See attached patch. When the shared memory segment cannot be mapped,
for example because GFORTRAN_SHARED_MEMORY_SIZE does not fit under the
address-space limit, a caf_shmem program dies with a segmentation fault
and gives no hint of the cause. shared_memory_init tested the mmap
result for NULL, which mmap does not return on failure. The patch
tests for MAP_FAILED first, so the supervisor and the images report the
mmap error and exit with status 1.
The change is one hunk in shared_memory_init in
libgfortran/caf/shmem/shared_memory.c. It also drops the old NULL
test, which a failed mmap never triggered.
There is no new test: caf.exp runs every test under each variant, and
the failure needs a resource limit. To reproduce:
program t
integer :: x[*]
x = this_image ()
sync all
if (this_image () == 1) print *, 'images:', num_images ()
end program
$ gfortran -fcoarray=lib t.f90 -lcaf_shmem -o t
$ (ulimit -v 8000000; GFORTRAN_NUM_IMAGES=4 \
GFORTRAN_SHARED_MEMORY_SIZE=64G ./t; echo $?)
Without the patch this ends in a segmentation fault (exit status 139).
With it the program prints "mmap failed: Cannot allocate memory" and
exits with status 1.
Regression tested on x86_64.
I plan to commit as simple if not obvious.
Regards,
Jerry
---
libgfortran: caf_shmem: check mmap against MAP_FAILED
shared_memory_init compared the result of mmap with NULL, but mmap
reports failure with MAP_FAILED. A failed mmap in the supervisor went
unnoticed until the first access to the shared memory segfaulted. In an
image it looked like a mapping at the wrong address, so the supervisor
restarted the image until it gave up with "no common state on shared
memory could be reached". Check for MAP_FAILED before comparing the
address, and exit with a message.
Assisted-by: Claude Opus 5.5
libgfortran/ChangeLog:
* caf/shmem/shared_memory.c (shared_memory_init): Check the result
of mmap against MAP_FAILED before comparing it with the requested
address.
>From b1f1a62d2e6ba98ab247bf55e072dc8b4e1c9da1 Mon Sep 17 00:00:00 2001
From: Jerry DeLisle <[email protected]>
Date: Tue, 17 Nov 2026 17:09:34 -0800
Subject: [PATCH] libgfortran: caf_shmem: check mmap against MAP_FAILED
shared_memory_init compared the result of mmap with NULL, but mmap
reports failure with MAP_FAILED. A failed mmap in the supervisor went
unnoticed until the first access to the shared memory segfaulted. In an
image it looked like a mapping at the wrong address, so the supervisor
restarted the image until it gave up with "no common state on shared
memory could be reached". Check for MAP_FAILED before comparing the
address, and exit with a message.
Assisted-by: Claude Opus 5.5
libgfortran/ChangeLog:
* caf/shmem/shared_memory.c (shared_memory_init): Check the result
of mmap against MAP_FAILED before comparing it with the requested
address.
---
libgfortran/caf/shmem/shared_memory.c | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)
diff --git a/libgfortran/caf/shmem/shared_memory.c b/libgfortran/caf/shmem/shared_memory.c
index 281336c80bc..dc78f66e820 100644
--- a/libgfortran/caf/shmem/shared_memory.c
+++ b/libgfortran/caf/shmem/shared_memory.c
@@ -256,6 +256,11 @@ shared_memory_init (shared_memory_act *mem, size_t size)
#ifdef HAVE_MMAP
mem->glbl.base
= mmap (base_ptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, mem->shm_fd, 0);
+ if (mem->glbl.base == MAP_FAILED)
+ {
+ perror ("mmap failed");
+ exit (1);
+ }
if (base_ptr && mem->glbl.base != base_ptr)
{
/* The supervisor will start us again. */
@@ -263,11 +268,6 @@ shared_memory_init (shared_memory_act *mem, size_t size)
free (local);
exit (210);
}
- else if (!base_ptr && !mem->glbl.base)
- {
- perror ("mmap failed");
- exit (1);
- }
#ifdef MADV_DONTDUMP
/* Keep the unlinked object, which counts as anonymous shared memory, out
of core dumps. */
--
2.56.0