Running a create operation with --write-zeros can randomly hang
forever waiting for child processes. This happens roughly on in
ten runs with when running with small (20MB) loop devices.
The bug is caused by the fact that signals can be coallesced into
one if they are not read by signalfd quick enough. So if two children
finish at exactly the same time, only one SIGCHLD will be received
by the parent.
To fix this, wait on all processes with WNOHANG every time a SIGCHLD
is recieved and exit when all processes have been waited on.
Reported-by: Xiao Ni <xni@xxxxxxxxxx>
Signed-off-by: Logan Gunthorpe <logang@xxxxxxxxxxxx>
---
Create.c | 28 +++++++++++++++-------------
1 file changed, 15 insertions(+), 13 deletions(-)
diff --git a/Create.c b/Create.c
index d033eb68f30c..4f992a22b7c9 100644
--- a/Create.c
+++ b/Create.c
@@ -178,6 +178,7 @@ static int wait_for_zero_forks(int *zero_pids, int count)
bool interrupted = false;
sigset_t sigset;
ssize_t s;
+ pid_t pid;
for (i = 0; i < count; i++)
if (zero_pids[i])
@@ -196,7 +197,7 @@ static int wait_for_zero_forks(int *zero_pids, int count)
return 1;
}
- while (1) {
+ while (wait_count) {
s = read(sfd, &fdsi, sizeof(fdsi));
if (s != sizeof(fdsi)) {
pr_err("Invalid signalfd read: %s\n", strerror(errno));
@@ -209,23 +210,24 @@ static int wait_for_zero_forks(int *zero_pids, int count)
pr_info("Interrupting zeroing processes, please wait...\n");
interrupted = true;
} else if (fdsi.ssi_signo == SIGCHLD) {
- if (!--wait_count)
- break;
+ for (i = 0; i < count; i++) {
+ if (!zero_pids[i])
+ continue;
+
+ pid = waitpid(zero_pids[i], &wstatus, WNOHANG);
+ if (pid <= 0)
+ continue;
+
+ zero_pids[i] = 0;
+ if (!WIFEXITED(wstatus) || WEXITSTATUS(wstatus))
+ ret = 1;
+ wait_count--;
+ }
}
}
close(sfd);
- for (i = 0; i < count; i++) {
- if (!zero_pids[i])
- continue;
-
- waitpid(zero_pids[i], &wstatus, 0);
- zero_pids[i] = 0;
- if (!WIFEXITED(wstatus) || WEXITSTATUS(wstatus))
- ret = 1;
- }
-
if (interrupted) {
pr_err("zeroing interrupted!\n");
return 1;
--
2.39.2
