*
*/
-#define EXPORT_SYMTAB
#define DEBUG_SUBSYSTEM S_RPC
+#include <linux/obd_support.h>
+#include <linux/obd_class.h>
#include <linux/lustre_net.h>
-extern int request_in_callback(ptl_event_t *ev, void *data);
-extern int ptl_handled_rpc(struct ptlrpc_service *service, void *start);
+extern int request_in_callback(ptl_event_t *ev);
static int ptlrpc_check_event(struct ptlrpc_service *svc,
struct ptlrpc_thread *thread, ptl_event_t *event)
ENTRY;
spin_lock(&svc->srv_lock);
- if (sigismember(&(current->pending.signal), SIGKILL) ||
- sigismember(&(current->pending.signal), SIGTERM) ||
- sigismember(&(current->pending.signal), SIGINT)) {
- thread->t_flags |= SVC_KILLED;
- GOTO(out, rc = 1);
- }
-
if (thread->t_flags & SVC_STOPPING)
GOTO(out, rc = 1);
+ LASSERT ((thread->t_flags & SVC_EVENT) == 0);
+
if (ptl_is_valid_handle(&svc->srv_eq_h)) {
int err;
err = PtlEQGet(svc->srv_eq_h, event);
}
if (err != PTL_EQ_EMPTY) {
- CDEBUG(D_NET, "BUG: PtlEQGet returned %d\n", rc);
+ CERROR("BUG: PtlEQGet returned %d\n", rc);
LBUG();
}
}
struct ptlrpc_service *
-ptlrpc_init_svc(__u32 bufsize, int req_portal, int rep_portal, char *uuid,
- svc_handler_t handler)
+ptlrpc_init_svc(__u32 bufsize, int nbuffs, int req_portal, int rep_portal,
+ obd_uuid_t uuid, svc_handler_t handler, char *name)
{
int err;
int rc, i;
RETURN(NULL);
}
+ service->srv_name = name;
spin_lock_init(&service->srv_lock);
INIT_LIST_HEAD(&service->srv_reqs);
INIT_LIST_HEAD(&service->srv_threads);
init_waitqueue_head(&service->srv_waitq);
service->srv_buf_size = bufsize;
+ service->srv_nbuffs = nbuffs;
service->srv_rep_portal = rep_portal;
service->srv_req_portal = req_portal;
service->srv_handler = handler;
err = kportal_uuid_to_peer(uuid, &service->srv_self);
if (err) {
- CERROR("cannot get peer for uuid '%s'", uuid);
- GOTO(err_free, NULL);
+ CERROR("cannot get peer for uuid '%s'\n", uuid);
+ OBD_FREE(service, sizeof(*service));
+ RETURN(NULL);
}
- service->srv_ring_length = RPC_RING_LENGTH;
-
- rc = PtlEQAlloc(service->srv_self.peer_ni, 128, request_in_callback,
- service, &(service->srv_eq_h));
+ /* NB We need exactly 1 event for each buffer we queue */
+ rc = PtlEQAlloc(service->srv_self.peer_ni, service->srv_nbuffs,
+ request_in_callback, &(service->srv_eq_h));
if (rc != PTL_OK) {
CERROR("PtlEQAlloc failed: %d\n", rc);
LBUG();
- GOTO(err_free, NULL);
+ OBD_FREE(service, sizeof(*service));
+ RETURN(NULL);
+ }
+
+ OBD_ALLOC(service->srv_rqbds, service->srv_nbuffs *
+ sizeof(struct ptlrpc_request_buffer_desc));
+ if (service->srv_rqbds == NULL) {
+ CERROR("no memory\n");
+ LBUG();
+ GOTO(failed, NULL);
}
- for (i = 0; i < service->srv_ring_length; i++) {
- OBD_ALLOC(service->srv_buf[i], service->srv_buf_size);
- if (service->srv_buf[i] == NULL) {
+ for (i = 0; i < service->srv_nbuffs; i++) {
+ struct ptlrpc_request_buffer_desc *rqbd =&service->srv_rqbds[i];
+
+ rqbd->rqbd_service = service;
+ ptl_set_inv_handle (&rqbd->rqbd_me_h);
+ OBD_ALLOC(rqbd->rqbd_buffer, service->srv_buf_size);
+ if (rqbd->rqbd_buffer == NULL) {
CERROR("no memory\n");
LBUG();
- GOTO(err_ring, NULL);
+ GOTO(failed, NULL);
}
- service->srv_ref_count[i] = 0;
- ptlrpc_link_svc_me(service, i);
+ ptlrpc_link_svc_me(rqbd);
}
CDEBUG(D_NET, "Starting service listening on portal %d\n",
service->srv_req_portal);
RETURN(service);
-err_ring:
- service->srv_ring_length = i;
- rpc_unregister_service(service); // XXX verify this is right
- PtlEQFree(service->srv_eq_h);
-err_free:
- OBD_FREE(service, sizeof(*service));
+failed:
+ ptlrpc_unregister_service(service);
return NULL;
}
static int handle_incoming_request(struct obd_device *obddev,
struct ptlrpc_service *svc,
- ptl_event_t *event)
+ ptl_event_t *event,
+ struct ptlrpc_request *request)
{
- struct ptlrpc_request request;
- struct lustre_peer peer;
- void *start;
+ struct ptlrpc_request_buffer_desc *rqbd = event->mem_desc.user_ptr;
int rc;
/* FIXME: If we move to an event-driven model, we should put the request
* on the stack of mds_handle instead. */
- start = event->mem_desc.start;
- memset(&request, 0, sizeof(request));
- request.rq_obd = obddev;
- request.rq_reqmsg = event->mem_desc.start + event->offset;
- request.rq_reqlen = event->mem_desc.length;
+ LASSERT ((event->mem_desc.options & PTL_MD_IOV) == 0);
+ LASSERT (rqbd->rqbd_service == svc);
+ LASSERT (rqbd->rqbd_buffer == event->mem_desc.start);
+ LASSERT (event->offset == 0);
+
+ request->rq_svc = svc;
+ request->rq_obd = obddev;
+ request->rq_xid = event->match_bits;
+ request->rq_reqmsg = event->mem_desc.start + event->offset;
+ request->rq_reqlen = event->mem_desc.length;
+
+ if (request->rq_reqlen < sizeof(struct lustre_msg)) {
+ CERROR("incomplete request (%d): ptl %d from "LPX64" xid "LPD64"\n",
+ request->rq_reqlen, svc->srv_req_portal,
+ event->initiator.nid, request->rq_xid);
+ spin_unlock(&svc->srv_lock);
+ RETURN(-EINVAL);
+ }
- if (request.rq_reqmsg->xid != event->match_bits)
+ if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
+ CERROR("wrong packet type received (type=%u)\n",
+ request->rq_reqmsg->type);
LBUG();
+ spin_unlock(&svc->srv_lock);
+ RETURN(-EINVAL);
+ }
+
+ if (request->rq_reqmsg->magic != PTLRPC_MSG_MAGIC) {
+ CERROR("wrong lustre_msg magic %d: ptl %d from "LPX64" xid "LPD64"\n",
+ request->rq_reqmsg->magic, svc->srv_req_portal,
+ event->initiator.nid, request->rq_xid);
+ spin_unlock(&svc->srv_lock);
+ RETURN(-EINVAL);
+ }
- CDEBUG(D_NET, "got req %d\n", request.rq_reqmsg->xid);
+ if (request->rq_reqmsg->version != PTLRPC_MSG_VERSION) {
+ CERROR("wrong lustre_msg version %d: ptl %d from "LPX64" xid "LPD64"\n",
+ request->rq_reqmsg->version, svc->srv_req_portal,
+ event->initiator.nid, request->rq_xid);
+ spin_unlock(&svc->srv_lock);
+ RETURN(-EINVAL);
+ }
- peer.peer_nid = event->initiator.nid;
+ CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
+
+ request->rq_peer.peer_nid = event->initiator.nid;
/* FIXME: this NI should be the incoming NI.
* We don't know how to find that from here. */
- peer.peer_ni = svc->srv_self.peer_ni;
-
- if (request.rq_reqmsg->conn) {
- request.rq_connection =
- (void *)(unsigned long)request.rq_reqmsg->conn;
- if (request.rq_reqmsg->token !=
- request.rq_connection->c_token) {
- struct ptlrpc_connection *tmp;
- tmp = ptlrpc_get_connection(&peer);
- CERROR("rq_reqmsg->conn: %p\n", request.rq_connection);
- CERROR("real connection: %p\n", tmp);
- CERROR("rq_reqmsg->token: %Lu\n",
- (unsigned long long)request.rq_reqmsg->token);
- CERROR("real token : %Lu\n",
- (unsigned long long)tmp->c_token);
- LBUG();
- }
- ptlrpc_connection_addref(request.rq_connection);
- } else {
- /*
- PHIL? should we perhaps only do this when
- we get an incoming connmgr_connect request?
- */
- request.rq_connection = ptlrpc_get_connection(&peer);
- if (!request.rq_connection)
- LBUG();
+ request->rq_peer.peer_ni = svc->srv_self.peer_ni;
+
+ request->rq_export = class_conn2export((struct lustre_handle *)
+ request->rq_reqmsg);
+
+ if (request->rq_export) {
+ request->rq_connection = request->rq_export->exp_connection;
+ ptlrpc_connection_addref(request->rq_connection);
}
spin_unlock(&svc->srv_lock);
- rc = svc->srv_handler(obddev, svc, &request);
- ptlrpc_put_connection(request.rq_connection);
- ptl_handled_rpc(svc, start);
+
+ rc = svc->srv_handler(request);
+ ptlrpc_put_connection(request->rq_connection);
+
+ ptlrpc_link_svc_me (rqbd);
return rc;
}
struct obd_device *obddev = data->dev;
struct ptlrpc_service *svc = data->svc;
struct ptlrpc_thread *thread = data->thread;
+ struct ptlrpc_request *request;
+ ptl_event_t *event;
ENTRY;
spin_unlock_irq(¤t->sigmask_lock);
sprintf(current->comm, data->name);
+ unlock_kernel();
/* Record that the thread is running */
thread->t_flags = SVC_RUNNING;
/* XXX maintain a list of all managed devices: insert here */
+ OBD_ALLOC(event, sizeof(*event));
+ LASSERT(event);
+ OBD_ALLOC(request, sizeof(*request));
+ LASSERT(request);
+
/* And now, loop forever on requests */
while (1) {
- ptl_event_t event;
-
wait_event(svc->srv_waitq,
- ptlrpc_check_event(svc, thread, &event));
+ ptlrpc_check_event(svc, thread, event));
spin_lock(&svc->srv_lock);
- if (thread->t_flags & SVC_SIGNAL) {
- thread->t_flags &= ~SVC_SIGNAL;
- spin_unlock(&svc->srv_lock);
- EXIT;
- break;
- }
if (thread->t_flags & SVC_STOPPING) {
thread->t_flags &= ~SVC_STOPPING;
EXIT;
break;
}
-
- if (thread->t_flags & SVC_EVENT) {
- thread->t_flags &= ~SVC_EVENT;
- rc = handle_incoming_request(obddev, svc, &event);
+
+ if (thread->t_flags & SVC_EVENT) {
+ LASSERT (event->sequence != 0);
+ rc = handle_incoming_request(obddev, svc, event,
+ request);
thread->t_flags &= ~SVC_EVENT;
continue;
}
break;
}
+ OBD_FREE(event, sizeof(*event));
+ OBD_FREE(request, sizeof(*request));
+
thread->t_flags = SVC_STOPPED;
wake_up(&thread->t_ctl_waitq);
CDEBUG(D_NET, "service thread exiting, process %d\n", current->pid);
static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
struct ptlrpc_thread *thread)
{
+ spin_lock(&svc->srv_lock);
thread->t_flags = SVC_STOPPING;
+ spin_unlock(&svc->srv_lock);
wake_up(&svc->srv_waitq);
- wait_event_interruptible(thread->t_ctl_waitq,
- (thread->t_flags & SVC_STOPPED));
+ wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED));
}
void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
CLONE_VM | CLONE_FS | CLONE_FILES);
if (rc < 0) {
CERROR("cannot start thread\n");
+ OBD_FREE(thread, sizeof(*thread));
RETURN(-EINVAL);
}
wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
RETURN(0);
}
-int rpc_unregister_service(struct ptlrpc_service *service)
+int ptlrpc_unregister_service(struct ptlrpc_service *service)
{
int rc, i;
- for (i = 0; i < service->srv_ring_length; i++) {
- if (ptl_is_valid_handle(&(service->srv_me_h[i]))) {
- rc = PtlMEUnlink(service->srv_me_h[i]);
+ /* NB service->srv_nbuffs gets set before we attempt (and possibly
+ * fail) to allocate srv_rqbds.
+ */
+ if (service->srv_rqbds != NULL) {
+ for (i = 0; i < service->srv_nbuffs; i++) {
+ struct ptlrpc_request_buffer_desc *rqbd =
+ &service->srv_rqbds[i];
+
+ if (rqbd->rqbd_buffer == NULL) /* no buffer allocated */
+ continue; /* => never initialised */
+
+ /* Buffer allocated => got linked */
+ LASSERT (ptl_is_valid_handle (&rqbd->rqbd_me_h));
+
+ rc = PtlMEUnlink(rqbd->rqbd_me_h);
if (rc)
CERROR("PtlMEUnlink failed: %d\n", rc);
- ptl_set_inv_handle(&(service->srv_me_h[i]));
+
+ OBD_FREE(rqbd->rqbd_buffer, service->srv_buf_size);
}
- if (service->srv_buf[i] != NULL)
- OBD_FREE(service->srv_buf[i], service->srv_buf_size);
- service->srv_buf[i] = NULL;
+ OBD_FREE(service->srv_rqbds, service->srv_nbuffs *
+ sizeof (struct ptlrpc_request_buffer_desc));
}
rc = PtlEQFree(service->srv_eq_h);
if (rc)
CERROR("PtlEQFree failed: %d\n", rc);
- return 0;
+ if (!list_empty(&service->srv_reqs)) {
+ // XXX reply with errors and clean up
+ CERROR("Request list not empty!\n");
+ rc = -EBUSY;
+ }
+
+ OBD_FREE(service, sizeof(*service));
+ if (rc)
+ LBUG();
+ return rc;
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff