/*! \brief
* Called at initialization to figure out version of the dbase we really have.
*
* This routine is called after replaying the log; it reads the restored labels.
*/
static int
InitializeDB(struct ubik_dbase *adbase)
{
afs_int32 code;
code = (*adbase->getlabel) (adbase, 0, &adbase->version);
if (code) {
/* try setting the label to a new value */
UBIK_VERSION_LOCK;
adbase->version.epoch = 1; /* value for newly-initialized db */
adbase->version.counter = 1;
code = (*adbase->setlabel) (adbase, 0, &adbase->version);
if (code) {
/* failed, try to set it back */
adbase->version.epoch = 0;
adbase->version.counter = 0;
(*adbase->setlabel) (adbase, 0, &adbase->version);
}
#ifdef AFS_PTHREAD_ENV
CV_BROADCAST(&adbase->version_cond);
#else
LWP_NoYieldSignal(&adbase->version);
#endif
UBIK_VERSION_UNLOCK;
}
return 0;
}
/**
* low-level thread entry point.
*
* @param[in] rock opaque pointer to thread worker object
*
* @return opaque return pointer from pool entry function
*
* @internal
*/
static void *
_afs_tp_worker_run(void * rock)
{
struct afs_thread_pool_worker * worker = rock;
struct afs_thread_pool * pool = worker->pool;
/* register worker with pool */
MUTEX_ENTER(&pool->lock);
queue_Append(&pool->thread_list, worker);
pool->nthreads++;
MUTEX_EXIT(&pool->lock);
/* call high-level entry point */
worker->ret = (*pool->entry)(pool, worker, pool->work_queue, pool->rock);
/* adjust pool live thread count */
MUTEX_ENTER(&pool->lock);
osi_Assert(pool->nthreads);
queue_Remove(worker);
pool->nthreads--;
if (!pool->nthreads) {
CV_BROADCAST(&pool->shutdown_cv);
pool->state = AFS_TP_STATE_STOPPED;
}
MUTEX_EXIT(&pool->lock);
_afs_tp_worker_free(worker);
return NULL;
}
afs_int32
canWrite(int fid)
{
#ifndef AFS_PTHREAD_ENV
afs_int32 code = 0;
#endif
extern dumpSyncP dumpSyncPtr;
ObtainWriteLock(&dumpSyncPtr->ds_lock);
/* let the pipe drain */
while (dumpSyncPtr->ds_bytes > 0) {
if (dumpSyncPtr->ds_readerStatus == DS_WAITING) {
dumpSyncPtr->ds_readerStatus = 0;
#ifdef AFS_PTHREAD_ENV
CV_BROADCAST(&dumpSyncPtr->ds_readerStatus_cond);
#else
code = LWP_SignalProcess(&dumpSyncPtr->ds_readerStatus);
if (code)
LogError(code, "canWrite: Signal delivery failed\n");
#endif
}
dumpSyncPtr->ds_writerStatus = DS_WAITING;
ReleaseWriteLock(&dumpSyncPtr->ds_lock);
#ifdef AFS_PTHREAD_ENV
MUTEX_ENTER(&dumpSyncPtr->ds_writerStatus_mutex);
CV_WAIT(&dumpSyncPtr->ds_writerStatus_cond, &dumpSyncPtr->ds_writerStatus_mutex);
MUTEX_EXIT(&dumpSyncPtr->ds_writerStatus_mutex);
#else
LWP_WaitProcess(&dumpSyncPtr->ds_writerStatus);
#endif
ObtainWriteLock(&dumpSyncPtr->ds_lock);
}
return (1);
}
/**
* wakeup all threads waiting in dequeue.
*
* @param[in] list list object
*
* @return operation status
* @retval 0 success
*
* @internal
*/
static int
_afs_wq_node_list_shutdown(struct afs_work_queue_node_list * list)
{
int ret = 0;
struct afs_work_queue_node *node, *nnode;
MUTEX_ENTER(&list->lock);
list->shutdown = 1;
for (queue_Scan(&list->list, node, nnode, afs_work_queue_node)) {
_afs_wq_node_state_change(node, AFS_WQ_NODE_STATE_ERROR);
queue_Remove(node);
node->qidx = AFS_WQ_NODE_LIST_NONE;
node->queue = NULL;
if (node->detached) {
/* if we are detached, we hold the reference on the node;
* otherwise, it is some other caller that holds the reference.
* So don't put the node if we are not detached; the node will
* get freed when someone else calls afs_wq_node_put */
afs_wq_node_put(node);
}
}
CV_BROADCAST(&list->cv);
MUTEX_EXIT(&list->lock);
return ret;
}
/**
* decrements queue->running_count, and signals waiters if appropriate.
*
* @param[in] queue queue to dec the running count of
*/
static void
_afs_wq_dec_running_count(struct afs_work_queue *queue)
{
MUTEX_ENTER(&queue->lock);
queue->running_count--;
if (queue->shutdown && queue->running_count == 0) {
/* if we've shut down, someone may be waiting for the running count
* to drop to 0 */
CV_BROADCAST(&queue->running_cv);
}
MUTEX_EXIT(&queue->lock);
}
/**
* change a node's state.
*
* @param[in] node node object
* @param[in] new_state new object state
*
* @return old state
*
* @pre node->lock held
*
* @internal
*/
static afs_wq_work_state_t
_afs_wq_node_state_change(struct afs_work_queue_node * node,
afs_wq_work_state_t new_state)
{
afs_wq_work_state_t old_state;
old_state = node->state;
node->state = new_state;
CV_BROADCAST(&node->state_cv);
return old_state;
}
static void *
SalvageChildReaperThread(void * args)
{
int slot, pid, status;
struct log_cleanup_node * cleanup;
MUTEX_ENTER(&worker_lock);
/* loop reaping our children */
while (1) {
/* wait() won't block unless we have children, so
* block on the cond var if we're childless */
while (current_workers == 0) {
CV_WAIT(&worker_cv, &worker_lock);
}
MUTEX_EXIT(&worker_lock);
cleanup = (struct log_cleanup_node *) malloc(sizeof(struct log_cleanup_node));
while (Reap_Child("salvageserver", &pid, &status) < 0) {
/* try to prevent livelock if something goes wrong */
sleep(1);
}
VOL_LOCK;
for (slot = 0; slot < Parallel; slot++) {
if (child_slot[slot] == pid)
break;
}
osi_Assert(slot < Parallel);
child_slot[slot] = 0;
VOL_UNLOCK;
SALVSYNC_doneWorkByPid(pid, status);
MUTEX_ENTER(&worker_lock);
if (cleanup) {
cleanup->pid = pid;
queue_Append(&log_cleanup_queue, cleanup);
CV_SIGNAL(&log_cleanup_queue.queue_change_cv);
}
/* ok, we've reaped a child */
current_workers--;
CV_BROADCAST(&worker_cv);
}
return NULL;
}
/**
* shutdown a work queue.
*
* @param[in] queue work queue object pointer
*
* @return operation status
* @retval 0 success
*/
int
afs_wq_shutdown(struct afs_work_queue * queue)
{
int ret = 0;
MUTEX_ENTER(&queue->lock);
if (queue->shutdown) {
/* already shutdown, do nothing */
MUTEX_EXIT(&queue->lock);
goto error;
}
queue->shutdown = 1;
ret = _afs_wq_node_list_shutdown(&queue->ready_list);
if (ret) {
goto error;
}
ret = _afs_wq_node_list_shutdown(&queue->blocked_list);
if (ret) {
goto error;
}
ret = _afs_wq_node_list_shutdown(&queue->done_list);
if (ret) {
goto error;
}
/* signal everyone that could be waiting, since these conditions will
* generally fail to signal on their own if we're shutdown, since no
* progress is being made */
CV_BROADCAST(&queue->pend_cv);
CV_BROADCAST(&queue->empty_cv);
MUTEX_EXIT(&queue->lock);
error:
return ret;
}
/**
* change VVGC partition state.
*
* @param[in] part disk partition object
* @param[in] state new state
*
* @pre VOL_LOCK is held
*
* @return old state
*
* @internal
*/
int
_VVGC_state_change(struct DiskPartition64 * part,
VVGCache_part_state_t state)
{
VVGCache_part_state_t old_state;
old_state = VVGCache.part[part->index].state;
VVGCache.part[part->index].state = state;
if (old_state != state) {
CV_BROADCAST(&VVGCache.part[part->index].cv);
}
return old_state;
}
void
haveWritten(afs_int32 nbytes)
{
#ifndef AFS_PTHREAD_ENV
afs_int32 code = 0;
#endif
extern dumpSyncP dumpSyncPtr;
dumpSyncPtr->ds_bytes += nbytes;
if (dumpSyncPtr->ds_readerStatus == DS_WAITING) {
dumpSyncPtr->ds_readerStatus = 0;
#ifdef AFS_PTHREAD_ENV
CV_BROADCAST(&dumpSyncPtr->ds_readerStatus_cond);
#else
code = LWP_SignalProcess(&dumpSyncPtr->ds_readerStatus);
if (code)
LogError(code, "haveWritten: Signal delivery failed\n");
#endif
}
ReleaseWriteLock(&dumpSyncPtr->ds_lock);
}
void
doneWriting(afs_int32 error)
{
#ifndef AFS_PTHREAD_ENV
afs_int32 code = 0;
#endif
/* wait for the reader */
ObtainWriteLock(&dumpSyncPtr->ds_lock);
while (dumpSyncPtr->ds_readerStatus != DS_WAITING) {
LogDebug(4, "doneWriting: waiting for Reader\n");
dumpSyncPtr->ds_writerStatus = DS_WAITING;
ReleaseWriteLock(&dumpSyncPtr->ds_lock);
#ifdef AFS_PTHREAD_ENV
MUTEX_ENTER(&dumpSyncPtr->ds_writerStatus_mutex);
CV_WAIT(&dumpSyncPtr->ds_writerStatus_cond, &dumpSyncPtr->ds_writerStatus_mutex);
MUTEX_EXIT(&dumpSyncPtr->ds_writerStatus_mutex);
#else
LWP_WaitProcess(&dumpSyncPtr->ds_writerStatus);
#endif
ObtainWriteLock(&dumpSyncPtr->ds_lock);
}
LogDebug(4, "doneWriting: setting done\n");
/* signal that we are done */
if (error)
dumpSyncPtr->ds_writerStatus = DS_DONE_ERROR;
else
dumpSyncPtr->ds_writerStatus = DS_DONE;
dumpSyncPtr->ds_readerStatus = 0;
#ifdef AFS_PTHREAD_ENV
CV_BROADCAST(&dumpSyncPtr->ds_readerStatus_cond);
#else
code = LWP_NoYieldSignal(&dumpSyncPtr->ds_readerStatus);
if (code)
LogError(code, "doneWriting: Signal delivery failed\n");
#endif
ReleaseWriteLock(&dumpSyncPtr->ds_lock);
}
/*!
* \brief Main interaction loop for the recovery manager
*
* The recovery light-weight process only runs when you're the
* synchronization site. It performs the following tasks, if and only
* if the prerequisite tasks have been performed successfully (it
* keeps track of which ones have been performed in its bit map,
* \p urecovery_state).
*
* First, it is responsible for probing that all servers are up. This
* is the only operation that must be performed even if this is not
* yet the sync site, since otherwise this site may not notice that
* enough other machines are running to even elect this guy to be the
* sync site.
*
* After that, the recovery process does nothing until the beacon and
* voting modules manage to get this site elected sync site.
*
* After becoming sync site, recovery first attempts to find the best
* database available in the network (it must do this in order to
* ensure finding the latest committed data). After finding the right
* database, it must fetch this dbase to the sync site.
*
* After fetching the dbase, it relabels it with a new version number,
* to ensure that everyone recognizes this dbase as the most recent
* dbase.
*
* One the dbase has been relabelled, this machine can start handling
* requests. However, the recovery module still has one more task:
* propagating the dbase out to everyone who is up in the network.
*/
void *
urecovery_Interact(void *dummy)
{
afs_int32 code, tcode;
struct ubik_server *bestServer = NULL;
struct ubik_server *ts;
int dbok, doingRPC, now;
afs_int32 lastProbeTime;
/* if we're the sync site, the best db version we've found yet */
static struct ubik_version bestDBVersion;
struct ubik_version tversion;
struct timeval tv;
int length, tlen, offset, file, nbytes;
struct rx_call *rxcall;
char tbuffer[1024];
struct ubik_stat ubikstat;
struct in_addr inAddr;
char hoststr[16];
char pbuffer[1028];
int fd = -1;
afs_int32 pass;
afs_pthread_setname_self("recovery");
/* otherwise, begin interaction */
urecovery_state = 0;
lastProbeTime = 0;
while (1) {
/* Run through this loop every 4 seconds */
tv.tv_sec = 4;
tv.tv_usec = 0;
#ifdef AFS_PTHREAD_ENV
select(0, 0, 0, 0, &tv);
#else
IOMGR_Select(0, 0, 0, 0, &tv);
#endif
ubik_dprint("recovery running in state %x\n", urecovery_state);
/* Every 30 seconds, check all the down servers and mark them
* as up if they respond. When a server comes up or found to
* not be current, then re-find the the best database and
* propogate it.
*/
if ((now = FT_ApproxTime()) > 30 + lastProbeTime) {
for (ts = ubik_servers, doingRPC = 0; ts; ts = ts->next) {
UBIK_BEACON_LOCK;
if (!ts->up) {
UBIK_BEACON_UNLOCK;
doingRPC = 1;
code = DoProbe(ts);
if (code == 0) {
UBIK_BEACON_LOCK;
ts->up = 1;
UBIK_BEACON_UNLOCK;
DBHOLD(ubik_dbase);
urecovery_state &= ~UBIK_RECFOUNDDB;
DBRELE(ubik_dbase);
}
} else {
UBIK_BEACON_UNLOCK;
DBHOLD(ubik_dbase);
if (!ts->currentDB)
urecovery_state &= ~UBIK_RECFOUNDDB;
DBRELE(ubik_dbase);
}
}
if (doingRPC)
now = FT_ApproxTime();
lastProbeTime = now;
}
/* Mark whether we are the sync site */
DBHOLD(ubik_dbase);
if (!ubeacon_AmSyncSite()) {
urecovery_state &= ~UBIK_RECSYNCSITE;
DBRELE(ubik_dbase);
continue; /* nothing to do */
}
urecovery_state |= UBIK_RECSYNCSITE;
/* If a server has just come up or if we have not found the
* most current database, then go find the most current db.
*/
if (!(urecovery_state & UBIK_RECFOUNDDB)) {
DBRELE(ubik_dbase);
bestServer = (struct ubik_server *)0;
bestDBVersion.epoch = 0;
bestDBVersion.counter = 0;
for (ts = ubik_servers; ts; ts = ts->next) {
UBIK_BEACON_LOCK;
if (!ts->up) {
UBIK_BEACON_UNLOCK;
continue; /* don't bother with these guys */
}
UBIK_BEACON_UNLOCK;
if (ts->isClone)
continue;
UBIK_ADDR_LOCK;
code = DISK_GetVersion(ts->disk_rxcid, &ts->version);
UBIK_ADDR_UNLOCK;
if (code == 0) {
/* perhaps this is the best version */
if (vcmp(ts->version, bestDBVersion) > 0) {
/* new best version */
bestDBVersion = ts->version;
bestServer = ts;
}
}
}
/* take into consideration our version. Remember if we,
* the sync site, have the best version. Also note that
* we may need to send the best version out.
*/
DBHOLD(ubik_dbase);
if (vcmp(ubik_dbase->version, bestDBVersion) >= 0) {
bestDBVersion = ubik_dbase->version;
bestServer = (struct ubik_server *)0;
urecovery_state |= UBIK_RECHAVEDB;
} else {
/* Clear the flag only when we know we have to retrieve
* the db. Because urecovery_AllBetter() looks at it.
*/
urecovery_state &= ~UBIK_RECHAVEDB;
}
urecovery_state |= UBIK_RECFOUNDDB;
urecovery_state &= ~UBIK_RECSENTDB;
}
if (!(urecovery_state & UBIK_RECFOUNDDB)) {
DBRELE(ubik_dbase);
continue; /* not ready */
}
/* If we, the sync site, do not have the best db version, then
* go and get it from the server that does.
*/
if ((urecovery_state & UBIK_RECHAVEDB) || !bestServer) {
urecovery_state |= UBIK_RECHAVEDB;
} else {
/* we don't have the best version; we should fetch it. */
urecovery_AbortAll(ubik_dbase);
/* Rx code to do the Bulk fetch */
file = 0;
offset = 0;
UBIK_ADDR_LOCK;
rxcall = rx_NewCall(bestServer->disk_rxcid);
ubik_print("Ubik: Synchronize database with server %s\n",
afs_inet_ntoa_r(bestServer->addr[0], hoststr));
UBIK_ADDR_UNLOCK;
code = StartDISK_GetFile(rxcall, file);
if (code) {
ubik_dprint("StartDiskGetFile failed=%d\n", code);
goto FetchEndCall;
}
nbytes = rx_Read(rxcall, (char *)&length, sizeof(afs_int32));
length = ntohl(length);
if (nbytes != sizeof(afs_int32)) {
ubik_dprint("Rx-read length error=%d\n", code = BULK_ERROR);
code = EIO;
goto FetchEndCall;
}
/* give invalid label during file transit */
UBIK_VERSION_LOCK;
tversion.epoch = 0;
code = (*ubik_dbase->setlabel) (ubik_dbase, file, &tversion);
UBIK_VERSION_UNLOCK;
if (code) {
ubik_dprint("setlabel io error=%d\n", code);
goto FetchEndCall;
}
snprintf(pbuffer, sizeof(pbuffer), "%s.DB%s%d.TMP",
ubik_dbase->pathName, (file<0)?"SYS":"",
(file<0)?-file:file);
fd = open(pbuffer, O_CREAT | O_RDWR | O_TRUNC, 0600);
if (fd < 0) {
code = errno;
goto FetchEndCall;
}
code = lseek(fd, HDRSIZE, 0);
if (code != HDRSIZE) {
close(fd);
goto FetchEndCall;
}
pass = 0;
while (length > 0) {
tlen = (length > sizeof(tbuffer) ? sizeof(tbuffer) : length);
#ifndef AFS_PTHREAD_ENV
if (pass % 4 == 0)
IOMGR_Poll();
#endif
nbytes = rx_Read(rxcall, tbuffer, tlen);
if (nbytes != tlen) {
ubik_dprint("Rx-read bulk error=%d\n", code = BULK_ERROR);
code = EIO;
close(fd);
goto FetchEndCall;
}
nbytes = write(fd, tbuffer, tlen);
pass++;
if (nbytes != tlen) {
code = UIOERROR;
close(fd);
goto FetchEndCall;
}
offset += tlen;
length -= tlen;
}
code = close(fd);
if (code)
goto FetchEndCall;
code = EndDISK_GetFile(rxcall, &tversion);
FetchEndCall:
tcode = rx_EndCall(rxcall, code);
if (!code)
code = tcode;
if (!code) {
/* we got a new file, set up its header */
urecovery_state |= UBIK_RECHAVEDB;
UBIK_VERSION_LOCK;
memcpy(&ubik_dbase->version, &tversion,
sizeof(struct ubik_version));
snprintf(tbuffer, sizeof(tbuffer), "%s.DB%s%d",
ubik_dbase->pathName, (file<0)?"SYS":"",
(file<0)?-file:file);
#ifdef AFS_NT40_ENV
snprintf(pbuffer, sizeof(pbuffer), "%s.DB%s%d.OLD",
ubik_dbase->pathName, (file<0)?"SYS":"",
(file<0)?-file:file);
code = unlink(pbuffer);
if (!code)
code = rename(tbuffer, pbuffer);
snprintf(pbuffer, sizeof(pbuffer), "%s.DB%s%d.TMP",
ubik_dbase->pathName, (file<0)?"SYS":"",
(file<0)?-file:file);
#endif
if (!code)
code = rename(pbuffer, tbuffer);
if (!code) {
(*ubik_dbase->open) (ubik_dbase, file);
/* after data is good, sync disk with correct label */
code =
(*ubik_dbase->setlabel) (ubik_dbase, 0,
&ubik_dbase->version);
}
UBIK_VERSION_UNLOCK;
#ifdef AFS_NT40_ENV
snprintf(pbuffer, sizeof(pbuffer), "%s.DB%s%d.OLD",
ubik_dbase->pathName, (file<0)?"SYS":"",
(file<0)?-file:file);
unlink(pbuffer);
#endif
}
if (code) {
unlink(pbuffer);
/*
* We will effectively invalidate the old data forever now.
* Unclear if we *should* but we do.
*/
UBIK_VERSION_LOCK;
ubik_dbase->version.epoch = 0;
ubik_dbase->version.counter = 0;
UBIK_VERSION_UNLOCK;
ubik_print("Ubik: Synchronize database failed (error = %d)\n",
code);
} else {
ubik_print("Ubik: Synchronize database completed\n");
urecovery_state |= UBIK_RECHAVEDB;
}
udisk_Invalidate(ubik_dbase, 0); /* data has changed */
#ifdef AFS_PTHREAD_ENV
CV_BROADCAST(&ubik_dbase->version_cond);
#else
LWP_NoYieldSignal(&ubik_dbase->version);
#endif
}
if (!(urecovery_state & UBIK_RECHAVEDB)) {
DBRELE(ubik_dbase);
continue; /* not ready */
}
/* If the database was newly initialized, then when we establish quorum, write
* a new label. This allows urecovery_AllBetter() to allow access for reads.
* Setting it to 2 also allows another site to come along with a newer
* database and overwrite this one.
*/
if (ubik_dbase->version.epoch == 1) {
urecovery_AbortAll(ubik_dbase);
UBIK_VERSION_LOCK;
version_globals.ubik_epochTime = 2;
ubik_dbase->version.epoch = version_globals.ubik_epochTime;
ubik_dbase->version.counter = 1;
code =
(*ubik_dbase->setlabel) (ubik_dbase, 0, &ubik_dbase->version);
UBIK_VERSION_UNLOCK;
udisk_Invalidate(ubik_dbase, 0); /* data may have changed */
#ifdef AFS_PTHREAD_ENV
CV_BROADCAST(&ubik_dbase->version_cond);
#else
LWP_NoYieldSignal(&ubik_dbase->version);
#endif
}
/* Check the other sites and send the database to them if they
* do not have the current db.
*/
if (!(urecovery_state & UBIK_RECSENTDB)) {
/* now propagate out new version to everyone else */
dbok = 1; /* start off assuming they all worked */
/*
* Check if a write transaction is in progress. We can't send the
* db when a write is in progress here because the db would be
* obsolete as soon as it goes there. Also, ops after the begin
* trans would reach the recepient and wouldn't find a transaction
* pending there. Frankly, I don't think it's possible to get past
* the write-lock above if there is a write transaction in progress,
* but then, it won't hurt to check, will it?
*/
if (ubik_dbase->flags & DBWRITING) {
struct timeval tv;
int safety = 0;
long cur_usec = 50000;
while ((ubik_dbase->flags & DBWRITING) && (safety < 500)) {
DBRELE(ubik_dbase);
/* sleep for a little while */
tv.tv_sec = 0;
tv.tv_usec = cur_usec;
#ifdef AFS_PTHREAD_ENV
select(0, 0, 0, 0, &tv);
#else
IOMGR_Select(0, 0, 0, 0, &tv);
#endif
cur_usec += 10000;
safety++;
DBHOLD(ubik_dbase);
}
}
for (ts = ubik_servers; ts; ts = ts->next) {
UBIK_ADDR_LOCK;
inAddr.s_addr = ts->addr[0];
UBIK_ADDR_UNLOCK;
UBIK_BEACON_LOCK;
if (!ts->up) {
UBIK_BEACON_UNLOCK;
ubik_dprint("recovery cannot send version to %s\n",
afs_inet_ntoa_r(inAddr.s_addr, hoststr));
dbok = 0;
continue;
}
UBIK_BEACON_UNLOCK;
ubik_dprint("recovery sending version to %s\n",
afs_inet_ntoa_r(inAddr.s_addr, hoststr));
if (vcmp(ts->version, ubik_dbase->version) != 0) {
ubik_dprint("recovery stating local database\n");
/* Rx code to do the Bulk Store */
code = (*ubik_dbase->stat) (ubik_dbase, 0, &ubikstat);
if (!code) {
length = ubikstat.size;
file = offset = 0;
UBIK_ADDR_LOCK;
rxcall = rx_NewCall(ts->disk_rxcid);
UBIK_ADDR_UNLOCK;
code =
StartDISK_SendFile(rxcall, file, length,
&ubik_dbase->version);
if (code) {
ubik_dprint("StartDiskSendFile failed=%d\n",
code);
goto StoreEndCall;
}
while (length > 0) {
tlen =
(length >
sizeof(tbuffer) ? sizeof(tbuffer) : length);
nbytes =
(*ubik_dbase->read) (ubik_dbase, file,
tbuffer, offset, tlen);
if (nbytes != tlen) {
ubik_dprint("Local disk read error=%d\n",
code = UIOERROR);
goto StoreEndCall;
}
nbytes = rx_Write(rxcall, tbuffer, tlen);
if (nbytes != tlen) {
ubik_dprint("Rx-write bulk error=%d\n", code =
BULK_ERROR);
goto StoreEndCall;
}
offset += tlen;
length -= tlen;
}
code = EndDISK_SendFile(rxcall);
StoreEndCall:
code = rx_EndCall(rxcall, code);
}
if (code == 0) {
/* we set a new file, process its header */
ts->version = ubik_dbase->version;
ts->currentDB = 1;
} else
dbok = 0;
} else {
/* mark file up to date */
ts->currentDB = 1;
}
}
if (dbok)
urecovery_state |= UBIK_RECSENTDB;
}
DBRELE(ubik_dbase);
}
return NULL;
}
/**
* execute a node on the queue.
*
* @param[in] queue work queue
* @param[in] rock opaque pointer (passed as third arg to callback func)
* @param[in] block allow blocking in dequeue
*
* @return operation status
* @retval 0 completed a work unit
*
* @internal
*/
static int
_afs_wq_do(struct afs_work_queue * queue,
void * rock,
int block)
{
int code, ret = 0;
struct afs_work_queue_node * node;
afs_wq_callback_func_t * cbf;
afs_wq_work_state_t next_state;
struct afs_work_queue_node_list * ql;
void * node_rock;
int detached = 0;
/* We can inc queue->running_count before actually pulling the node off
* of the ready_list, since running_count only really matters when we are
* shut down. If we get shut down before we pull the node off of
* ready_list, but after we inc'd running_count,
* _afs_wq_node_list_dequeue should return immediately with EINTR,
* in which case we'll dec running_count, so it's as if we never inc'd it
* in the first place. */
MUTEX_ENTER(&queue->lock);
if (queue->shutdown) {
MUTEX_EXIT(&queue->lock);
return EINTR;
}
queue->running_count++;
MUTEX_EXIT(&queue->lock);
ret = _afs_wq_node_list_dequeue(&queue->ready_list,
&node,
AFS_WQ_NODE_STATE_RUNNING,
block);
if (ret) {
_afs_wq_dec_running_count(queue);
goto error;
}
cbf = node->cbf;
node_rock = node->rock;
detached = node->detached;
if (cbf != NULL) {
MUTEX_EXIT(&node->lock);
code = (*cbf)(queue, node, queue->rock, node_rock, rock);
MUTEX_ENTER(&node->lock);
if (code == 0) {
next_state = AFS_WQ_NODE_STATE_DONE;
ql = &queue->done_list;
} else if (code == AFS_WQ_ERROR_RESCHEDULE) {
if (node->error_count) {
next_state = AFS_WQ_NODE_STATE_ERROR;
ql = &queue->done_list;
} else if (node->block_count) {
next_state = AFS_WQ_NODE_STATE_BLOCKED;
ql = &queue->blocked_list;
} else {
next_state = AFS_WQ_NODE_STATE_SCHEDULED;
ql = &queue->ready_list;
}
} else {
next_state = AFS_WQ_NODE_STATE_ERROR;
ql = &queue->done_list;
}
} else {
next_state = AFS_WQ_NODE_STATE_DONE;
code = 0;
ql = &queue->done_list;
}
_afs_wq_dec_running_count(queue);
node->retcode = code;
if ((next_state == AFS_WQ_NODE_STATE_DONE) ||
(next_state == AFS_WQ_NODE_STATE_ERROR)) {
MUTEX_ENTER(&queue->lock);
if (queue->drain && queue->pend_count == queue->opts.pend_lothresh) {
/* signal other threads if we're about to below the low
* pending-tasks threshold */
queue->drain = 0;
CV_SIGNAL(&queue->pend_cv);
}
if (queue->pend_count == 1) {
/* signal other threads if we're about to become 'empty' */
CV_BROADCAST(&queue->empty_cv);
}
queue->pend_count--;
MUTEX_EXIT(&queue->lock);
}
ret = _afs_wq_node_state_wait_busy(node);
if (ret) {
goto error;
}
/* propagate scheduling changes down through dependencies */
ret = _afs_wq_dep_propagate(node, next_state);
if (ret) {
goto error;
}
ret = _afs_wq_node_state_wait_busy(node);
if (ret) {
goto error;
}
if (detached &&
((next_state == AFS_WQ_NODE_STATE_DONE) ||
(next_state == AFS_WQ_NODE_STATE_ERROR))) {
_afs_wq_node_state_change(node, next_state);
_afs_wq_node_put_r(node, 1);
} else {
ret = _afs_wq_node_list_enqueue(ql,
node,
next_state);
}
error:
return ret;
}
static void
SalvageServer(int argc, char **argv)
{
int pid, ret;
struct SalvageQueueNode * node;
pthread_t tid;
pthread_attr_t attrs;
int slot;
VolumePackageOptions opts;
/* All entries to the log will be appended. Useful if there are
* multiple salvagers appending to the log.
*/
CheckLogFile((char *)AFSDIR_SERVER_SALSRVLOG_FILEPATH);
#ifndef AFS_NT40_ENV
#ifdef AFS_LINUX20_ENV
fcntl(fileno(logFile), F_SETFL, O_APPEND); /* Isn't this redundant? */
#else
fcntl(fileno(logFile), F_SETFL, FAPPEND); /* Isn't this redundant? */
#endif
#endif
setlinebuf(logFile);
fprintf(logFile, "%s\n", cml_version_number);
LogCommandLine(argc, argv, "Online Salvage Server",
SalvageVersion, "Starting OpenAFS", Log);
/* Get and hold a lock for the duration of the salvage to make sure
* that no other salvage runs at the same time. The routine
* VInitVolumePackage2 (called below) makes sure that a file server or
* other volume utilities don't interfere with the salvage.
*/
/* even demand attach online salvager
* still needs this because we don't want
* a stand-alone salvager to conflict with
* the salvager daemon */
ObtainSharedSalvageLock();
child_slot = (int *) malloc(Parallel * sizeof(int));
osi_Assert(child_slot != NULL);
memset(child_slot, 0, Parallel * sizeof(int));
/* initialize things */
VOptDefaults(salvageServer, &opts);
if (VInitVolumePackage2(salvageServer, &opts)) {
Log("Shutting down: errors encountered initializing volume package\n");
Exit(1);
}
DInit(10);
queue_Init(&pending_q);
queue_Init(&log_cleanup_queue);
MUTEX_INIT(&worker_lock, "worker", MUTEX_DEFAULT, 0);
CV_INIT(&worker_cv, "worker", CV_DEFAULT, 0);
CV_INIT(&log_cleanup_queue.queue_change_cv, "queuechange", CV_DEFAULT, 0);
osi_Assert(pthread_attr_init(&attrs) == 0);
/* start up the reaper and log cleaner threads */
osi_Assert(pthread_attr_setdetachstate(&attrs, PTHREAD_CREATE_DETACHED) == 0);
osi_Assert(pthread_create(&tid,
&attrs,
&SalvageChildReaperThread,
NULL) == 0);
osi_Assert(pthread_create(&tid,
&attrs,
&SalvageLogCleanupThread,
NULL) == 0);
osi_Assert(pthread_create(&tid,
&attrs,
&SalvageLogScanningThread,
NULL) == 0);
/* loop forever serving requests */
while (1) {
node = SALVSYNC_getWork();
osi_Assert(node != NULL);
Log("dispatching child to salvage volume %u...\n",
node->command.sop.parent);
VOL_LOCK;
/* find a slot */
for (slot = 0; slot < Parallel; slot++) {
if (!child_slot[slot])
break;
}
osi_Assert (slot < Parallel);
do_fork:
pid = Fork();
if (pid == 0) {
VOL_UNLOCK;
ret = DoSalvageVolume(node, slot);
Exit(ret);
} else if (pid < 0) {
Log("failed to fork child worker process\n");
sleep(1);
goto do_fork;
} else {
child_slot[slot] = pid;
node->pid = pid;
VOL_UNLOCK;
MUTEX_ENTER(&worker_lock);
current_workers++;
/* let the reaper thread know another worker was spawned */
CV_BROADCAST(&worker_cv);
/* if we're overquota, wait for the reaper */
while (current_workers >= Parallel) {
CV_WAIT(&worker_cv, &worker_lock);
}
MUTEX_EXIT(&worker_lock);
}
}
}
/**
* acquire a lock on a file on local disk.
*
* @param[in] dl the VDiskLock structure corresponding to the file on disk
* @param[in] locktype READ_LOCK if you want a read lock, or WRITE_LOCK if
* you want a write lock
* @param[in] nonblock 0 to wait for conflicting locks to clear before
* obtaining the lock; 1 to fail immediately if a
* conflicting lock is held by someone else
*
* @return operation status
* @retval 0 success
* @retval EBUSY someone else is holding a conflicting lock and nonblock=1 was
* specified
* @retval EIO error acquiring file lock
*
* @note DAFS only
*
* @note while normal fcntl-y locks on Unix systems generally only work per-
* process, this interface also deals with locks between threads in the
* process in addition to different processes acquiring the lock
*/
int
VGetDiskLock(struct VDiskLock *dl, int locktype, int nonblock)
{
int code = 0;
osi_Assert(locktype == READ_LOCK || locktype == WRITE_LOCK);
if (nonblock) {
if (locktype == READ_LOCK) {
ObtainReadLockNoBlock(&dl->rwlock, code);
} else {
ObtainWriteLockNoBlock(&dl->rwlock, code);
}
if (code) {
return EBUSY;
}
} else if (locktype == READ_LOCK) {
ObtainReadLock(&dl->rwlock);
} else {
ObtainWriteLock(&dl->rwlock);
}
MUTEX_ENTER(&dl->mutex);
if ((dl->flags & VDISKLOCK_ACQUIRING)) {
/* Some other thread is waiting to acquire an fs lock. If nonblock=1,
* we can return immediately, since we know we'll need to wait to
* acquire. Otherwise, wait for the other thread to finish acquiring
* the fs lock */
if (nonblock) {
code = EBUSY;
} else {
while ((dl->flags & VDISKLOCK_ACQUIRING)) {
CV_WAIT(&dl->cv, &dl->mutex);
}
}
}
if (code == 0 && !(dl->flags & VDISKLOCK_ACQUIRED)) {
/* no other thread holds the lock on the actual file; so grab one */
/* first try, don't block on the lock to see if we can get it without
* waiting */
code = VLockFileLock(dl->lockfile, dl->offset, locktype, 1);
if (code == EBUSY && !nonblock) {
/* mark that we are waiting on the fs lock */
dl->flags |= VDISKLOCK_ACQUIRING;
MUTEX_EXIT(&dl->mutex);
code = VLockFileLock(dl->lockfile, dl->offset, locktype, nonblock);
MUTEX_ENTER(&dl->mutex);
dl->flags &= ~VDISKLOCK_ACQUIRING;
if (code == 0) {
dl->flags |= VDISKLOCK_ACQUIRED;
}
CV_BROADCAST(&dl->cv);
}
}
if (code) {
if (locktype == READ_LOCK) {
ReleaseReadLock(&dl->rwlock);
} else {
ReleaseWriteLock(&dl->rwlock);
}
} else {
/* successfully got the lock, so inc the number of unlocks we need
* to do before we can unlock the actual file */
++dl->lockers;
}
MUTEX_EXIT(&dl->mutex);
return code;
}