26,18 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Whamcloud, Inc. */ /* * This file is part of Lustre, http://www.lustre.org/ * Lustre is a trademark of Sun Microsystems, Inc. */ +/** Implementation of client-side PortalRPC interfaces */ + #define DEBUG_SUBSYSTEM S_RPC #ifndef __KERNEL__ #include @@ -50,6 +54,9 @@ #include "ptlrpc_internal.h" +/** + * Initialize passed in client structure \a cl. + */ void ptlrpc_init_client(int req_portal, int rep_portal, char *name, struct ptlrpc_client *cl) { @@ -58,6 +65,9 @@ void ptlrpc_init_client(int req_portal, int rep_portal, char *name, cl->cli_name = name; } +/** + * Return PortalRPC connection for remore uud \a uuid + */ struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid) { struct ptlrpc_connection *c; @@ -67,7 +77,7 @@ struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid) err = ptlrpc_uuid_to_peer(uuid, &peer, &self); if (err != 0) { - CERROR("cannot find peer %s!\n", uuid->uuid); + CNETERR("cannot find peer %s!\n", uuid->uuid); return NULL; } @@ -82,7 +92,11 @@ struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid) return c; } -static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal) +/** + * Allocate and initialize new bulk descriptor + * Returns pointer to the descriptor or NULL on error. + */ +struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal) { struct ptlrpc_bulk_desc *desc; @@ -90,7 +104,7 @@ static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal if (!desc) return NULL; - spin_lock_init(&desc->bd_lock); + cfs_spin_lock_init(&desc->bd_lock); cfs_waitq_init(&desc->bd_waitq); desc->bd_max_iov = npages; desc->bd_iov_count = 0; @@ -101,6 +115,13 @@ static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal return desc; } +/** + * Prepare bulk descriptor for specified outgoing request \a req that + * can fit \a npages * 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, int npages, int type, int portal) { @@ -126,31 +147,11 @@ struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req, return desc; } -struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req, - int npages, int type, int portal) -{ - struct obd_export *exp = req->rq_export; - struct ptlrpc_bulk_desc *desc; - - ENTRY; - LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK); - - desc = new_bulk(npages, type, portal); - if (desc == NULL) - RETURN(NULL); - - desc->bd_export = class_export_get(exp); - desc->bd_req = req; - - desc->bd_cbid.cbid_fn = server_bulk_callback; - desc->bd_cbid.cbid_arg = desc; - - /* NB we don't assign rq_bulk here; server-side requests are - * re-used, and the handler frees the bulk desc explicitly. */ - - return desc; -} - +/** + * 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, cfs_page_t *page, int pageoffset, int len) { @@ -162,11 +163,17 @@ void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc, desc->bd_nob += len; + cfs_page_pin(page); ptlrpc_add_bulk_page(desc, page, pageoffset, len); } +/** + * 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 i; ENTRY; LASSERT(desc != NULL); @@ -181,12 +188,18 @@ void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc) else class_import_put(desc->bd_import); + for (i = 0; i < desc->bd_iov_count ; i++) + cfs_page_unpin(desc->bd_iov[i].kiov_page); + OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc, bd_iov[desc->bd_max_iov])); EXIT; } -/* Set server timelimit for this req */ +/** + * Set server timelimit for this req, i.e. how long are we willing to wait + * for reply before timing out this request. + */ void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req) { __u32 serv_est; @@ -197,6 +210,13 @@ void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req) if (AT_OFF) { /* non-AT settings */ + /** + * \a imp_server_timeout means this is reverse import and + * we send (currently only) ASTs to the client and cannot afford + * to wait too long for the reply, otherwise the other client + * (because of which we are sending this request) would + * timeout waiting for us + */ req->rq_timeout = req->rq_import->imp_server_timeout ? obd_timeout / 2 : obd_timeout; } else { @@ -228,7 +248,7 @@ static void ptlrpc_at_adj_service(struct ptlrpc_request *req, idx = import_at_get_index(req->rq_import, req->rq_request_portal); /* max service estimates are tracked on the server side, so just keep minimal history here */ - oldse = at_add(&at->iat_service_estimate[idx], serv_est); + oldse = at_measured(&at->iat_service_estimate[idx], serv_est); if (oldse != 0) CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d " "has changed from %d to %d\n", @@ -260,7 +280,7 @@ static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req, CFS_DURATION_T"\n", service_time, cfs_time_sub(now, req->rq_sent)); - oldnl = at_add(&at->iat_net_latency, nl); + oldnl = at_measured(&at->iat_net_latency, nl); if (oldnl != 0) CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) " "has changed from %d to %d\n", @@ -290,7 +310,7 @@ static int unpack_reply(struct ptlrpc_request *req) return 0; } -/* +/** * Handle an early reply message, called with the rq_lock held. * If anything goes wrong just ignore it - same as if it never happened */ @@ -302,11 +322,11 @@ static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req) ENTRY; req->rq_early = 0; - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); rc = sptlrpc_cli_unwrap_early_reply(req, &early_req); if (rc) { - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); RETURN(rc); } @@ -321,7 +341,7 @@ static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req) sptlrpc_cli_finish_early_reply(early_req); - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); if (rc == 0) { /* Adjust the local timeout for this req */ @@ -344,49 +364,57 @@ static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req) RETURN(rc); } +/** + * Wind down request pool \a pool. + * Frees all requests from the pool too + */ void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool) { - struct list_head *l, *tmp; + cfs_list_t *l, *tmp; struct ptlrpc_request *req; LASSERT(pool != NULL); - spin_lock(&pool->prp_lock); - list_for_each_safe(l, tmp, &pool->prp_req_list) { - req = list_entry(l, struct ptlrpc_request, rq_list); - list_del(&req->rq_list); + cfs_spin_lock(&pool->prp_lock); + cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) { + req = cfs_list_entry(l, struct ptlrpc_request, rq_list); + cfs_list_del(&req->rq_list); LASSERT(req->rq_reqbuf); LASSERT(req->rq_reqbuf_len == pool->prp_rq_size); - OBD_FREE(req->rq_reqbuf, pool->prp_rq_size); + OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size); OBD_FREE(req, sizeof(*req)); } - spin_unlock(&pool->prp_lock); + cfs_spin_unlock(&pool->prp_lock); OBD_FREE(pool, sizeof(*pool)); } +/** + * 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 i; int size = 1; - while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD) + while (size < pool->prp_rq_size) size <<= 1; - LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size, + LASSERTF(cfs_list_empty(&pool->prp_req_list) || + size == pool->prp_rq_size, "Trying to change pool size with nonempty pool " "from %d to %d bytes\n", pool->prp_rq_size, size); - spin_lock(&pool->prp_lock); + cfs_spin_lock(&pool->prp_lock); pool->prp_rq_size = size; for (i = 0; i < num_rq; i++) { struct ptlrpc_request *req; struct lustre_msg *msg; - spin_unlock(&pool->prp_lock); + cfs_spin_unlock(&pool->prp_lock); OBD_ALLOC(req, sizeof(struct ptlrpc_request)); if (!req) return; - OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD); + OBD_ALLOC_LARGE(msg, size); if (!msg) { OBD_FREE(req, sizeof(struct ptlrpc_request)); return; @@ -394,13 +422,21 @@ void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq) 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); + cfs_spin_lock(&pool->prp_lock); + cfs_list_add_tail(&req->rq_list, &pool->prp_req_list); } - spin_unlock(&pool->prp_lock); + cfs_spin_unlock(&pool->prp_lock); return; } +/** + * Create and initialize new request pool with given attributes: + * \a num_rq - initial number of requests to create for the pool + * \a msgsize - maximum message size possible for requests in thid pool + * \a populate_pool - function to be called when more requests need to be added + * to the pool + * Returns pointer to newly created pool or NULL on error. + */ struct ptlrpc_request_pool * ptlrpc_init_rq_pool(int num_rq, int msgsize, void (*populate_pool)(struct ptlrpc_request_pool *, int)) @@ -414,14 +450,14 @@ ptlrpc_init_rq_pool(int num_rq, int msgsize, /* Request next power of two for the allocation, because internally kernel would do exactly this */ - spin_lock_init(&pool->prp_lock); + cfs_spin_lock_init(&pool->prp_lock); CFS_INIT_LIST_HEAD(&pool->prp_req_list); - pool->prp_rq_size = msgsize; + pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD; pool->prp_populate = populate_pool; populate_pool(pool, num_rq); - if (list_empty(&pool->prp_req_list)) { + if (cfs_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; @@ -429,6 +465,9 @@ ptlrpc_init_rq_pool(int num_rq, int msgsize, return pool; } +/** + * Fetches one request from pool \a pool + */ static struct ptlrpc_request * ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool) { @@ -438,21 +477,21 @@ ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool) if (!pool) return NULL; - spin_lock(&pool->prp_lock); + cfs_spin_lock(&pool->prp_lock); /* See if we have anything in a pool, and bail out if nothing, * in writeout path, where this matters, this is safe to do, because * nothing is lost in this case, and when some in-flight requests * complete, this code will be called again. */ - if (unlikely(list_empty(&pool->prp_req_list))) { - spin_unlock(&pool->prp_lock); + if (unlikely(cfs_list_empty(&pool->prp_req_list))) { + cfs_spin_unlock(&pool->prp_lock); return NULL; } - request = list_entry(pool->prp_req_list.next, struct ptlrpc_request, - rq_list); - list_del_init(&request->rq_list); - spin_unlock(&pool->prp_lock); + request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request, + rq_list); + cfs_list_del_init(&request->rq_list); + cfs_spin_unlock(&pool->prp_lock); LASSERT(request->rq_reqbuf); LASSERT(request->rq_pool); @@ -466,15 +505,18 @@ ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool) return request; } +/** + * Returns freed \a request to pool. + */ static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request) { struct ptlrpc_request_pool *pool = request->rq_pool; - spin_lock(&pool->prp_lock); - LASSERT(list_empty(&request->rq_list)); + cfs_spin_lock(&pool->prp_lock); + LASSERT(cfs_list_empty(&request->rq_list)); LASSERT(!request->rq_receiving_reply); - list_add_tail(&request->rq_list, &pool->prp_req_list); - spin_unlock(&pool->prp_lock); + cfs_list_add_tail(&request->rq_list, &pool->prp_req_list); + cfs_spin_unlock(&pool->prp_lock); } static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request, @@ -523,7 +565,7 @@ static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request, ptlrpc_at_set_req_timeout(request); - spin_lock_init(&request->rq_lock); + cfs_spin_lock_init(&request->rq_lock); CFS_INIT_LIST_HEAD(&request->rq_list); CFS_INIT_LIST_HEAD(&request->rq_timed_list); CFS_INIT_LIST_HEAD(&request->rq_replay_list); @@ -532,8 +574,9 @@ static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request, CFS_INIT_LIST_HEAD(&request->rq_history_list); CFS_INIT_LIST_HEAD(&request->rq_exp_list); cfs_waitq_init(&request->rq_reply_waitq); + cfs_waitq_init(&request->rq_set_waitq); request->rq_xid = ptlrpc_next_xid(); - atomic_set(&request->rq_refcount, 1); + cfs_atomic_set(&request->rq_refcount, 1); lustre_msg_set_opc(request->rq_reqmsg, opcode); @@ -558,12 +601,22 @@ int ptlrpc_request_bufs_pack(struct ptlrpc_request *request, } EXPORT_SYMBOL(ptlrpc_request_bufs_pack); +/** + * Pack request buffers for network transfer, performing necessary encryption + * steps if necessary. + */ int ptlrpc_request_pack(struct ptlrpc_request *request, __u32 version, int opcode) { return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL); } +/** + * Helper function to allocate new request on import \a imp + * and possibly using existing request from pool \a pool if provided. + * Returns allocated request structure with import field filled or + * NULL on error. + */ static inline struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp, struct ptlrpc_request_pool *pool) @@ -591,6 +644,12 @@ struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp, return request; } +/** + * Helper function for creating a request. + * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits + * buffer structures according to capsule template \a format. + * Returns allocated request structure pointer or NULL on error. + */ static struct ptlrpc_request * ptlrpc_request_alloc_internal(struct obd_import *imp, struct ptlrpc_request_pool * pool, @@ -607,12 +666,20 @@ ptlrpc_request_alloc_internal(struct obd_import *imp, return request; } +/** + * Allocate new request structure for import \a imp and initialize its + * buffer structure according to capsule template \a format. + */ struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp, const struct req_format *format) { return ptlrpc_request_alloc_internal(imp, NULL, format); } +/** + * Allocate new request structure for import \a imp from pool \a pool and + * initialize its buffer structure according to capsule template \a format. + */ struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp, struct ptlrpc_request_pool * pool, const struct req_format *format) @@ -620,6 +687,10 @@ struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp, return ptlrpc_request_alloc_internal(imp, pool, format); } +/** + * For requests not from pool, free memory of the request structure. + * For requests obtained from a pool earlier, return request back to pool. + */ void ptlrpc_request_free(struct ptlrpc_request *request) { if (request->rq_pool) @@ -628,6 +699,13 @@ void ptlrpc_request_free(struct ptlrpc_request *request) OBD_FREE_PTR(request); } +/** + * Allocate new request for operatione \a opcode and immediatelly pack it for + * network transfer. + * Only used for simple requests like OBD_PING where the only important + * part of the request is operation itself. + * Returns allocated request or NULL on error. + */ struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp, const struct req_format *format, __u32 version, int opcode) @@ -645,6 +723,13 @@ struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp, return req; } +/** + * 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, @@ -667,6 +752,9 @@ ptlrpc_prep_req_pool(struct obd_import *imp, return request; } +/** + * 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) @@ -675,6 +763,13 @@ ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count, NULL); } +/** + * Allocate "fake" request that would not be sent anywhere in the end. + * Only used as a hack because we have no other way of performing + * async actions in lustre between layers. + * Used on MDS to request object preallocations from more than one OST at a + * time. + */ struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp, unsigned int timeout, ptlrpc_interpterer_t interpreter) @@ -707,35 +802,50 @@ struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp, request->rq_no_delay = request->rq_no_resend = 1; request->rq_fake = 1; - spin_lock_init(&request->rq_lock); + cfs_spin_lock_init(&request->rq_lock); CFS_INIT_LIST_HEAD(&request->rq_list); CFS_INIT_LIST_HEAD(&request->rq_replay_list); CFS_INIT_LIST_HEAD(&request->rq_set_chain); CFS_INIT_LIST_HEAD(&request->rq_history_list); CFS_INIT_LIST_HEAD(&request->rq_exp_list); cfs_waitq_init(&request->rq_reply_waitq); + cfs_waitq_init(&request->rq_set_waitq); request->rq_xid = ptlrpc_next_xid(); - atomic_set(&request->rq_refcount, 1); + cfs_atomic_set(&request->rq_refcount, 1); RETURN(request); } +/** + * Indicate that processing of "fake" request is finished. + */ void ptlrpc_fakereq_finished(struct ptlrpc_request *req) { - /* if we kill request before timeout - need adjust counter */ - if (req->rq_phase == RQ_PHASE_RPC) { - struct ptlrpc_request_set *set = req->rq_set; + struct ptlrpc_request_set *set = req->rq_set; + int wakeup = 0; - if (set) - set->set_remaining --; - } + /* hold ref on the request to prevent others (ptlrpcd) to free it */ + ptlrpc_request_addref(req); + cfs_list_del_init(&req->rq_list); + + /* if we kill request before timeout - need adjust counter */ + if (req->rq_phase == RQ_PHASE_RPC && set != NULL && + cfs_atomic_dec_and_test(&set->set_remaining)) + wakeup = 1; ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE); - list_del_init(&req->rq_list); -} + /* Only need to call wakeup once when to be empty. */ + if (wakeup) + cfs_waitq_signal(&set->set_waitq); + ptlrpc_req_finished(req); +} +/** + * Allocate and initialize new request set structure. + * 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; @@ -744,61 +854,82 @@ struct ptlrpc_request_set *ptlrpc_prep_set(void) OBD_ALLOC(set, sizeof *set); if (!set) RETURN(NULL); + cfs_atomic_set(&set->set_refcount, 1); CFS_INIT_LIST_HEAD(&set->set_requests); cfs_waitq_init(&set->set_waitq); - set->set_remaining = 0; - spin_lock_init(&set->set_new_req_lock); + cfs_atomic_set(&set->set_new_count, 0); + cfs_atomic_set(&set->set_remaining, 0); + cfs_spin_lock_init(&set->set_new_req_lock); CFS_INIT_LIST_HEAD(&set->set_new_requests); CFS_INIT_LIST_HEAD(&set->set_cblist); RETURN(set); } -/* Finish with this set; opposite of prep_set. */ +/** + * Wind down and free request set structure previously allocated with + * ptlrpc_prep_set. + * Ensures that all requests on the set have completed and removes + * all requests from the request list in a set. + * If any unsent request happen to be on the list, pretends that they got + * an error in flight and calls their completion handler. + */ void ptlrpc_set_destroy(struct ptlrpc_request_set *set) { - struct list_head *tmp; - struct list_head *next; + cfs_list_t *tmp; + cfs_list_t *next; int expected_phase; int n = 0; ENTRY; /* Requests on the set should either all be completed, or all be new */ - expected_phase = (set->set_remaining == 0) ? + expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ? RQ_PHASE_COMPLETE : RQ_PHASE_NEW; - list_for_each (tmp, &set->set_requests) { + cfs_list_for_each (tmp, &set->set_requests) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); LASSERT(req->rq_phase == expected_phase); n++; } - LASSERTF(set->set_remaining == 0 || set->set_remaining == n, "%d / %d\n", - set->set_remaining, n); + LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 || + cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n", + cfs_atomic_read(&set->set_remaining), n); - list_for_each_safe(tmp, next, &set->set_requests) { + cfs_list_for_each_safe(tmp, next, &set->set_requests) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_set_chain); - list_del_init(&req->rq_set_chain); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); + cfs_list_del_init(&req->rq_set_chain); LASSERT(req->rq_phase == expected_phase); if (req->rq_phase == RQ_PHASE_NEW) { ptlrpc_req_interpret(NULL, req, -EBADR); - set->set_remaining--; + cfs_atomic_dec(&set->set_remaining); } + cfs_spin_lock(&req->rq_lock); req->rq_set = NULL; + req->rq_invalid_rqset = 0; + cfs_spin_unlock(&req->rq_lock); + ptlrpc_req_finished (req); } - LASSERT(set->set_remaining == 0); + LASSERT(cfs_atomic_read(&set->set_remaining) == 0); - OBD_FREE(set, sizeof(*set)); + ptlrpc_reqset_put(set); EXIT; } +/** + * Add a callback function \a fn to the set. + * This function would be called when all requests on this set are completed. + * The function will be passed \a data argument. + */ int ptlrpc_set_add_cb(struct ptlrpc_request_set *set, set_interpreter_func fn, void *data) { @@ -810,49 +941,64 @@ int ptlrpc_set_add_cb(struct ptlrpc_request_set *set, cbdata->psc_interpret = fn; cbdata->psc_data = data; - list_add_tail(&cbdata->psc_item, &set->set_cblist); + cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist); RETURN(0); } +/** + * Add a new request to the general purpose request set. + * Assumes request reference from the caller. + */ void ptlrpc_set_add_req(struct ptlrpc_request_set *set, struct ptlrpc_request *req) { + LASSERT(cfs_list_empty(&req->rq_set_chain)); + /* The set takes over the caller's request reference */ - list_add_tail(&req->rq_set_chain, &set->set_requests); + cfs_list_add_tail(&req->rq_set_chain, &set->set_requests); req->rq_set = set; - set->set_remaining++; + cfs_atomic_inc(&set->set_remaining); + req->rq_queued_time = cfs_time_current(); } /** - * Lock so many callers can add things, the context that owns the set - * is supposed to notice these and move them into the set proper. + * Add a request to a request with dedicated server thread + * and wake the thread to make any necessary processing. + * Currently only used for ptlrpcd. */ -int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc, +void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc, struct ptlrpc_request *req) { struct ptlrpc_request_set *set = pc->pc_set; + int count, i; - /* - * Let caller know that we stopped and will not handle this request. - * It needs to take care itself of request. - */ - if (test_bit(LIOD_STOP, &pc->pc_flags)) - return -EALREADY; + LASSERT(req->rq_set == NULL); + LASSERT(cfs_test_bit(LIOD_STOP, &pc->pc_flags) == 0); - spin_lock(&set->set_new_req_lock); + cfs_spin_lock(&set->set_new_req_lock); /* * The set takes over the caller's request reference. */ - list_add_tail(&req->rq_set_chain, &set->set_new_requests); req->rq_set = set; - spin_unlock(&set->set_new_req_lock); + req->rq_queued_time = cfs_time_current(); + cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests); + count = cfs_atomic_inc_return(&set->set_new_count); + cfs_spin_unlock(&set->set_new_req_lock); - cfs_waitq_signal(&set->set_waitq); - return 0; + /* Only need to call wakeup once for the first entry. */ + if (count == 1) { + cfs_waitq_signal(&set->set_waitq); + + /* XXX: It maybe unnecessary to wakeup all the partners. But to + * guarantee the async RPC can be processed ASAP, we have + * no other better choice. It maybe fixed in future. */ + for (i = 0; i < pc->pc_npartners; i++) + cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq); + } } -/* +/** * Based on the current state of the import, determine if the request * can be sent, is an error, or should be delayed. * @@ -880,20 +1026,18 @@ static int ptlrpc_import_delay_req(struct obd_import *imp, } else if (imp->imp_state == LUSTRE_IMP_CLOSED) { DEBUG_REQ(D_ERROR, req, "IMP_CLOSED "); *status = -EIO; + } else if (ptlrpc_send_limit_expired(req)) { + /* probably doesn't need to be a D_ERROR after initial testing */ + DEBUG_REQ(D_ERROR, req, "send limit expired "); + *status = -EIO; } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING && imp->imp_state == LUSTRE_IMP_CONNECTING) { /* allow CONNECT even if import is invalid */ ; - if (atomic_read(&imp->imp_inval_count) != 0) { + if (cfs_atomic_read(&imp->imp_inval_count) != 0) { DEBUG_REQ(D_ERROR, req, "invalidate in flight"); *status = -EIO; } - } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) || - imp->imp_obd->obd_no_recov) { - /* If the import has been invalidated (such as by an OST - * failure), and if the import(MGC) tried all of its connection - * list (Bug 13464), the request must fail with -ESHUTDOWN. - * This indicates the requests should be discarded; an -EIO - * may result in a resend of the request. */ + } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) { if (!imp->imp_deactive) DEBUG_REQ(D_ERROR, req, "IMP_INVALID"); *status = -ESHUTDOWN; /* bz 12940 */ @@ -902,7 +1046,7 @@ static int ptlrpc_import_delay_req(struct obd_import *imp, *status = -EIO; } else if (req->rq_send_state != imp->imp_state) { /* invalidate in progress - any requests should be drop */ - if (atomic_read(&imp->imp_inval_count) != 0) { + if (cfs_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) { @@ -915,13 +1059,24 @@ static int ptlrpc_import_delay_req(struct obd_import *imp, RETURN(delay); } - -/* Conditionally suppress specific console messages */ +/** + * 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. + */ static int ptlrpc_console_allow(struct ptlrpc_request *req) { - __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); + __u32 opc; int err; + /* Fake requests include no rq_reqmsg */ + if (req->rq_fake) + return 0; + + 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)) && @@ -940,6 +1095,10 @@ static int ptlrpc_console_allow(struct ptlrpc_request *req) return 1; } +/** + * Check request processing status. + * Returns the status. + */ static int ptlrpc_check_status(struct ptlrpc_request *req) { int err; @@ -982,7 +1141,9 @@ static int ptlrpc_check_status(struct ptlrpc_request *req) } /** - * save pre-versions for replay + * save pre-versions of objects into request for replay. + * Versions are obtained from server reply. + * used for VBR. */ static void ptlrpc_save_versions(struct ptlrpc_request *req) { @@ -1004,6 +1165,10 @@ static void ptlrpc_save_versions(struct ptlrpc_request *req) /** * Callback function called when client receives RPC reply for \a req. + * Returns 0 on success or error code. + * The return alue would be assigned to req->rq_status by the caller + * as request processing status. + * This function also decides if the request needs to be saved for later replay. */ static int after_reply(struct ptlrpc_request *req) { @@ -1014,14 +1179,26 @@ static int after_reply(struct ptlrpc_request *req) long timediff; ENTRY; - LASSERT(!req->rq_receiving_reply); - LASSERT(obd); - LASSERT(req->rq_nob_received <= req->rq_repbuf_len); + LASSERT(obd != NULL); + /* repbuf must be unlinked */ + LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink); + + if (req->rq_reply_truncate) { + if (ptlrpc_no_resend(req)) { + DEBUG_REQ(D_ERROR, req, "reply buffer overflow," + " expected: %d, actual size: %d", + req->rq_nob_received, req->rq_repbuf_len); + RETURN(-EOVERFLOW); + } - if (req->rq_reply_truncate && !req->rq_no_resend) { - req->rq_resend = 1; sptlrpc_cli_free_repbuf(req); - req->rq_replen = req->rq_nob_received; + /* Pass the required reply buffer size (include + * space for early reply). + * NB: no need to roundup because alloc_repbuf + * will roundup it */ + req->rq_replen = req->rq_nob_received; + req->rq_nob_received = 0; + req->rq_resend = 1; RETURN(0); } @@ -1029,7 +1206,6 @@ static int after_reply(struct ptlrpc_request *req) * NB Until this point, the whole of the incoming message, * including buflens, status etc is in the sender's byte order. */ - rc = sptlrpc_cli_unwrap_reply(req); if (rc) { DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc); @@ -1046,7 +1222,7 @@ static int after_reply(struct ptlrpc_request *req) if (rc) RETURN(rc); - do_gettimeofday(&work_start); + cfs_gettimeofday(&work_start); timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL); if (obd->obd_svc_stats != NULL) { lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR, @@ -1061,7 +1237,8 @@ static int after_reply(struct ptlrpc_request *req) RETURN(-EPROTO); } - OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val); + if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING) + CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val); ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg)); ptlrpc_at_adj_net_latency(req, lustre_msg_get_service_time(req->rq_repmsg)); @@ -1100,7 +1277,7 @@ static int after_reply(struct ptlrpc_request *req) } if (imp->imp_replayable) { - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); /* * No point in adding already-committed requests to the replay * list, we will just remove them immediately. b=9829 @@ -1113,9 +1290,9 @@ static int after_reply(struct ptlrpc_request *req) ptlrpc_save_versions(req); ptlrpc_retain_replayable_request(req, imp); } else if (req->rq_commit_cb != NULL) { - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); req->rq_commit_cb(req); - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); } /* @@ -1126,56 +1303,67 @@ static int after_reply(struct ptlrpc_request *req) lustre_msg_get_last_committed(req->rq_repmsg); } ptlrpc_free_committed(imp); - spin_unlock(&imp->imp_lock); + + if (req->rq_transno > imp->imp_peer_committed_transno) + ptlrpc_pinger_commit_expected(imp); + + cfs_spin_unlock(&imp->imp_lock); } RETURN(rc); } +/** + * Helper function to send request \a req over the network for the first time + * Also adjusts request phase. + * Returns 0 on success or error code. + */ static int ptlrpc_send_new_req(struct ptlrpc_request *req) { - struct obd_import *imp; + struct obd_import *imp = req->rq_import; int rc; ENTRY; LASSERT(req->rq_phase == RQ_PHASE_NEW); - if (req->rq_sent && (req->rq_sent > cfs_time_current_sec())) + if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) && + (!req->rq_generation_set || + req->rq_import_generation == imp->imp_generation)) RETURN (0); ptlrpc_rqphase_move(req, RQ_PHASE_RPC); - imp = req->rq_import; - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); - req->rq_import_generation = imp->imp_generation; + if (!req->rq_generation_set) + req->rq_import_generation = imp->imp_generation; if (ptlrpc_import_delay_req(imp, req, &rc)) { - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); req->rq_waiting = 1; - spin_unlock(&req->rq_lock); + cfs_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), ptlrpc_import_state_name(req->rq_send_state), ptlrpc_import_state_name(imp->imp_state)); - LASSERT(list_empty(&req->rq_list)); - list_add_tail(&req->rq_list, &imp->imp_delayed_list); - atomic_inc(&req->rq_import->imp_inflight); - spin_unlock(&imp->imp_lock); + LASSERT(cfs_list_empty(&req->rq_list)); + cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list); + cfs_atomic_inc(&req->rq_import->imp_inflight); + cfs_spin_unlock(&imp->imp_lock); RETURN(0); } if (rc != 0) { - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); req->rq_status = rc; ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); RETURN(rc); } - LASSERT(list_empty(&req->rq_list)); - list_add_tail(&req->rq_list, &imp->imp_sending_list); - atomic_inc(&req->rq_import->imp_inflight); - spin_unlock(&imp->imp_lock); + LASSERT(cfs_list_empty(&req->rq_list)); + cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list); + cfs_atomic_inc(&req->rq_import->imp_inflight); + cfs_spin_unlock(&imp->imp_lock); lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid()); @@ -1206,20 +1394,27 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req) RETURN(0); } -/* this sends any unsent RPCs in @set and returns TRUE if all are sent */ +/** + * this sends any unsent RPCs in \a set and returns 1 if all are sent + * and no more replies are expected. + * (it is possible to get less replies than requests sent e.g. due to timed out + * requests or requests that we had trouble to send out) + */ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) { - struct list_head *tmp; + cfs_list_t *tmp; int force_timer_recalc = 0; ENTRY; - if (set->set_remaining == 0) + if (cfs_atomic_read(&set->set_remaining) == 0) RETURN(1); - list_for_each(tmp, &set->set_requests) { + cfs_list_for_each(tmp, &set->set_requests) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); struct obd_import *imp = req->rq_import; + int unregistered = 0; int rc = 0; if (req->rq_phase == RQ_PHASE_NEW && @@ -1293,10 +1488,21 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) if (ptlrpc_client_recv_or_unlink(req) || ptlrpc_client_bulk_active(req)) continue; + /* If there is no need to resend, fail it now. */ + if (req->rq_no_resend) { + if (req->rq_status == 0) + req->rq_status = -EIO; + ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); + GOTO(interpret, req->rq_status); + } else { + continue; + } } if (req->rq_err) { + cfs_spin_lock(&req->rq_lock); req->rq_replied = 0; + cfs_spin_unlock(&req->rq_lock); if (req->rq_status == 0) req->rq_status = -EIO; ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); @@ -1327,13 +1533,15 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) if (!ptlrpc_unregister_reply(req, 1)) continue; - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); if (ptlrpc_import_delay_req(imp, req, &status)){ /* put on delay list - only if we wait * recovery finished - before send */ - list_del_init(&req->rq_list); - list_add_tail(&req->rq_list, &imp->imp_delayed_list); - spin_unlock(&imp->imp_lock); + cfs_list_del_init(&req->rq_list); + cfs_list_add_tail(&req->rq_list, + &imp-> \ + imp_delayed_list); + cfs_spin_unlock(&imp->imp_lock); continue; } @@ -1341,29 +1549,33 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) req->rq_status = status; ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); GOTO(interpret, req->rq_status); } - if (req->rq_no_resend && !req->rq_wait_ctx) { + if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) { req->rq_status = -ENOTCONN; ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); GOTO(interpret, req->rq_status); } - list_del_init(&req->rq_list); - list_add_tail(&req->rq_list, + cfs_list_del_init(&req->rq_list); + cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list); - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); + cfs_spin_lock(&req->rq_lock); req->rq_waiting = 0; + cfs_spin_unlock(&req->rq_lock); - if (req->rq_timedout||req->rq_resend) { + if (req->rq_timedout || req->rq_resend) { /* This is re-sending anyways, * let's mark req as resend. */ + cfs_spin_lock(&req->rq_lock); req->rq_resend = 1; + cfs_spin_unlock(&req->rq_lock); if (req->rq_bulk) { __u64 old_xid; @@ -1387,15 +1599,21 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) if (status) { if (req->rq_err) { req->rq_status = status; + cfs_spin_lock(&req->rq_lock); req->rq_wait_ctx = 0; + cfs_spin_unlock(&req->rq_lock); force_timer_recalc = 1; } else { + cfs_spin_lock(&req->rq_lock); req->rq_wait_ctx = 1; + cfs_spin_unlock(&req->rq_lock); } continue; } else { + cfs_spin_lock(&req->rq_lock); req->rq_wait_ctx = 0; + cfs_spin_unlock(&req->rq_lock); } rc = ptl_send_rpc(req, 0); @@ -1403,33 +1621,41 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) DEBUG_REQ(D_HA, req, "send failed (%d)", rc); force_timer_recalc = 1; + cfs_spin_lock(&req->rq_lock); req->rq_net_err = 1; + cfs_spin_unlock(&req->rq_lock); } /* need to reset the timeout */ force_timer_recalc = 1; } - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); if (ptlrpc_client_early(req)) { ptlrpc_at_recv_early_reply(req); - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); continue; } /* Still waiting for a reply? */ if (ptlrpc_client_recv(req)) { - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); continue; } /* Did we actually receive a reply? */ if (!ptlrpc_client_replied(req)) { - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); continue; } - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); + + /* unlink from net because we are going to + * swab in-place of reply buffer */ + unregistered = ptlrpc_unregister_reply(req, 1); + if (!unregistered) + continue; req->rq_status = after_reply(req); if (req->rq_resend) @@ -1440,7 +1666,7 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) * process the reply. Similarly if the RPC returned * an error, and therefore the bulk will never arrive. */ - if (req->rq_bulk == NULL || req->rq_status != 0) { + if (req->rq_bulk == NULL || req->rq_status < 0) { ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); GOTO(interpret, req->rq_status); } @@ -1458,7 +1684,7 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) * was good after getting the REPLY for her GET or * the ACK for her PUT. */ DEBUG_REQ(D_ERROR, req, "bulk transfer failed"); - LBUG(); + req->rq_status = -EIO; } ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET); @@ -1468,8 +1694,11 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) /* This moves to "unregistering" phase we need to wait for * reply unlink. */ - if (!ptlrpc_unregister_reply(req, 1)) + if (!unregistered && !ptlrpc_unregister_reply(req, 1)) { + /* start async bulk unlink too */ + ptlrpc_unregister_bulk(req, 1); continue; + } if (!ptlrpc_unregister_bulk(req, 1)) continue; @@ -1490,44 +1719,49 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set) libcfs_nid2str(imp->imp_connection->c_peer.nid), req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1); - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); /* Request already may be not on sending or delaying list. This * may happen in the case of marking it erroneous for the case * ptlrpc_import_delay_req(req, status) find it impossible to * allow sending this rpc and returns *status != 0. */ - if (!list_empty(&req->rq_list)) { - list_del_init(&req->rq_list); - atomic_dec(&imp->imp_inflight); + if (!cfs_list_empty(&req->rq_list)) { + cfs_list_del_init(&req->rq_list); + cfs_atomic_dec(&imp->imp_inflight); } - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); - set->set_remaining--; + cfs_atomic_dec(&set->set_remaining); cfs_waitq_broadcast(&imp->imp_recovery_waitq); } /* If we hit an error, we want to recover promptly. */ - RETURN(set->set_remaining == 0 || force_timer_recalc); + RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc); } -/* Return 1 if we should give up, else 0 */ +/** + * Time out request \a req. is \a async_unlink is set, that means do not wait + * until LNet actually confirms network buffer unlinking. + * Return 1 if we should give up further retrying attempts or 0 otherwise. + */ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) { struct obd_import *imp = req->rq_import; int rc = 0; ENTRY; - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); req->rq_timedout = 1; - spin_unlock(&req->rq_lock); - - DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req, - "Request x"LPU64" sent from %s to NID %s "CFS_DURATION_T"s " - "ago has %s ("CFS_DURATION_T"s prior to deadline).\n", - req->rq_xid, imp ? imp->imp_obd->obd_name : "", - imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "", - cfs_time_sub(cfs_time_current_sec(), req->rq_sent), - req->rq_net_err ? "failed due to network error" : "timed out", - cfs_time_sub(req->rq_deadline, req->rq_sent)); + cfs_spin_unlock(&req->rq_lock); + + DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req, "Request " + " sent has %s: [sent "CFS_DURATION_T"/" + "real "CFS_DURATION_T"]", + 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)) ? + "timed out for sent delay" : "timed out for slow reply"), + req->rq_sent, req->rq_real_sent); if (imp != NULL && obd_debug_peer_on_timeout) LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer); @@ -1546,7 +1780,7 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) if (req->rq_fake) RETURN(1); - atomic_inc(&imp->imp_timeouts); + cfs_atomic_inc(&imp->imp_timeouts); /* The DLM server doesn't want recovery run on its imports. */ if (imp->imp_dlm_fake) @@ -1560,16 +1794,16 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)", ptlrpc_import_state_name(req->rq_send_state), ptlrpc_import_state_name(imp->imp_state)); - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); req->rq_status = -ETIMEDOUT; req->rq_err = 1; - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); RETURN(1); } /* if a request can't be resent we can't wait for an answer after the timeout */ - if (req->rq_no_resend) { + if (ptlrpc_no_resend(req)) { DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:"); rc = 1; } @@ -1579,10 +1813,15 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink) RETURN(rc); } +/** + * Time out all uncompleted requests in request set pointed by \a data + * Callback used when waiting on sets with l_wait_event. + * Always returns 1. + */ int ptlrpc_expired_set(void *data) { struct ptlrpc_request_set *set = data; - struct list_head *tmp; + cfs_list_t *tmp; time_t now = cfs_time_current_sec(); ENTRY; @@ -1591,9 +1830,10 @@ int ptlrpc_expired_set(void *data) /* * A timeout expired. See which reqs it applies to... */ - list_for_each (tmp, &set->set_requests) { + cfs_list_for_each (tmp, &set->set_requests) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); /* don't expire request waiting for context */ if (req->rq_wait_ctx) @@ -1622,24 +1862,32 @@ int ptlrpc_expired_set(void *data) RETURN(1); } +/** + * Sets rq_intr flag in \a req under spinlock. + */ void ptlrpc_mark_interrupted(struct ptlrpc_request *req) { - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); req->rq_intr = 1; - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); } +/** + * 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) { struct ptlrpc_request_set *set = data; - struct list_head *tmp; + cfs_list_t *tmp; LASSERT(set != NULL); - CERROR("INTERRUPTED SET %p\n", set); + CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set); - list_for_each(tmp, &set->set_requests) { + cfs_list_for_each(tmp, &set->set_requests) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); if (req->rq_phase != RQ_PHASE_RPC && req->rq_phase != RQ_PHASE_UNREGISTERING) @@ -1654,7 +1902,7 @@ void ptlrpc_interrupted_set(void *data) */ int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set) { - struct list_head *tmp; + cfs_list_t *tmp; time_t now = cfs_time_current_sec(); int timeout = 0; struct ptlrpc_request *req; @@ -1663,8 +1911,8 @@ int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set) SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */ - list_for_each(tmp, &set->set_requests) { - req = list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_for_each(tmp, &set->set_requests) { + req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain); /* * Request in-flight? @@ -1699,19 +1947,25 @@ int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set) RETURN(timeout); } +/** + * Send all unset request from the set and then wait untill all + * requests in the set complete (either get a reply, timeout, get an + * error or otherwise be interrupted). + * Returns 0 on success or error code otherwise. + */ int ptlrpc_set_wait(struct ptlrpc_request_set *set) { - struct list_head *tmp; + cfs_list_t *tmp; struct ptlrpc_request *req; struct l_wait_info lwi; int rc, timeout; ENTRY; - if (list_empty(&set->set_requests)) + if (cfs_list_empty(&set->set_requests)) RETURN(0); - list_for_each(tmp, &set->set_requests) { - req = list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_for_each(tmp, &set->set_requests) { + req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain); if (req->rq_phase == RQ_PHASE_NEW) (void)ptlrpc_send_new_req(req); } @@ -1745,6 +1999,23 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi); + /* 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()) { + cfs_sigset_t blocked_sigs = + cfs_block_sigsinv(LUSTRE_FATAL_SIGS); + + /* In fact we only interrupt for the "fatal" signals + * like SIGINT or SIGKILL. We still ignore less + * important signals since ptlrpc set is not easily + * reentrant from userspace again */ + if (cfs_signal_pending()) + ptlrpc_interrupted_set(set); + cfs_block_sigs(blocked_sigs); + } + LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT); /* -EINTR => all requests have been flagged rq_intr so next @@ -1754,13 +2025,22 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) * EINTR. * I don't really care if we go once more round the loop in * the error cases -eeb. */ - } while (rc != 0 || set->set_remaining != 0); + if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) { + cfs_list_for_each(tmp, &set->set_requests) { + req = cfs_list_entry(tmp, struct ptlrpc_request, + rq_set_chain); + cfs_spin_lock(&req->rq_lock); + req->rq_invalid_rqset = 1; + cfs_spin_unlock(&req->rq_lock); + } + } + } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0); - LASSERT(set->set_remaining == 0); + LASSERT(cfs_atomic_read(&set->set_remaining) == 0); rc = 0; - list_for_each(tmp, &set->set_requests) { - req = list_entry(tmp, struct ptlrpc_request, rq_set_chain); + cfs_list_for_each(tmp, &set->set_requests) { + req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain); LASSERT(req->rq_phase == RQ_PHASE_COMPLETE); if (req->rq_status != 0) @@ -1775,9 +2055,9 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) struct ptlrpc_set_cbdata *cbdata, *n; int err; - list_for_each_entry_safe(cbdata, n, + cfs_list_for_each_entry_safe(cbdata, n, &set->set_cblist, psc_item) { - list_del_init(&cbdata->psc_item); + cfs_list_del_init(&cbdata->psc_item); err = cbdata->psc_interpret(set, cbdata->psc_data, rc); if (err && !rc) rc = err; @@ -1788,6 +2068,14 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set) RETURN(rc); } +/** + * Helper fuction for request freeing. + * Called when request count reached zero and request needs to be freed. + * Removes request from all sorts of sending/replay lists it might be on, + * frees network buffers if any are present. + * If \a locked is set, that means caller is already holding import imp_lock + * and so we no longer need to reobtain it (for certain lists manipulations) + */ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) { ENTRY; @@ -1798,11 +2086,10 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) LASSERTF(!request->rq_receiving_reply, "req %p\n", request); LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */ - LASSERTF(list_empty(&request->rq_list), "req %p\n", request); - LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request); - LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request); + LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request); + LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request); + LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request); LASSERTF(!request->rq_replay, "req %p\n", request); - LASSERT(request->rq_cli_ctx || request->rq_fake); req_capsule_fini(&request->rq_pill); @@ -1810,14 +2097,14 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) * request->rq_reqmsg to NULL while osc_close is dereferencing it. */ if (request->rq_import != NULL) { if (!locked) - spin_lock(&request->rq_import->imp_lock); - list_del_init(&request->rq_replay_list); + cfs_spin_lock(&request->rq_import->imp_lock); + cfs_list_del_init(&request->rq_replay_list); if (!locked) - spin_unlock(&request->rq_import->imp_lock); + cfs_spin_unlock(&request->rq_import->imp_lock); } - LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request); + LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request); - if (atomic_read(&request->rq_refcount) != 0) { + if (cfs_atomic_read(&request->rq_refcount) != 0) { DEBUG_REQ(D_ERROR, request, "freeing request with nonzero refcount"); LBUG(); @@ -1850,12 +2137,22 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked) } 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. + */ void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request) { LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock); (void)__ptlrpc_req_finished(request, 1); } +/** + * 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. + */ static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked) { ENTRY; @@ -1870,9 +2167,9 @@ static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked) } DEBUG_REQ(D_INFO, request, "refcount now %u", - atomic_read(&request->rq_refcount) - 1); + cfs_atomic_read(&request->rq_refcount) - 1); - if (atomic_dec_and_test(&request->rq_refcount)) { + if (cfs_atomic_dec_and_test(&request->rq_refcount)) { __ptlrpc_free_req(request, locked); RETURN(1); } @@ -1880,21 +2177,29 @@ static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked) RETURN(0); } +/** + * Drops one reference count for a request. + */ void ptlrpc_req_finished(struct ptlrpc_request *request) { __ptlrpc_req_finished(request, 0); } +/** + * Returns xid of a \a request + */ __u64 ptlrpc_req_xid(struct ptlrpc_request *request) { return request->rq_xid; } EXPORT_SYMBOL(ptlrpc_req_xid); -/* Disengage the client's reply buffer from the network +/** + * Disengage the client's reply buffer from the network * NB does _NOT_ unregister any client-side bulk. * IDEMPOTENT, but _not_ safe against concurrent callers. * 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) { @@ -1905,7 +2210,7 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) /* * Might sleep. */ - LASSERT(!in_interrupt()); + LASSERT(!cfs_in_interrupt()); /* * Let's setup deadline for reply unlink. @@ -1969,10 +2274,17 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async) RETURN(0); } -/* caller must hold imp->imp_lock */ +/** + * Iterates through replay_list on import and prunes + * all requests have transno smaller than last_committed for the + * import and don't have rq_replay set. + * Since requests are sorted in transno order, stops when meetign first + * transno bigger than last_committed. + * caller must hold imp->imp_lock + */ void ptlrpc_free_committed(struct obd_import *imp) { - struct list_head *tmp, *saved; + cfs_list_t *tmp, *saved; struct ptlrpc_request *req; struct ptlrpc_request *last_req = NULL; /* temporary fire escape */ ENTRY; @@ -1984,7 +2296,7 @@ void ptlrpc_free_committed(struct obd_import *imp) if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked && imp->imp_generation == imp->imp_last_generation_checked) { - CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n", + CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n", imp->imp_obd->obd_name, imp->imp_peer_committed_transno); EXIT; return; @@ -1995,8 +2307,9 @@ void ptlrpc_free_committed(struct obd_import *imp) imp->imp_last_transno_checked = imp->imp_peer_committed_transno; imp->imp_last_generation_checked = imp->imp_generation; - list_for_each_safe(tmp, saved, &imp->imp_replay_list) { - req = list_entry(tmp, struct ptlrpc_request, rq_replay_list); + cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) { + req = cfs_list_entry(tmp, struct ptlrpc_request, + rq_replay_list); /* XXX ok to remove when 1357 resolved - rread 05/29/03 */ LASSERT(req != last_req); @@ -2022,15 +2335,15 @@ void ptlrpc_free_committed(struct obd_import *imp) break; } - DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")", + DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")", imp->imp_peer_committed_transno); free_req: - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); req->rq_replay = 0; - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); if (req->rq_commit_cb != NULL) req->rq_commit_cb(req); - list_del_init(&req->rq_replay_list); + cfs_list_del_init(&req->rq_replay_list); __ptlrpc_req_finished(req, 1); } @@ -2045,13 +2358,19 @@ void ptlrpc_cleanup_client(struct obd_import *imp) return; } +/** + * Schedule previously sent request for resend. + * For bulk requests we assign new xid (to avoid problems with + * lost replies and therefore several transfers landing into same buffer + * from different sending attempts). + */ void ptlrpc_resend_req(struct ptlrpc_request *req) { DEBUG_REQ(D_HA, req, "going to resend"); lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 }); req->rq_status = -EAGAIN; - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); req->rq_resend = 1; req->rq_net_err = 0; req->rq_timedout = 0; @@ -2064,7 +2383,7 @@ void ptlrpc_resend_req(struct ptlrpc_request *req) old_xid, req->rq_xid); } ptlrpc_client_wake_req(req); - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); } /* XXX: this function and rq_status are currently unused */ @@ -2073,24 +2392,31 @@ void ptlrpc_restart_req(struct ptlrpc_request *req) DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request"); req->rq_status = -ERESTARTSYS; - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); req->rq_restart = 1; req->rq_timedout = 0; ptlrpc_client_wake_req(req); - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); } +/** + * Grab additional reference on a request \a req + */ struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req) { ENTRY; - atomic_inc(&req->rq_refcount); + cfs_atomic_inc(&req->rq_refcount); RETURN(req); } +/** + * Add a request to import replay_list. + * Must be called under imp_lock + */ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, struct obd_import *imp) { - struct list_head *tmp; + cfs_list_t *tmp; LASSERT_SPIN_LOCKED(&imp->imp_lock); @@ -2104,7 +2430,7 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT); /* don't re-add requests that have been replayed */ - if (!list_empty(&req->rq_replay_list)) + if (!cfs_list_empty(&req->rq_replay_list)) return; lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY); @@ -2112,9 +2438,10 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, LASSERT(imp->imp_replayable); /* Balanced in ptlrpc_free_committed, usually. */ ptlrpc_request_addref(req); - list_for_each_prev(tmp, &imp->imp_replay_list) { + cfs_list_for_each_prev(tmp, &imp->imp_replay_list) { struct ptlrpc_request *iter = - list_entry(tmp, struct ptlrpc_request, rq_replay_list); + cfs_list_entry(tmp, struct ptlrpc_request, + rq_replay_list); /* We may have duplicate transnos if we create and then * open a file, or for closes retained if to match creating @@ -2131,13 +2458,17 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req, continue; } - list_add(&req->rq_replay_list, &iter->rq_replay_list); + cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list); return; } - list_add(&req->rq_replay_list, &imp->imp_replay_list); + cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list); } +/** + * Send request and wait until it completes. + * Returns request processing status. + */ int ptlrpc_queue_wait(struct ptlrpc_request *req) { struct ptlrpc_request_set *set; @@ -2170,6 +2501,11 @@ struct ptlrpc_replay_async_args { int praa_old_status; }; +/** + * Callback used for replayed requests reply processing. + * In case of succesful reply calls registeresd request replay callback. + * In case of error restart replay process. + */ static int ptlrpc_replay_interpret(const struct lu_env *env, struct ptlrpc_request *req, void * data, int rc) @@ -2178,7 +2514,7 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, struct obd_import *imp = req->rq_import; ENTRY; - atomic_dec(&imp->imp_replay_inflight); + cfs_atomic_dec(&imp->imp_replay_inflight); if (!ptlrpc_client_replied(req)) { CERROR("request replay timed out, restarting recovery\n"); @@ -2194,10 +2530,11 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) { /** replay was failed due to version mismatch */ DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n"); - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); imp->imp_vbr_failed = 1; imp->imp_no_lock_replay = 1; - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); + lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status); } else { /** The transno had better not change over replay. */ LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) == @@ -2208,14 +2545,23 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, lustre_msg_get_transno(req->rq_repmsg)); } - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); /** if replays by version then gap was occur on server, no trust to locks */ if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY) imp->imp_no_lock_replay = 1; imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg); - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); LASSERT(imp->imp_last_replay_transno); + /* 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, + lustre_msg_get_transno(req->rq_reqmsg)); + GOTO(out, rc = -EINVAL); + } + DEBUG_REQ(D_HA, req, "got rep"); /* let the callback do fixups, possibly including in the request */ @@ -2236,13 +2582,9 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, * Errors while replay can set transno to 0, but * imp_last_replay_transno shouldn't be set to 0 anyway */ - if (req->rq_transno > 0) { - spin_lock(&imp->imp_lock); - LASSERT(req->rq_transno <= imp->imp_last_replay_transno); - imp->imp_last_replay_transno = req->rq_transno; - spin_unlock(&imp->imp_lock); - } else + if (req->rq_transno == 0) CERROR("Transno is 0 during replay!\n"); + /* continue with recovery */ rc = ptlrpc_import_recovery_state_machine(imp); out: @@ -2250,19 +2592,22 @@ static int ptlrpc_replay_interpret(const struct lu_env *env, if (rc != 0) /* this replay failed, so restart recovery */ - ptlrpc_connect_import(imp, NULL); + ptlrpc_connect_import(imp); RETURN(rc); } +/** + * Prepares and queues request for replay. + * Adds it to ptlrpcd queue for actual sending. + * Returns 0 on success. + */ int ptlrpc_replay_req(struct ptlrpc_request *req) { struct ptlrpc_replay_async_args *aa; ENTRY; LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY); - /* Not handling automatic bulk replay yet (or ever?) */ - LASSERT(req->rq_bulk == NULL); LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args)); aa = ptlrpc_req_async_args(req); @@ -2280,58 +2625,65 @@ int ptlrpc_replay_req(struct ptlrpc_request *req) /* Readjust the timeout for current conditions */ ptlrpc_at_set_req_timeout(req); + /* Tell server the net_latency, so the server can calculate how long + * it should wait for next replay */ + lustre_msg_set_service_time(req->rq_reqmsg, + ptlrpc_at_get_net_latency(req)); DEBUG_REQ(D_HA, req, "REPLAY"); - atomic_inc(&req->rq_import->imp_replay_inflight); + cfs_atomic_inc(&req->rq_import->imp_replay_inflight); ptlrpc_request_addref(req); /* ptlrpcd needs a ref */ - ptlrpcd_add_req(req, PSCOPE_OTHER); + ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1); RETURN(0); } +/** + * Aborts all in-flight request on import \a imp sending and delayed lists + */ void ptlrpc_abort_inflight(struct obd_import *imp) { - struct list_head *tmp, *n; + cfs_list_t *tmp, *n; ENTRY; /* Make sure that no new requests get processed for this import. * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing * this flag and then putting requests on sending_list or delayed_list. */ - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); /* XXX locking? Maybe we should remove each request with the list * locked? Also, how do we know if the requests on the list are * being freed at this time? */ - list_for_each_safe(tmp, n, &imp->imp_sending_list) { + cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_list); + cfs_list_entry(tmp, struct ptlrpc_request, rq_list); DEBUG_REQ(D_RPCTRACE, req, "inflight"); - spin_lock (&req->rq_lock); + cfs_spin_lock (&req->rq_lock); if (req->rq_import_generation < imp->imp_generation) { req->rq_err = 1; req->rq_status = -EINTR; ptlrpc_client_wake_req(req); } - spin_unlock (&req->rq_lock); + cfs_spin_unlock (&req->rq_lock); } - list_for_each_safe(tmp, n, &imp->imp_delayed_list) { + cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) { struct ptlrpc_request *req = - list_entry(tmp, struct ptlrpc_request, rq_list); + cfs_list_entry(tmp, struct ptlrpc_request, rq_list); DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req"); - spin_lock (&req->rq_lock); + cfs_spin_lock (&req->rq_lock); if (req->rq_import_generation < imp->imp_generation) { req->rq_err = 1; req->rq_status = -EINTR; ptlrpc_client_wake_req(req); } - spin_unlock (&req->rq_lock); + cfs_spin_unlock (&req->rq_lock); } /* Last chance to free reqs left on the replay list, but we @@ -2339,38 +2691,43 @@ void ptlrpc_abort_inflight(struct obd_import *imp) if (imp->imp_replayable) ptlrpc_free_committed(imp); - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); EXIT; } +/** + * Abort all uncompleted requests in request set \a set + */ void ptlrpc_abort_set(struct ptlrpc_request_set *set) { - struct list_head *tmp, *pos; + cfs_list_t *tmp, *pos; LASSERT(set != NULL); - list_for_each_safe(pos, tmp, &set->set_requests) { + cfs_list_for_each_safe(pos, tmp, &set->set_requests) { struct ptlrpc_request *req = - list_entry(pos, struct ptlrpc_request, rq_set_chain); + cfs_list_entry(pos, struct ptlrpc_request, + rq_set_chain); - spin_lock(&req->rq_lock); + cfs_spin_lock(&req->rq_lock); if (req->rq_phase != RQ_PHASE_RPC) { - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); continue; } req->rq_err = 1; req->rq_status = -EINTR; ptlrpc_client_wake_req(req); - spin_unlock(&req->rq_lock); + cfs_spin_unlock(&req->rq_lock); } } static __u64 ptlrpc_last_xid; -static spinlock_t ptlrpc_last_xid_lock; +static cfs_spinlock_t ptlrpc_last_xid_lock; -/* Initialize the XID for the node. This is common among all requests on +/** + * Initialize the XID for the node. This is common among all requests on * this node, and only requires the property that it is monotonically * increasing. It does not need to be sequential. Since this is also used * as the RDMA match bits, it is important that a single client NOT have @@ -2389,9 +2746,9 @@ void ptlrpc_init_xid(void) { time_t now = cfs_time_current_sec(); - spin_lock_init(&ptlrpc_last_xid_lock); + cfs_spin_lock_init(&ptlrpc_last_xid_lock); if (now < YEAR_2004) { - ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid)); + cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid)); ptlrpc_last_xid >>= 2; ptlrpc_last_xid |= (1ULL << 61); } else { @@ -2399,23 +2756,30 @@ void ptlrpc_init_xid(void) } } +/** + * Increase xid and returns resultng new value to the caller. + */ __u64 ptlrpc_next_xid(void) { __u64 tmp; - spin_lock(&ptlrpc_last_xid_lock); + cfs_spin_lock(&ptlrpc_last_xid_lock); tmp = ++ptlrpc_last_xid; - spin_unlock(&ptlrpc_last_xid_lock); + cfs_spin_unlock(&ptlrpc_last_xid_lock); return tmp; } +/** + * Get a glimpse at what next xid value might have been. + * Returns possible next xid. + */ __u64 ptlrpc_sample_next_xid(void) { #if BITS_PER_LONG == 32 /* need to avoid possible word tearing on 32-bit systems */ __u64 tmp; - spin_lock(&ptlrpc_last_xid_lock); + cfs_spin_lock(&ptlrpc_last_xid_lock); tmp = ptlrpc_last_xid + 1; - spin_unlock(&ptlrpc_last_xid_lock); + cfs_spin_unlock(&ptlrpc_last_xid_lock); return tmp; #else /* No need to lock, since returned value is racy anyways */ @@ -2423,3 +2787,136 @@ __u64 ptlrpc_sample_next_xid(void) #endif } EXPORT_SYMBOL(ptlrpc_sample_next_xid); + +/** + * Functions for operating ptlrpc workers. + * + * A ptlrpc work is a function which will be running inside ptlrpc context. + * The callback shouldn't sleep otherwise it will block that ptlrpcd thread. + * + * 1. after a work is created, it can be used many times, that is: + * handler = ptlrpcd_alloc_work(); + * ptlrpcd_queue_work(); + * + * queue it again when necessary: + * ptlrpcd_queue_work(); + * ptlrpcd_destroy_work(); + * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but + * it will only be queued once in any time. Also as its name implies, it may + * have delay before it really runs by ptlrpcd thread. + */ +struct ptlrpc_work_async_args { + __u64 magic; + int (*cb)(const struct lu_env *, void *); + void *cbdata; +}; + +#define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */ + +static int work_interpreter(const struct lu_env *env, + struct ptlrpc_request *req, void *data, int rc) +{ + struct ptlrpc_work_async_args *arg = data; + + LASSERT(arg->magic == PTLRPC_WORK_MAGIC); + LASSERT(arg->cb != NULL); + + return arg->cb(env, arg->cbdata); +} + +/** + * Create a work for ptlrpc. + */ +void *ptlrpcd_alloc_work(struct obd_import *imp, + int (*cb)(const struct lu_env *, void *), void *cbdata) +{ + struct ptlrpc_request *req = NULL; + struct ptlrpc_work_async_args *args; + ENTRY; + + cfs_might_sleep(); + + if (cb == NULL) + RETURN(ERR_PTR(-EINVAL)); + + /* copy some code from deprecated fakereq. */ + OBD_ALLOC_PTR(req); + if (req == NULL) { + CERROR("ptlrpc: run out of memory!\n"); + RETURN(ERR_PTR(-ENOMEM)); + } + + req->rq_send_state = LUSTRE_IMP_FULL; + req->rq_type = PTL_RPC_MSG_REQUEST; + req->rq_import = class_import_get(imp); + req->rq_export = NULL; + req->rq_interpret_reply = work_interpreter; + /* don't want reply */ + req->rq_receiving_reply = 0; + req->rq_must_unlink = 0; + req->rq_no_delay = req->rq_no_resend = 1; + + cfs_spin_lock_init(&req->rq_lock); + CFS_INIT_LIST_HEAD(&req->rq_list); + CFS_INIT_LIST_HEAD(&req->rq_replay_list); + CFS_INIT_LIST_HEAD(&req->rq_set_chain); + CFS_INIT_LIST_HEAD(&req->rq_history_list); + CFS_INIT_LIST_HEAD(&req->rq_exp_list); + cfs_waitq_init(&req->rq_reply_waitq); + cfs_waitq_init(&req->rq_set_waitq); + cfs_atomic_set(&req->rq_refcount, 1); + + CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args)); + args = ptlrpc_req_async_args(req); + args->magic = PTLRPC_WORK_MAGIC; + args->cb = cb; + args->cbdata = cbdata; + + RETURN(req); +} +EXPORT_SYMBOL(ptlrpcd_alloc_work); + +void ptlrpcd_destroy_work(void *handler) +{ + struct ptlrpc_request *req = handler; + + if (req) + ptlrpc_req_finished(req); +} +EXPORT_SYMBOL(ptlrpcd_destroy_work); + +int ptlrpcd_queue_work(void *handler) +{ + struct ptlrpc_request *req = handler; + + /* + * Check if the req is already being queued. + * + * Here comes a trick: it lacks a way of checking if a req is being + * processed reliably in ptlrpc. Here I have to use refcount of req + * for this purpose. This is okay because the caller should use this + * req as opaque data. - Jinshan + */ + LASSERT(cfs_atomic_read(&req->rq_refcount) > 0); + if (cfs_atomic_read(&req->rq_refcount) > 1) + return -EBUSY; + + if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */ + cfs_atomic_dec(&req->rq_refcount); + return -EBUSY; + } + + /* re-initialize the req */ + req->rq_timeout = obd_timeout; + req->rq_sent = cfs_time_current_sec(); + 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); + return 0; +} +EXPORT_SYMBOL(ptlrpcd_queue_work);