-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
* 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 <errno.h>
#include "ptlrpc_internal.h"
+static int ptlrpc_send_new_req(struct ptlrpc_request *req);
+
+/**
+ * Initialize passed in client structure \a cl.
+ */
void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
struct ptlrpc_client *cl)
{
cl->cli_reply_portal = rep_portal;
cl->cli_name = name;
}
+EXPORT_SYMBOL(ptlrpc_init_client);
+/**
+ * Return PortalRPC connection for remore uud \a uuid
+ */
struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
{
struct ptlrpc_connection *c;
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;
}
return c;
}
+EXPORT_SYMBOL(ptlrpc_uuid_to_connection);
-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;
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)
{
return desc;
}
+EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
-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;
-}
-
-void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
- cfs_page_t *page, int pageoffset, int len)
+/**
+ * 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, int pin)
{
LASSERT(desc->bd_iov_count < desc->bd_max_iov);
LASSERT(page != NULL);
desc->bd_nob += len;
+ if (pin)
+ cfs_page_pin(page);
+
ptlrpc_add_bulk_page(desc, page, pageoffset, len);
}
+EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
-void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
+/**
+ * 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 i;
ENTRY;
LASSERT(desc != NULL);
else
class_import_put(desc->bd_import);
+ if (unpin) {
+ 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;
}
+EXPORT_SYMBOL(__ptlrpc_free_bulk);
-/* 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;
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 {
reqmsg*/
lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
}
+EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
/* Adjust max service estimate based on server value */
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",
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",
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
*/
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)
{
cfs_list_t *l, *tmp;
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));
}
cfs_spin_unlock(&pool->prp_lock);
OBD_FREE(pool, sizeof(*pool));
}
+EXPORT_SYMBOL(ptlrpc_free_rq_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(cfs_list_empty(&pool->prp_req_list) ||
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;
cfs_spin_unlock(&pool->prp_lock);
return;
}
+EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
+/**
+ * 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))
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);
}
return pool;
}
+EXPORT_SYMBOL(ptlrpc_init_rq_pool);
+/**
+ * Fetches one request from pool \a pool
+ */
static struct ptlrpc_request *
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;
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();
cfs_atomic_set(&request->rq_refcount, 1);
}
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);
-}
+ int rc;
+ rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
+ if (rc)
+ return rc;
+
+ /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
+ * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
+ * have to send old ptlrpc_body to keep interoprability with these
+ * clients.
+ *
+ * Only three kinds of server->client RPCs so far:
+ * - LDLM_BL_CALLBACK
+ * - LDLM_CP_CALLBACK
+ * - LDLM_GL_CALLBACK
+ *
+ * XXX This should be removed whenever we drop the interoprability with
+ * the these old clients.
+ */
+ if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
+ opcode == LDLM_GL_CALLBACK)
+ req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
+ sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
+
+ return rc;
+}
+EXPORT_SYMBOL(ptlrpc_request_pack);
+/**
+ * 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)
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,
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);
}
+EXPORT_SYMBOL(ptlrpc_request_alloc);
+/**
+ * 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)
{
return ptlrpc_request_alloc_internal(imp, pool, format);
}
+EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
+/**
+ * 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)
else
OBD_FREE_PTR(request);
}
+EXPORT_SYMBOL(ptlrpc_request_free);
+/**
+ * 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)
}
return req;
}
+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,
}
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 "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)
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();
cfs_atomic_set(&request->rq_refcount, 1);
RETURN(request);
}
+EXPORT_SYMBOL(ptlrpc_prep_fakereq);
+/**
+ * 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);
- cfs_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);
+}
+EXPORT_SYMBOL(ptlrpc_fakereq_finished);
+/**
+ * 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;
+ struct ptlrpc_request_set *set;
+
+ ENTRY;
+ 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);
+ 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);
+ set->set_max_inflight = UINT_MAX;
+ set->set_producer = NULL;
+ set->set_producer_arg = NULL;
+ set->set_rc = 0;
+
+ RETURN(set);
+}
+EXPORT_SYMBOL(ptlrpc_prep_set);
- ENTRY;
- OBD_ALLOC(set, sizeof *set);
- if (!set)
- RETURN(NULL);
- CFS_INIT_LIST_HEAD(&set->set_requests);
- cfs_waitq_init(&set->set_waitq);
- 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);
+/**
+ * Allocate and initialize new request set structure with flow control
+ * extension. This extension allows to control the number of requests in-flight
+ * for the whole set. A callback function to generate requests must be provided
+ * and the request set will keep the number of requests sent over the wire to
+ * @max_inflight.
+ * Returns a pointer to the newly allocated set structure or NULL on error.
+ */
+struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
+ void *arg)
+
+{
+ struct ptlrpc_request_set *set;
+
+ set = ptlrpc_prep_set();
+ if (!set)
+ RETURN(NULL);
+
+ set->set_max_inflight = max;
+ set->set_producer = func;
+ set->set_producer_arg = arg;
- RETURN(set);
+ RETURN(set);
}
+EXPORT_SYMBOL(ptlrpc_prep_fcset);
-/* 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)
{
cfs_list_t *tmp;
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;
cfs_list_for_each (tmp, &set->set_requests) {
struct ptlrpc_request *req =
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);
cfs_list_for_each_safe(tmp, next, &set->set_requests) {
struct ptlrpc_request *req =
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;
}
+EXPORT_SYMBOL(ptlrpc_set_destroy);
+/**
+ * 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)
{
RETURN(0);
}
+EXPORT_SYMBOL(ptlrpc_set_add_cb);
+/**
+ * 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)
{
- /* The set takes over the caller's request reference */
- cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
- req->rq_set = set;
- set->set_remaining++;
+ LASSERT(cfs_list_empty(&req->rq_set_chain));
+
+ /* The set takes over the caller's request reference */
+ cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
+ req->rq_set = set;
+ cfs_atomic_inc(&set->set_remaining);
+ req->rq_queued_time = cfs_time_current();
+
+ if (req->rq_reqmsg != NULL)
+ lustre_msg_set_jobid(req->rq_reqmsg, NULL);
+
+ if (set->set_producer != NULL)
+ /* If the request set has a producer callback, the RPC must be
+ * sent straight away */
+ ptlrpc_send_new_req(req);
}
+EXPORT_SYMBOL(ptlrpc_set_add_req);
/**
- * 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 (cfs_test_bit(LIOD_STOP, &pc->pc_flags))
- return -EALREADY;
+ LASSERT(req->rq_set == NULL);
+ LASSERT(cfs_test_bit(LIOD_STOP, &pc->pc_flags) == 0);
cfs_spin_lock(&set->set_new_req_lock);
/*
* The set takes over the caller's request reference.
*/
- cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
req->rq_set = set;
+ 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);
+ }
}
+EXPORT_SYMBOL(ptlrpc_set_add_new_req);
-/*
+/**
* Based on the current state of the import, determine if the request
* can be sent, is an error, or should be delayed.
*
} else if (imp->imp_state == LUSTRE_IMP_NEW) {
DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
*status = -EIO;
- LBUG();
} 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 */ ;
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 */
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)) &&
return 1;
}
+/**
+ * Check request processing status.
+ * Returns the status.
+ */
static int ptlrpc_check_status(struct ptlrpc_request *req)
{
int err;
if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
struct obd_import *imp = req->rq_import;
__u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
- LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
- " with %s. The %s operation failed with %d\n",
- libcfs_nid2str(imp->imp_connection->c_peer.nid),
- ll_opcode2str(opc), err);
- RETURN(err < 0 ? err : -EINVAL);
- }
-
- if (err < 0) {
- DEBUG_REQ(D_INFO, req, "status is %d", err);
- } else if (err > 0) {
- /* XXX: translate this error from net to host */
- DEBUG_REQ(D_INFO, req, "status is %d", err);
- }
-
- if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
- struct obd_import *imp = req->rq_import;
- __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
-
if (ptlrpc_console_allow(req))
LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
"communicating with %s. The %s "
libcfs_nid2str(
imp->imp_connection->c_peer.nid),
ll_opcode2str(opc), err);
-
RETURN(err < 0 ? err : -EINVAL);
}
+ if (err < 0) {
+ DEBUG_REQ(D_INFO, req, "status is %d", err);
+ } else if (err > 0) {
+ /* XXX: translate this error from net to host */
+ DEBUG_REQ(D_INFO, req, "status is %d", err);
+ }
+
RETURN(err);
}
/**
- * 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)
{
/**
* 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)
{
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);
}
* 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);
RETURN(rc);
}
+ /* 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();
+
+ DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
+ req->rq_resend = 1;
+ req->rq_nr_resend++;
+
+ /* Readjust the timeout for current conditions */
+ ptlrpc_at_set_req_timeout(req);
+ /* delay resend to give a chance to the server to get ready.
+ * The delay is increased by 1s on every resend and is capped to
+ * the current request timeout (i.e. obd_timeout if AT is off,
+ * or AT service time x 125% + 5s, see at_est2timeout) */
+ if (req->rq_nr_resend > req->rq_timeout)
+ req->rq_sent = now + req->rq_timeout;
+ else
+ req->rq_sent = now + req->rq_nr_resend;
+ }
+
/*
* Security layer unwrap might ask resend this request.
*/
}
if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
- OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
+ 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));
lustre_msg_get_last_committed(req->rq_repmsg);
}
ptlrpc_free_committed(imp);
+
+ 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;
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)) {
cfs_spin_lock(&req->rq_lock);
RETURN(0);
}
-/* this sends any unsent RPCs in @set and returns TRUE if all are sent */
+static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
+{
+ int remaining, rc;
+ ENTRY;
+
+ LASSERT(set->set_producer != NULL);
+
+ remaining = cfs_atomic_read(&set->set_remaining);
+
+ /* populate the ->set_requests list with requests until we
+ * reach the maximum number of RPCs in flight for this set */
+ while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
+ rc = set->set_producer(set, set->set_producer_arg);
+ if (rc == -ENOENT) {
+ /* no more RPC to produce */
+ set->set_producer = NULL;
+ set->set_producer_arg = NULL;
+ RETURN(0);
+ }
+ }
+
+ RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
+}
+
+/**
+ * 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)
{
- cfs_list_t *tmp;
+ cfs_list_t *tmp, *next;
int force_timer_recalc = 0;
ENTRY;
- if (set->set_remaining == 0)
+ if (cfs_atomic_read(&set->set_remaining) == 0)
RETURN(1);
- cfs_list_for_each(tmp, &set->set_requests) {
+ cfs_list_for_each_safe(tmp, next, &set->set_requests) {
struct ptlrpc_request *req =
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 &&
/* delayed send - skip */
if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
- continue;
+ continue;
+
+ /* delayed resend - skip */
+ if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
+ req->rq_sent > cfs_time_current_sec())
+ continue;
if (!(req->rq_phase == RQ_PHASE_RPC ||
req->rq_phase == RQ_PHASE_BULK ||
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);
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);
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;
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);
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;
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)
continue;
* 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);
}
* 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);
/* 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;
ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
- CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
- "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
- imp->imp_obd->obd_uuid.uuid,
- req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
- req->rq_xid,
- libcfs_nid2str(imp->imp_connection->c_peer.nid),
- req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
+ CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
+ "Completed RPC pname:cluuid:pid:xid:nid:"
+ "opc %s:%s:%d:"LPU64":%s:%d\n",
+ cfs_curproc_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));
cfs_spin_lock(&imp->imp_lock);
/* Request already may be not on sending or delaying list. This
}
cfs_spin_unlock(&imp->imp_lock);
- set->set_remaining--;
+ cfs_atomic_dec(&set->set_remaining);
cfs_waitq_broadcast(&imp->imp_recovery_waitq);
+
+ if (set->set_producer) {
+ /* produce a new request if possible */
+ if (ptlrpc_set_producer(set) > 0)
+ force_timer_recalc = 1;
+
+ /* free the request that has just been completed
+ * in order not to pollute set->set_requests */
+ cfs_list_del_init(&req->rq_set_chain);
+ cfs_spin_lock(&req->rq_lock);
+ req->rq_set = NULL;
+ req->rq_invalid_rqset = 0;
+ cfs_spin_unlock(&req->rq_lock);
+
+ /* record rq_status to compute the final status later */
+ if (req->rq_status != 0)
+ set->set_rc = req->rq_status;
+ ptlrpc_req_finished(req);
+ }
}
/* 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);
}
+EXPORT_SYMBOL(ptlrpc_check_set);
-/* 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;
req->rq_timedout = 1;
cfs_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));
+ 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);
/* 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;
}
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;
*/
RETURN(1);
}
+EXPORT_SYMBOL(ptlrpc_expired_set);
+/**
+ * Sets rq_intr flag in \a req under spinlock.
+ */
void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
{
cfs_spin_lock(&req->rq_lock);
req->rq_intr = 1;
cfs_spin_unlock(&req->rq_lock);
}
+EXPORT_SYMBOL(ptlrpc_mark_interrupted);
+/**
+ * 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;
cfs_list_t *tmp;
LASSERT(set != NULL);
- CERROR("INTERRUPTED SET %p\n", set);
+ CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
cfs_list_for_each(tmp, &set->set_requests) {
struct ptlrpc_request *req =
ptlrpc_mark_interrupted(req);
}
}
+EXPORT_SYMBOL(ptlrpc_interrupted_set);
/**
* Get the smallest timeout in the set; this does NOT set a timeout.
if (req->rq_phase == RQ_PHASE_NEW)
deadline = req->rq_sent;
+ else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
+ deadline = req->rq_sent;
else
deadline = req->rq_sent + req->rq_timeout;
}
RETURN(timeout);
}
+EXPORT_SYMBOL(ptlrpc_set_next_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)
{
cfs_list_t *tmp;
int rc, timeout;
ENTRY;
+ if (set->set_producer)
+ (void)ptlrpc_set_producer(set);
+ else
+ 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);
+ }
+
if (cfs_list_empty(&set->set_requests))
RETURN(0);
- 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);
- }
-
do {
timeout = ptlrpc_set_next_timeout(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_restore_sigs(blocked_sigs);
+ }
+
LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
/* -EINTR => all requests have been flagged rq_intr so next
* 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;
+ rc = set->set_rc; /* rq_status of already freed requests if any */
cfs_list_for_each(tmp, &set->set_requests) {
req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
RETURN(rc);
}
+EXPORT_SYMBOL(ptlrpc_set_wait);
+/**
+ * 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;
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);
}
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);
}
+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.
+ */
static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
{
ENTRY;
RETURN(0);
}
+/**
+ * Drops one reference count for a request.
+ */
void ptlrpc_req_finished(struct ptlrpc_request *request)
{
__ptlrpc_req_finished(request, 0);
}
+EXPORT_SYMBOL(ptlrpc_req_finished);
+/**
+ * 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)
{
}
RETURN(0);
}
+EXPORT_SYMBOL(ptlrpc_unregister_reply);
-/* 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)
{
cfs_list_t *tmp, *saved;
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;
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:
cfs_spin_lock(&req->rq_lock);
EXIT;
return;
}
+EXPORT_SYMBOL(ptlrpc_cleanup_client);
+/**
+ * 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");
ptlrpc_client_wake_req(req);
cfs_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);
cfs_spin_unlock(&req->rq_lock);
}
+EXPORT_SYMBOL(ptlrpc_restart_req);
+/**
+ * Grab additional reference on a request \a req
+ */
struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
{
ENTRY;
cfs_atomic_inc(&req->rq_refcount);
RETURN(req);
}
+EXPORT_SYMBOL(ptlrpc_request_addref);
+/**
+ * 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)
{
cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
}
+EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
+/**
+ * Send request and wait until it completes.
+ * Returns request processing status.
+ */
int ptlrpc_queue_wait(struct ptlrpc_request *req)
{
struct ptlrpc_request_set *set;
RETURN(rc);
}
+EXPORT_SYMBOL(ptlrpc_queue_wait);
struct ptlrpc_replay_async_args {
int praa_old_state;
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)
imp->imp_vbr_failed = 1;
imp->imp_no_lock_replay = 1;
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) ==
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 */
* 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) {
- cfs_spin_lock(&imp->imp_lock);
- LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
- imp->imp_last_replay_transno = req->rq_transno;
- cfs_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:
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);
/* 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");
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);
}
+EXPORT_SYMBOL(ptlrpc_replay_req);
+/**
+ * Aborts all in-flight request on import \a imp sending and delayed lists
+ */
void ptlrpc_abort_inflight(struct obd_import *imp)
{
cfs_list_t *tmp, *n;
cfs_spin_lock (&req->rq_lock);
if (req->rq_import_generation < imp->imp_generation) {
req->rq_err = 1;
- req->rq_status = -EINTR;
+ req->rq_status = -EIO;
ptlrpc_client_wake_req(req);
}
cfs_spin_unlock (&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;
+ req->rq_status = -EIO;
ptlrpc_client_wake_req(req);
}
cfs_spin_unlock (&req->rq_lock);
EXIT;
}
+EXPORT_SYMBOL(ptlrpc_abort_inflight);
+/**
+ * Abort all uncompleted requests in request set \a set
+ */
void ptlrpc_abort_set(struct ptlrpc_request_set *set)
{
cfs_list_t *tmp, *pos;
static __u64 ptlrpc_last_xid;
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
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 {
}
}
+/**
+ * Increase xid and returns resultng new value to the caller.
+ */
__u64 ptlrpc_next_xid(void)
{
__u64 tmp;
cfs_spin_unlock(&ptlrpc_last_xid_lock);
return tmp;
}
+EXPORT_SYMBOL(ptlrpc_next_xid);
+/**
+ * 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
#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);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff