Extends:
- RFC0015: Job Control and Monitoring APIs
- Adds attributes to register files and directories for cleanup upon client termination
Register Cleanup of Files and Directories
Application processes frequently need to create temporary files and directories (e.g., for shared memory backing) that need to be cleaned up upon termination. This RFC provides a mechanism by which the process can register files or directories for post-termination cleanup by the PMIx server.
[ATTRIBUTES][EXTENSION]
[APPROVED]
Copyright (c) 2017-2018 Intel, Inc. All rights reserved.
This document is subject to all provisions relating to code contributions to the PMIx community as defined in the community's LICENSE file. Code Components extracted from this document must include the License text as described in that file.
Application processes frequently need to create temporary files and directories (e.g., for shared memory backing) that need to be cleaned up upon termination. This RFC provides a mechanism by which the process can register files or directories for post-termination cleanup by the PMIx server.
Processes will register files/directories for post-termination removal using the PMIx_Job_control_nb interface via the following attributes:
#define PMIX_REGISTER_CLEANUP "pmix.reg.cleanup" // (char*) comma-delimited list of files to
// be removed upon process termination
#define PMIX_REGISTER_CLEANUP_DIR "pmix.reg.cleanupdir" // (char*) comma-delimited list of directories to
// be removed upon process termination
#define PMIX_CLEANUP_IGNORE "pmix.cleanup.ignore" // (char*) comma-delimited list of files or
// directories that are not to be removed
All directives in a request will pertain to all registered cleanup directories included in that request. Directives supported for this request include:
#define PMIX_CLEANUP_RECURSIVE "pmix.cleanup.recurse" // (bool) recursively traverse subdirectories under the
// specified one(s)
#define PMIX_CLEANUP_LEAVE_TOPDIR "pmix.cleanup.lvtop" // (bool) when recursively cleaning subdirs, do not remove
// the top-level directory (the one given in the
// cleanup request)
Multiple requests can be issued by a given client and/or in a single call to PMIx_Job_control_nb - the PMIx server will aggregate them upon receipt. Duplicate file requests will return an error if they conflict regarding removal vs ignore, but will otherwise return success (with the duplicate entry dropped). Any duplicate directory-targeted requests will be de-duplicated by the server using the following conflict resolution rules:
-
a directive to recursively traverse subdirectories will take precedence over any duplicate request that does not include that directive
-
a directive to leave the top-level directory will take precedence over any duplicate request that does not so specify
Note that the PMIx server is not responsible for checking overlap between directory-targeted requests. Thus, a request involving a directory that is underneath another directory involvced in a request will be treated as non-duplicate.
The PMIx_Job_control_nb API takes an array of pmix_proc_t structures to identify the target processes to be impacted by the requested control operation. Registering for cleanup shall be strictly a LOCAL operation. Hence, any such requests shall not be relayed to the local host RM for processing, but instead will be handled by the local PMIx server itself.
Requests that specify the requesting process for the pmix_proc_t argument shall refer solely to the requesting process itself. The PMIx server shall implement such requests upon termination of that process. Requests that include a given namespace with PMIX_RANK_WILDCARD shall be executed upon termination of all local clients from that nspace. Likewise, a request that provides a NULL for the pmix_proc_t argument shall be executed upon termination of all local clients from the nspace of the requestor.
Note that filenames and directories not including an absolute path are considered ambiguous and will result in return of an error - this is done to avoid confusion over location versus the working directory of the PMIx server.
Requests to cleanup subdirectories will proceed by:
-
remove all specified files - i.e., non-directories;
-
traverse the directory tree under each specified target location, removing all files not specified as "to be ignored". Note that requests specifying a file as "to be removed" that is also specified (either as part of the same request, or in either an earlier or later request) as "to be ignored" is considered contradictory and will return an error for the contradicting request; and
-
directories shall be removed if empty, starting from the deepest part of the directory tree and working upwards.
Files and/or directories that cannot be removed (e.g., due to a permissions problem) will be silently ignored, and only files and/or directories with matching effective uid and gid of the requesting client process will be considered - thus, files owned by other users and groups will be automatically ignored. Printed output indicating that this has occurred may be available from the PMIx library implementation (e.g., in the case of the PMIx Reference Implementation, upon setting PMIX_DEBUG to a value of at least 10 in the PMIx server's local environment).
#include <pmix.h>
#include <pmix_common.h>
typedef struct {
pthread_mutex_t mutex;
pthread_cond_t cond;
volatile bool active;
pmix_status_t status;
} mylock_t;
#define DEBUG_CONSTRUCT_LOCK(l) \
do { \
pthread_mutex_init(&(l)->mutex, NULL); \
pthread_cond_init(&(l)->cond, NULL); \
(l)->active = true; \
(l)->status = PMIX_SUCCESS; \
} while(0)
#define DEBUG_DESTRUCT_LOCK(l) \
do { \
pthread_mutex_destroy(&(l)->mutex); \
pthread_cond_destroy(&(l)->cond); \
} while(0)
#define DEBUG_WAIT_THREAD(lck) \
do { \
pthread_mutex_lock(&(lck)->mutex); \
while ((lck)->active) { \
pthread_cond_wait(&(lck)->cond, &(lck)->mutex); \
} \
pthread_mutex_unlock(&(lck)->mutex); \
} while(0)
#define DEBUG_WAKEUP_THREAD(lck) \
do { \
pthread_mutex_lock(&(lck)->mutex); \
(lck)->active = false; \
pthread_cond_broadcast(&(lck)->cond); \
pthread_mutex_unlock(&(lck)->mutex); \
} while(0)
static void cbfunc(pmix_status_t status,
pmix_info_t *info, size_t ninfo,
void *cbdata,
pmix_release_cbfunc_t release_fn,
void *release_cbdata)
{
mylock_t *lock = (mylock_t*)cbdata;
lock->status = status;
/* let the library release the data and cleanup from
* the operation */
if (NULL != release_fn) {
release_fn(release_cbdata);
}
/* release the block */
DEBUG_WAKEUP_THREAD(lock);
}
int main (int argc, char **argv)
{
pmix_proc_t myproc, proc;
pmix_info_t *info;
mylock_t lock;
PMIx_Init(&myproc, NULL, 0);
/* request to cleanup a process-specific tmp directory tree and a specific shmem backing file */
PMIX_INFO_CREATE(info, 5);
PMIX_INFO_LOAD(&info[0], PMIX_REGISTER_CLEANUP_DIR, "/mytmpdir", PMIX_STRING);
PMIX_INFO_LOAD(&info[1], PMIX_REGISTER_CLEANUP, "/tmp/dev_shm/mybackfile", PMIX_STRING);
/* recursively cleanup subdirectories */
PMIX_INFO_LOAD(&info[2], PMIX_CLEANUP_RECURSIVE, NULL, PMIX_BOOL);
/* ignore a file used for debugging output, if present */
PMIX_INFO_LOAD(&info[3], PMIX_CLEANUP_IGNORE, "/mytmpdir/output-124.txt", PMIX_STRING);
/* ignore a file containing contact info, if present */
PMIX_INFO_LOAD(&info[4], PMIX_CLEANUP_IGNORE, "/mytmpdir/subdir/contact.info", PMIX_STRING);
DEBUG_CONSTRUCT_LOCK(&lock);
/* pass the request - note that this request will be serviced upon
* my termination */
rc = PMIx_Job_control_nb(&myproc, 1, info, 5, cbfunc, (void*)&lock);
if (PMIX_SUCCESS != rc) {
fprintf(stderr, "Job control failed\n");
PMIX_INFO_FREE(info, 5);
DEBUG_DESTRUCT_LOCK(&lock);
goto done;
}
DEBUG_WAIT_THREAD(&lock);
rc = lock.status;
PMIX_INFO_FREE(info, 5);
DEBUG_DESTRUCT_LOCK(&lock);
/* now register to have my job-level tmp directory tree
* cleaned up. Any overlapping requests by my peers will
* be de-duplicated by the server so this cleanup occurs
* only once when all local job clients have terminated */
PMIX_INFO_CREATE(info, 2);
PMIX_INFO_LOAD(&info[0], PMIX_REGISTER_CLEANUP_DIR, "/myjobtmpdir", PMIX_STRING);
/* recursively cleanup subdirectories */
PMIX_INFO_LOAD(&info[1], PMIX_CLEANUP_RECURSIVE, NULL, PMIX_BOOL);
/* we want this to refer to the entire local job, not just us */
PMIX_PROC_CONSTRUCT(&proc);
(void)strncpy(proc.nspace, myproc.nspace, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
rc = PMIx_Job_control_nb(&proc, 1, info, 2, cbfunc, (void*)&lock);
if (PMIX_SUCCESS != rc) {
fprintf(stderr, "Job control failed\n");
PMIX_INFO_FREE(info, 2);
DEBUG_DESTRUCT_LOCK(&lock);
goto done;
}
DEBUG_WAIT_THREAD(&lock);
rc = lock.status;
PMIX_INFO_FREE(info, 2);
DEBUG_DESTRUCT_LOCK(&lock);
...
done:
PMIx_Finalize(NULL, 0);
return rc;
}
Prototype implementation available in PMIx master repo in Pull Request Add registration for cleanup support, and in Open MPI PR Update to PMIx v3.0 PR for cleanup registration
Ralph H. Castain
Intel, Inc.
Github: rhc54