* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
*/
-#include <linux/config.h>
-#include <linux/module.h>
-#include <linux/kmod.h>
-
#define DEBUG_SUBSYSTEM S_RPC
-
+#ifdef __KERNEL__
+# include <linux/config.h>
+# include <linux/module.h>
+# include <linux/kmod.h>
+#else
+# include <liblustre.h>
+#endif
+
+#include <linux/obd_support.h>
#include <linux/lustre_ha.h>
#include <linux/lustre_net.h>
+#include <linux/lustre_import.h>
+#include <linux/lustre_export.h>
#include <linux/obd.h>
+#include <linux/obd_ost.h>
+#include <linux/obd_class.h>
+#include <linux/obd_lov.h> /* for IOC_LOV_SET_OSC_ACTIVE */
+
+#include "ptlrpc_internal.h"
-int ptlrpc_reconnect_import(struct obd_import *imp, int rq_opc,
- struct ptlrpc_request **reqptr)
+static int ptlrpc_recover_import_no_retry(struct obd_import *, char *);
+
+void ptlrpc_run_recovery_over_upcall(struct obd_device *obd)
{
- struct obd_device *obd = imp->imp_obd;
- struct client_obd *cli = &obd->u.cli;
- int size[] = { sizeof(cli->cl_target_uuid), sizeof(obd->obd_uuid) };
- char *tmp[] = {cli->cl_target_uuid.uuid, obd->obd_uuid.uuid};
- struct ptlrpc_connection *conn = imp->imp_connection;
- struct ptlrpc_request *req;
- struct obd_export *ldlmexp;
- struct lustre_handle old_hdl;
+ char *argv[4];
+ char *envp[3];
int rc;
+ ENTRY;
- req = ptlrpc_prep_req(imp, rq_opc, 2, size, tmp);
- if (!req)
- RETURN(-ENOMEM);
- req->rq_level = LUSTRE_CONN_NEW;
- req->rq_replen = lustre_msg_size(0, NULL);
- /*
- * This address is the export that represents our client-side LDLM
- * service (for ASTs). We should only have one on this list, so we
- * just grab the first one.
- *
- * XXX tear down export, call class_obd_connect?
- */
- ldlmexp = list_entry(obd->obd_exports.next, struct obd_export,
- exp_obd_chain);
- req->rq_reqmsg->addr = (__u64)(unsigned long)ldlmexp;
- req->rq_reqmsg->cookie = ldlmexp->exp_cookie;
- rc = ptlrpc_queue_wait(req);
- if (rc) {
- CERROR("cannot connect to %s@%s: rc = %d\n",
- cli->cl_target_uuid.uuid, conn->c_remote_uuid.uuid, rc);
- GOTO(out_disc, rc);
- }
- if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT) {
- memset(&old_hdl, 0, sizeof(old_hdl));
- if (!memcmp(&old_hdl.addr, &req->rq_repmsg->addr,
- sizeof (old_hdl.addr)) &&
- !memcmp(&old_hdl.cookie, &req->rq_repmsg->cookie,
- sizeof (old_hdl.cookie))) {
- CERROR("%s@%s didn't like our handle "LPX64"/"LPX64
- ", failed\n", cli->cl_target_uuid.uuid,
- conn->c_remote_uuid.uuid,
- (__u64)(unsigned long)ldlmexp,
- ldlmexp->exp_cookie);
- GOTO(out_disc, rc = -ENOTCONN);
- }
+ argv[0] = obd_lustre_upcall;
+ argv[1] = "RECOVERY_OVER";
+ argv[2] = obd->obd_uuid.uuid;
+ argv[3] = NULL;
+
+ envp[0] = "HOME=/";
+ envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
+ envp[2] = NULL;
- old_hdl.addr = req->rq_repmsg->addr;
- old_hdl.cookie = req->rq_repmsg->cookie;
- if (memcmp(&imp->imp_handle, &old_hdl, sizeof(old_hdl))) {
- CERROR("%s@%s changed handle from "LPX64"/"LPX64
- " to "LPX64"/"LPX64"; "
- "copying, but this may foreshadow disaster\n",
- cli->cl_target_uuid.uuid,
- conn->c_remote_uuid.uuid,
- old_hdl.addr, old_hdl.cookie,
- imp->imp_handle.addr, imp->imp_handle.cookie);
- imp->imp_handle.addr = req->rq_repmsg->addr;
- imp->imp_handle.cookie = req->rq_repmsg->cookie;
- GOTO(out_disc, rc = 0);
- }
+ rc = USERMODEHELPER(argv[0], argv, envp);
+ if (rc < 0) {
+ CERROR("Error invoking recovery upcall %s %s %s: %d; check "
+ "/proc/sys/lustre/upcall\n",
+ argv[0], argv[1], argv[2], rc);
- CERROR("reconnected to %s@%s after partition\n",
- cli->cl_target_uuid.uuid, conn->c_remote_uuid.uuid);
- GOTO(out_disc, rc = 0);
+ } else {
+ CWARN("Invoked upcall %s %s %s\n",
+ argv[0], argv[1], argv[2]);
}
-
- old_hdl = imp->imp_handle;
- imp->imp_handle.addr = req->rq_repmsg->addr;
- imp->imp_handle.cookie = req->rq_repmsg->cookie;
- CERROR("reconnected to %s@%s ("LPX64"/"LPX64", was "LPX64"/"
- LPX64")!\n", cli->cl_target_uuid.uuid, conn->c_remote_uuid.uuid,
- imp->imp_handle.addr, imp->imp_handle.cookie,
- old_hdl.addr, old_hdl.cookie);
- GOTO(out_disc, rc = 0);
-
- out_disc:
- *reqptr = req;
- return rc;
}
-int ptlrpc_run_recovery_upcall(struct ptlrpc_connection *conn)
+void ptlrpc_run_failed_import_upcall(struct obd_import* imp)
{
- char *argv[3];
+#ifdef __KERNEL__
+ unsigned long flags;
+ char *argv[7];
char *envp[3];
int rc;
-
ENTRY;
- argv[0] = obd_recovery_upcall;
- argv[1] = conn->c_remote_uuid.uuid;
- argv[2] = NULL;
+
+ spin_lock_irqsave(&imp->imp_lock, flags);
+ if (imp->imp_state == LUSTRE_IMP_CLOSED) {
+ spin_unlock_irqrestore(&imp->imp_lock, flags);
+ EXIT;
+ return;
+ }
+ spin_unlock_irqrestore(&imp->imp_lock, flags);
+
+ argv[0] = obd_lustre_upcall;
+ argv[1] = "FAILED_IMPORT";
+ argv[2] = imp->imp_target_uuid.uuid;
+ argv[3] = imp->imp_obd->obd_name;
+ argv[4] = imp->imp_connection->c_remote_uuid.uuid;
+ argv[5] = imp->imp_obd->obd_uuid.uuid;
+ argv[6] = NULL;
envp[0] = "HOME=/";
envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
envp[2] = NULL;
- rc = call_usermodehelper(argv[0], argv, envp);
+ rc = USERMODEHELPER(argv[0], argv, envp);
if (rc < 0) {
- CERROR("Error invoking recovery upcall %s for %s: %d\n",
- argv[0], argv[1], rc);
- CERROR("Check /proc/sys/lustre/recovery_upcall?\n");
+ CERROR("Error invoking recovery upcall %s %s %s %s %s: %d; "
+ "check /proc/sys/lustre/upcall\n",
+ argv[0], argv[1], argv[2], argv[3], argv[4],rc);
+
} else {
- CERROR("Invoked upcall %s for connection %s\n",
- argv[0], argv[1]);
+ CWARN("Invoked upcall %s %s %s %s %s\n",
+ argv[0], argv[1], argv[2], argv[3], argv[4]);
+ }
+#else
+ if (imp->imp_state == LUSTRE_IMP_CLOSED) {
+ EXIT;
+ return;
}
+ ptlrpc_recover_import(imp, NULL);
+#endif
+}
- /*
- * We don't want to make this a "failed" recovery, because the system
- * administrator -- or, perhaps, tester -- may well be able to rescue
- * things by running the correct upcall.
- */
- RETURN(0);
+/* This might block waiting for the upcall to start, so it should
+ * not be called from a thread that shouldn't block. (Like ptlrpcd) */
+void ptlrpc_initiate_recovery(struct obd_import *imp)
+{
+ ENTRY;
+
+ LASSERT (obd_lustre_upcall != NULL);
+
+ if (strcmp(obd_lustre_upcall, "DEFAULT") == 0) {
+ CDEBUG(D_HA, "%s: starting recovery without upcall\n",
+ imp->imp_target_uuid.uuid);
+ ptlrpc_connect_import(imp, NULL);
+ }
+ else if (strcmp(obd_lustre_upcall, "NONE") == 0) {
+ CDEBUG(D_HA, "%s: recovery disabled\n",
+ imp->imp_target_uuid.uuid);
+ }
+ else {
+ CDEBUG(D_HA, "%s: calling upcall to start recovery\n",
+ imp->imp_target_uuid.uuid);
+ ptlrpc_run_failed_import_upcall(imp);
+ }
+
+ EXIT;
}
-int ptlrpc_replay(struct obd_import *imp)
+int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
{
int rc = 0;
struct list_head *tmp, *pos;
- struct ptlrpc_request *req;
+ struct ptlrpc_request *req = NULL;
unsigned long flags;
- __u64 committed = imp->imp_peer_committed_transno;
+ __u64 last_transno;
ENTRY;
+ *inflight = 0;
+
/* It might have committed some after we last spoke, so make sure we
* get rid of them now.
*/
spin_lock_irqsave(&imp->imp_lock, flags);
-
ptlrpc_free_committed(imp);
+ last_transno = imp->imp_last_replay_transno;
+ spin_unlock_irqrestore(&imp->imp_lock, flags);
- CDEBUG(D_HA, "import %p from %s has committed "LPD64"\n",
- imp, imp->imp_obd->u.cli.cl_target_uuid.uuid, committed);
+ CDEBUG(D_HA, "import %p from %s committed "LPU64" last "LPU64"\n",
+ imp, imp->imp_target_uuid.uuid, imp->imp_peer_committed_transno,
+ last_transno);
+
+ /* Do I need to hold a lock across this iteration? We shouldn't be
+ * racing with any additions to the list, because we're in recovery
+ * and are therefore not processing additional requests to add. Calls
+ * to ptlrpc_free_committed might commit requests, but nothing "newer"
+ * than the one we're replaying (it can't be committed until it's
+ * replayed, and we're doing that here). l_f_e_safe protects against
+ * problems with the current request being committed, in the unlikely
+ * event of that race. So, in conclusion, I think that it's safe to
+ * perform this list-walk without the imp_lock held.
+ *
+ * But, the {mdc,osc}_replay_open callbacks both iterate
+ * request lists, and have comments saying they assume the
+ * imp_lock is being held by ptlrpc_replay, but it's not. it's
+ * just a little race...
+ */
+ list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
+
+ /* If need to resend, stop on the matching one first. It's
+ possible though it's already been committed, so in that case
+ we'll just continue with replay */
+ if (imp->imp_resend_replay &&
+ req->rq_transno == last_transno) {
+ lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
+ break;
+ }
- list_for_each(tmp, &imp->imp_replay_list) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
- DEBUG_REQ(D_HA, req, "RETAINED: ");
- }
+ if (req->rq_transno > last_transno) {
+ imp->imp_last_replay_transno = req->rq_transno;
+ break;
+ }
- list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
+ req = NULL;
+ }
- DEBUG_REQ(D_HA, req, "REPLAY:");
+ imp->imp_resend_replay = 0;
- /* XXX locking WRT failure during replay? */
+ if (req != NULL) {
rc = ptlrpc_replay_req(req);
-
if (rc) {
- CERROR("recovery replay error %d for req "LPD64"\n",
- rc, req->rq_xid);
- GOTO(out, rc);
+ CERROR("recovery replay error %d for req "
+ LPD64"\n", rc, req->rq_xid);
+ RETURN(rc);
}
+ *inflight = 1;
}
+ RETURN(rc);
+}
- out:
+int ptlrpc_resend(struct obd_import *imp)
+{
+ struct list_head *tmp, *pos;
+ struct ptlrpc_request *req;
+ unsigned long flags;
+
+ ENTRY;
+
+ /* As long as we're in recovery, nothing should be added to the sending
+ * list, so we don't need to hold the lock during this iteration and
+ * resend process.
+ */
+ /* Well... what if lctl recover is called twice at the same time?
+ */
+ spin_lock_irqsave(&imp->imp_lock, flags);
+ if (imp->imp_state != LUSTRE_IMP_RECOVER) {
+ spin_unlock_irqrestore(&imp->imp_lock, flags);
+ RETURN(-1);
+ }
spin_unlock_irqrestore(&imp->imp_lock, flags);
- return rc;
+
+ list_for_each_safe(tmp, pos, &imp->imp_sending_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_list);
+ ptlrpc_resend_req(req);
+ }
+
+ RETURN(0);
}
-#define NO_RESEND 0 /* No action required. */
-#define RESEND 1 /* Resend required. */
-#define RESEND_IGNORE 2 /* Resend, ignore the reply (already saw it). */
-#define RESTART 3 /* Have to restart the call, sorry! */
+void ptlrpc_wake_delayed(struct obd_import *imp)
+{
+ unsigned long flags;
+ struct list_head *tmp, *pos;
+ struct ptlrpc_request *req;
+
+ spin_lock_irqsave(&imp->imp_lock, flags);
+ list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
+ req = list_entry(tmp, struct ptlrpc_request, rq_list);
+
+ DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
+ ptlrpc_wake_client_req(req);
+ }
+ spin_unlock_irqrestore(&imp->imp_lock, flags);
+}
-static int resend_type(struct ptlrpc_request *req, __u64 committed)
+void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
{
- if (req->rq_transno && req->rq_transno < committed) {
- if (req->rq_flags & PTL_RPC_FL_REPLIED) {
- /* Saw the reply and it was committed, no biggie. */
- DEBUG_REQ(D_HA, req, "NO_RESEND");
- return NO_RESEND;
+ int rc;
+ struct obd_import *imp= failed_req->rq_import;
+ unsigned long flags;
+ ENTRY;
+
+ CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
+ imp->imp_obd->obd_name,
+ imp->imp_target_uuid.uuid,
+ imp->imp_connection->c_remote_uuid.uuid);
+
+ if (ptlrpc_set_import_discon(imp)) {
+ if (!imp->imp_replayable) {
+ CDEBUG(D_HA, "import %s@%s for %s not replayable, "
+ "auto-deactivating\n",
+ imp->imp_target_uuid.uuid,
+ imp->imp_connection->c_remote_uuid.uuid,
+ imp->imp_obd->obd_name);
+ ptlrpc_deactivate_import(imp);
}
- /* Request committed, but no reply: have to restart. */
- return RESTART;
- }
- if (req->rq_flags & PTL_RPC_FL_REPLIED) {
- /* Saw reply, so resend and ignore new reply. */
- return RESEND_IGNORE;
+ rc = ptlrpc_connect_import(imp, NULL);
}
- /* Didn't see reply either, so resend. */
- return RESEND;
+ /* Wait for recovery to complete and resend. If evicted, then
+ this request will be errored out later.*/
+ spin_lock_irqsave(&failed_req->rq_lock, flags);
+ failed_req->rq_resend = 1;
+ spin_unlock_irqrestore(&failed_req->rq_lock, flags);
+ EXIT;
}
-int ptlrpc_resend(struct obd_import *imp)
+/*
+ * This should only be called by the ioctl interface, currently
+ * with the lctl deactivate and activate commands.
+ */
+int ptlrpc_set_import_active(struct obd_import *imp, int active)
{
+ struct obd_device *obd = imp->imp_obd;
int rc = 0;
- struct list_head *tmp, *pos;
- struct ptlrpc_request *req;
- unsigned long flags;
- __u64 committed = imp->imp_peer_committed_transno;
- ENTRY;
+ LASSERT(obd);
- spin_lock_irqsave(&imp->imp_lock, flags);
- list_for_each(tmp, &imp->imp_sending_list) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
- DEBUG_REQ(D_HA, req, "SENDING: ");
+ /* When deactivating, mark import invalid, and abort in-flight
+ * requests. */
+ if (!active) {
+ ptlrpc_invalidate_import(imp, 0);
+ imp->imp_deactive = 1;
}
- list_for_each_safe(tmp, pos, &imp->imp_sending_list) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
+ /* When activating, mark import valid, and attempt recovery */
+ if (active) {
+ imp->imp_deactive = 0;
+ CDEBUG(D_HA, "setting import %s VALID\n",
+ imp->imp_target_uuid.uuid);
+ rc = ptlrpc_recover_import(imp, NULL);
+ }
- switch(resend_type(req, committed)) {
- case NO_RESEND:
- break;
+ RETURN(rc);
+}
- case RESTART:
- DEBUG_REQ(D_HA, req, "RESTART:");
- ptlrpc_restart_req(req);
- break;
+int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid)
+{
+ int rc;
+ ENTRY;
- case RESEND_IGNORE:
- DEBUG_REQ(D_HA, req, "RESEND_IGNORE:");
- rc = ptlrpc_replay_req(req);
- if (rc) {
- DEBUG_REQ(D_ERROR, req, "error %d resending:",
- rc);
- ptlrpc_restart_req(req); /* might as well */
- }
- break;
+ /* force import to be disconnected. */
+ ptlrpc_set_import_discon(imp);
- case RESEND:
- DEBUG_REQ(D_HA, req, "RESEND:");
- ptlrpc_resend_req(req);
- break;
+ rc = ptlrpc_recover_import_no_retry(imp, new_uuid);
- default:
- LBUG();
- }
- }
+ RETURN(rc);
+}
+int ptlrpc_import_in_recovery(struct obd_import *imp)
+{
+ unsigned long flags;
+ int in_recovery = 1;
+ spin_lock_irqsave(&imp->imp_lock, flags);
+ if (imp->imp_state == LUSTRE_IMP_FULL ||
+ imp->imp_state == LUSTRE_IMP_CLOSED ||
+ imp->imp_state == LUSTRE_IMP_DISCON)
+ in_recovery = 0;
spin_unlock_irqrestore(&imp->imp_lock, flags);
- RETURN(rc);
+ return in_recovery;
}
-void ptlrpc_wake_delayed(struct obd_import *imp)
+static int ptlrpc_recover_import_no_retry(struct obd_import *imp,
+ char *new_uuid)
{
+ int rc;
unsigned long flags;
- struct list_head *tmp, *pos;
- struct ptlrpc_request *req;
+ int in_recovery = 0;
+ struct l_wait_info lwi;
+ ENTRY;
spin_lock_irqsave(&imp->imp_lock, flags);
- list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
- DEBUG_REQ(D_HA, req, "waking:");
- wake_up(&req->rq_wait_for_rep);
+ if (imp->imp_state != LUSTRE_IMP_DISCON) {
+ in_recovery = 1;
}
spin_unlock_irqrestore(&imp->imp_lock, flags);
+
+ if (in_recovery == 1)
+ RETURN(-EALREADY);
+
+ rc = ptlrpc_connect_import(imp, new_uuid);
+ if (rc)
+ RETURN(rc);
+
+ CDEBUG(D_HA, "%s: recovery started, waiting\n",
+ imp->imp_target_uuid.uuid);
+
+ lwi = LWI_TIMEOUT(MAX(obd_timeout * HZ, 1), NULL, NULL);
+ rc = l_wait_event(imp->imp_recovery_waitq,
+ !ptlrpc_import_in_recovery(imp), &lwi);
+ CDEBUG(D_HA, "%s: recovery finished\n",
+ imp->imp_target_uuid.uuid);
+
+ RETURN(rc);
+}
+
+void ptlrpc_fail_export(struct obd_export *exp)
+{
+ int rc, already_failed;
+ unsigned long flags;
+
+ spin_lock_irqsave(&exp->exp_lock, flags);
+ already_failed = exp->exp_failed;
+ exp->exp_failed = 1;
+ spin_unlock_irqrestore(&exp->exp_lock, flags);
+
+ if (already_failed) {
+ CDEBUG(D_HA, "disconnecting dead export %p/%s; skipping\n",
+ exp, exp->exp_client_uuid.uuid);
+ return;
+ }
+
+ CDEBUG(D_HA, "disconnecting export %p/%s\n",
+ exp, exp->exp_client_uuid.uuid);
+
+ if (obd_dump_on_timeout)
+ portals_debug_dumplog();
+
+ /* Most callers into obd_disconnect are removing their own reference
+ * (request, for example) in addition to the one from the hash table.
+ * We don't have such a reference here, so make one. */
+ class_export_get(exp);
+ rc = obd_disconnect(exp, 0);
+ if (rc)
+ CERROR("disconnecting export %p failed: %d\n", exp, rc);
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff