From af6b702bd96a7f16c5e4321a6f4dc1d0f05a720d Mon Sep 17 00:00:00 2001
From: Mark Grondona <mgrondona@llnl.gov>
Date: Wed, 23 Oct 2002 20:49:37 +0000
Subject: [PATCH] o typo fixes. better shm_init() failure handling.
---
src/slurmd/elan_interconnect.c | 3 ---
src/slurmd/io.c | 49 +++++++++++++++++-----------------
src/slurmd/mgr.c | 2 ++
src/slurmd/shm.c | 48 ++++++++++++++++++++-------------
src/slurmd/slurmd.c | 3 ++-
5 files changed, 58 insertions(+), 47 deletions(-)
diff --git a/src/slurmd/elan_interconnect.c b/src/slurmd/elan_interconnect.c
index 2032eb604f3..9a9739069f4 100644
--- a/src/slurmd/elan_interconnect.c
+++ b/src/slurmd/elan_interconnect.c
@@ -52,8 +52,6 @@ _wait_and_destroy_prg(qsw_jobinfo_t qsw_job, pid_t pid)
int i = 0;
int sleeptime = 1;
- shm_init();
-
debug3("waiting to destory program description...");
again:
if (waitpid(pid, NULL, 0) < 0) {
@@ -80,7 +78,6 @@ _wait_and_destroy_prg(qsw_jobinfo_t qsw_job, pid_t pid)
sleep(sleeptime*=2);
}
- shm_fini();
exit(0);
return SLURM_SUCCESS;
}
diff --git a/src/slurmd/io.c b/src/slurmd/io.c
index 5ca27c814df..11894afc89d 100644
--- a/src/slurmd/io.c
+++ b/src/slurmd/io.c
@@ -63,9 +63,8 @@ typedef enum slurmd_io_tupe {
CLIENT_STDOUT,
} slurmd_io_type_t;
-static char *slurmd_io_str[] =
+static char *_io_str[] =
{
- "domain socket",
"task stderr",
"task stdout",
"task stdin",
@@ -332,8 +331,8 @@ _validate_task_out(struct io_info *t, int type)
while ((r = list_next(i))) {
if (r->type != type) {
fatal("_validate_io: %s reader is %s",
- slurmd_io_str[t->type],
- slurmd_io_str[r->type]);
+ _io_str[t->type],
+ _io_str[r->type]);
}
}
list_iterator_destroy(i);
@@ -357,13 +356,13 @@ _validate_task_in(struct io_info *t)
while ((r = list_next(i)) != NULL) {
if (r->magic != IO_MAGIC) {
error("_validate_io: %s writer is invalid",
- slurmd_io_str[t->type]);
+ _io_str[t->type]);
return 0;
}
if (r->type != CLIENT_STDOUT) {
error("_validate_io: %s writer is %s",
- slurmd_io_str[t->type],
- slurmd_io_str[r->type]);
+ _io_str[t->type],
+ _io_str[r->type]);
retval = 0;
}
}
@@ -386,7 +385,7 @@ _validate_client_stdout(struct io_info *client)
while ((t = list_next(i))) {
if (t->type != TASK_STDIN) {
error("_validate_io: client stdin reader is %s",
- slurmd_io_str[t->type]);
+ _io_str[t->type]);
retval = 0;
}
}
@@ -396,7 +395,7 @@ _validate_client_stdout(struct io_info *client)
while ((t = list_next(i))) {
if (t->type != TASK_STDOUT) {
error("_validate_io: client stdout writer is %s",
- slurmd_io_str[t->type]);
+ _io_str[t->type]);
retval = 0;
}
}
@@ -421,7 +420,7 @@ _validate_client_stderr(struct io_info *client)
while ((t = list_next(i))) {
if (t->type != TASK_STDERR) {
error("_validate_io: client stderr writer is %s",
- slurmd_io_str[t->type]);
+ _io_str[t->type]);
retval = 0;
}
}
@@ -691,8 +690,7 @@ _readable(io_obj_t *obj)
xassert(io->magic == IO_MAGIC);
- if ((rc = (!io->disconnected && !io->eof && (obj->fd > 0))))
- debug3("readable %s", slurmd_io_str[io->type]);
+ rc = (!io->disconnected && !io->eof && (obj->fd > 0));
return rc;
}
@@ -707,8 +705,7 @@ _writable(io_obj_t *obj)
rc = (!io->disconnected
&& ((cbuf_used(io->buf) > 0) || io->eof));
- if (rc)
- debug3("writable %s", slurmd_io_str[io->type]);
+
return rc;
}
@@ -725,7 +722,7 @@ _write(io_obj_t *obj, List objs)
return 0;
verbose("Need to write %ld bytes to %s %d",
- cbuf_used(io->buf), slurmd_io_str[io->type], io->id);
+ cbuf_used(io->buf), _io_str[io->type], io->id);
if (io->eof && (cbuf_used(io->buf) == 0)) {
@@ -741,16 +738,18 @@ _write(io_obj_t *obj, List objs)
}
while ((n = cbuf_read_fd(io->buf, obj->fd, -1)) < 0) {
- int local_errno = errno;
if ((errno == EAGAIN) || (errno == EWOULDBLOCK))
continue;
- error("task <%ld> write failed: %s", io->id,
- slurm_strerror(local_errno));
+ error("task <%ld> write failed: %m", io->id);
+ if (io->type == CLIENT_STDERR || io->type == CLIENT_STDOUT)
+ _io_disconnect_client(io);
+ else
+ _shutdown_task_obj(io);
return -1;
}
verbose("Wrote %d bytes to %s %d",
- n, slurmd_io_str[io->type], io->id);
+ n, _io_str[io->type], io->id);
return 0;
}
@@ -796,7 +795,7 @@ _task_read(io_obj_t *obj, List objs)
goto again;
if ((errno == EAGAIN) || (errno == EWOULDBLOCK)) {
error("%s %d: read returned EAGAIN",
- slurmd_io_str[t->type], t->id);
+ _io_str[t->type], t->id);
return 0;
}
error("Unable to read from task %ld fd %d errno %d %m",
@@ -804,7 +803,7 @@ _task_read(io_obj_t *obj, List objs)
return -1;
}
verbose("read %d bytes from %s %d",
- n, slurmd_io_str[t->type], t->id);
+ n, _io_str[t->type], t->id);
if (n == 0) { /* got eof */
verbose("got eof on task %ld", t->id);
@@ -821,7 +820,7 @@ _task_read(io_obj_t *obj, List objs)
while((r = list_next(i))) {
n = cbuf_write(r->buf, (void *) buf, n, NULL);
verbose("wrote %ld bytes into %s buf", n,
- slurmd_io_str[r->type]);
+ _io_str[r->type]);
}
list_iterator_destroy(i);
@@ -834,7 +833,7 @@ _task_error(io_obj_t *obj, List objs)
struct io_info *t = (struct io_info *) obj->arg;
xassert(t->magic == IO_MAGIC);
- error("error on %s %d", slurmd_io_str[t->type], t->id);
+ error("error on %s %d", _io_str[t->type], t->id);
_shutdown_task_obj(t);
obj->fd = -1;
list_delete_all(objs, (ListFindF) find_obj, obj);
@@ -867,7 +866,7 @@ _client_read(io_obj_t *obj, List objs)
return -1;
}
- debug("read %d bytes from %s %d", n, slurmd_io_str[client->type],
+ debug("read %d bytes from %s %d", n, _io_str[client->type],
client->id);
if (n == 0) { /* got eof, disconnect this client */
@@ -904,7 +903,7 @@ _client_error(io_obj_t *obj, List objs)
xassert(io->magic == IO_MAGIC);
- error("%s task %d", slurmd_io_str[io->type], io->id);
+ error("%s task %d", _io_str[io->type], io->id);
return 0;
}
diff --git a/src/slurmd/mgr.c b/src/slurmd/mgr.c
index a2885ab179a..1d21e197fff 100644
--- a/src/slurmd/mgr.c
+++ b/src/slurmd/mgr.c
@@ -61,6 +61,8 @@ mgr_launch_tasks(launch_tasks_request_msg_t *msg)
slurmd_job_t *job;
log_reinit();
+
+ /* New process, so we must reinit shm */
if (shm_init() < 0)
return SLURM_ERROR;
if (!(job = job_create(msg)))
diff --git a/src/slurmd/shm.c b/src/slurmd/shm.c
index da6b43d25a5..f3d8646b1e9 100644
--- a/src/slurmd/shm.c
+++ b/src/slurmd/shm.c
@@ -121,7 +121,6 @@ int
shm_fini(void)
{
int destroy = 0;
- info("process %ld detaching from shm", getpid());
xassert(slurmd_shm != NULL);
_shm_lock();
if (--slurmd_shm->users == 0)
@@ -316,17 +315,20 @@ shm_signal_step(uint32_t jobid, uint32_t stepid, uint32_t signal)
int retval = SLURM_SUCCESS;
int i;
job_step_t *s;
- task_t *t;
+ task_t *t, *tlast;
_shm_lock();
if ((i = _shm_find_step(jobid, stepid)) >= 0) {
s = &slurmd_shm->step[i];
+ tlast = NULL;
for (t = s->task_list; t; t = t->next) {
+ xassert(t != tlast);
if (t->pid > 0 && kill(t->pid, signo) < 0) {
- error("kill %d.%d pid %ld: %m",
- jobid, stepid, (long)t->pid);
+ error("kill %d.%d task %d pid %ld: %m",
+ jobid, stepid, t->id, (long)t->pid);
retval = errno;
}
+ tlast = t;
}
} else
retval = ESRCH;
@@ -550,6 +552,7 @@ shm_add_task(uint32_t jobid, uint32_t stepid, task_t *task)
slurm_seterrno_ret(ESRCH);
}
s = &slurmd_shm->step[i];
+ debug2("adding task %d to step %d.%d", task->id, jobid, stepid);
if (_shm_find_task_in_step(s, task->id)) {
_shm_unlock();
slurm_seterrno_ret(EEXIST);
@@ -578,9 +581,9 @@ _shm_find_task_in_step(job_step_t *s, int taskid)
task_t *t = NULL;
for (t = s->task_list; t && t->used; t = t->next) {
if (t->id == taskid)
- break;
+ return t;
}
- return t;
+ return NULL;
}
static task_t *
@@ -588,8 +591,10 @@ _shm_alloc_task(void)
{
int i;
for (i = 0; i < MAX_TASKS; i++) {
- if (!slurmd_shm->task[i].used)
+ if (!slurmd_shm->task[i].used) {
+ slurmd_shm->task[i].used = true;
return &slurmd_shm->task[i];
+ }
}
return NULL;
}
@@ -599,6 +604,7 @@ _shm_task_copy(task_t *to, task_t *from)
{
*to = *from;
/* next and step are not valid for copying */
+ to->used = true;
to->next = NULL;
to->job_step = NULL;
}
@@ -606,12 +612,9 @@ _shm_task_copy(task_t *to, task_t *from)
static void
_shm_step_copy(job_step_t *to, job_step_t *from)
{
- task_t *t = NULL;
- if (to->task_list)
- t = to->task_list;
*to = *from;
to->state = SLURMD_JOB_ALLOCATED;
- to->task_list = t; /* addition of tasks is another step */
+ to->task_list = NULL; /* addition of tasks is another step */
}
static void
@@ -640,9 +643,14 @@ _shm_create()
key_t key = ftok(".", 'a');
if ((shmid = shmget(key, sizeof(slurmd_shm_t), oflags)) < 0) {
- if ((shmid = shmget(key, sizeof(slurmd_shm_t), 0600)) < 0)
- error("shmget: %m");
- return SLURM_ERROR;
+ if ((shmid = shmget(key, sizeof(slurmd_shm_t), 0600)) < 0) {
+ if (errno == EINVAL) {
+ error("shm_init: Existing shm invalid. "
+ "Please remove.");
+ } else
+ error("shmget: %m");
+ return SLURM_ERROR;
+ }
}
slurmd_shm = shmat(shmid, NULL, 0);
@@ -663,13 +671,13 @@ _shm_attach()
key_t key = ftok(".", 'a');
if ((shmid = shmget(key, sizeof(slurmd_shm_t), oflags)) < 0)
- fatal("shm_attach: %m");
+ return SLURM_ERROR;
slurmd_shm = shmat(shmid, NULL, 0);
if (slurmd_shm == (void *)-1 || !slurmd_shm)
- fatal("shmat: %m");
+ return SLURM_ERROR;
- return 1;
+ return SLURM_SUCCESS;
}
/*
@@ -701,7 +709,11 @@ _shm_reopen()
}
/* Attach to shared memory region */
- _shm_attach();
+ if ((_shm_attach() < 0) && (_shm_create() < 0)) {
+ error("shm_create(): %m");
+ return SLURM_FAILURE;
+ }
+
/* Lock and unlock semaphore to ensure data is initialized */
_shm_lock();
diff --git a/src/slurmd/slurmd.c b/src/slurmd/slurmd.c
index 6c1cb963f25..e6dafb1b71e 100644
--- a/src/slurmd/slurmd.c
+++ b/src/slurmd/slurmd.c
@@ -128,7 +128,8 @@ int main(int argc, char *argv[])
}
/* shared memory init */
- slurmd_init();
+ if (slurmd_init() < 0)
+ exit (1);
if ((rc = getnodename(node_name, MAX_NAME_LEN)))
fatal("getnodename: %m");
--
GitLab