[PATCH] libgfortran: [PR127684] caf_shmem: SYNC IMAGES hangs when an, image of the set terminates
Jerry D
jvdelisle2@gmail.com
Tue Oct 6 00:58:16 GMT 2026
On 10/5/26 10:04 AM, Jerry D wrote:
> See the attached patch. In my campaign to identify caf related bugs I discovered
> this one.
>
> Regression tested under heavy loads.
>
> OK for mainline?
>
> Regards,
>
> Jerry
I am working on ways to create LLM assisted patches more human readable. Using
Claude, and I suspect other LLMs, one can set up rules and guidelines to follow.
The attached V2 of the subject patch is essentially identical with an improved
commit message that better explanations of the problem and the fix. There is a
bit of a guide here on how to review all this as well.
How the hang happens: image 1 waits in SYNC IMAGES ([2, 3]) when image
3 stops. Image 2 then executes SYNC IMAGES ([1, 3]), finds image 3
stopped and returns with STAT_STOPPED_IMAGE without synchronizing, so
image 1 is not released. FAIL IMAGE leads to the same hang.
The point to check is the order in sync_table_terminated: under the
table lock it releases the waiting images and takes back their
unmatched counts, and only then stores the new status. Outside the
WIN32 supervisor, which this patch does not change, every change of an
image's status to stopped or failed goes through it.
Reading guide: start with sync.h, the new waiting and aborted fields
and what sync_table and sync_table_terminated promise. Then those two
functions in sync.c. In shmem.c, _gfortran_caf_sync_images drops its
early return for a terminated image and sets STAT= only when sync_table
reports a terminated image. The other shmem.c and supervisor.c hunks
route the status changes through sync_table_terminated, which replaces
the supervisor's wake-up of all waiting images. set_table_parts and
the sync_init changes only lay out the enlarged table.
Regression tested on x86_64.
OK for mainline?
Regards,
Jerry
PS Feedback on the format of this email is greatly appreciated.
---
libgfortran: [PR127684] caf_shmem: SYNC IMAGES hang on
terminated image
SYNC IMAGES hung when an image of the image set stopped or failed while
another image of the set was waiting in the statement.
An image that found a stopped or failed image in its set returned at once
with STAT=, without synchronizing, so an image already waiting for it was
never released. The supervisor also woke the waiting images without
holding the table lock, so a wakeup could be lost.
Each image now records the set it waits for. Images that stop or fail,
and the POSIX supervisor when it reaps one, set the status through the
new sync_table_terminated, which holds the table lock. For a stopped
image it releases the images waiting for it and takes back the counts
they raised that no partner has matched, before it stores the status,
so an image that sees the stop cannot pair with the aborted statement.
A failed image is skipped and the other images of the set synchronize
(F2023 11.7.11).
Assisted-by: Claude Opus 5.5
PR libfortran/127684
libgfortran/ChangeLog:
* caf/shmem.c (_gfortran_caf_sync_images): Do not return early for
a stopped or failed image. Check the images only when one of them
terminated.
(mark_stopped): Use sync_table_terminated.
(_gfortran_caf_fail_image): Likewise.
* caf/shmem/supervisor.c (supervisor_main_loop): Likewise. Do not
signal all images waiting in sync_table.
* caf/shmem/sync.c (set_table_parts): New function.
(sync_init): Use it.
(sync_init_supervisor): Allocate the image sets of the waiting
images and the aborted flags.
(stopped_image_in): New function.
(sync_table): Return whether an image of the set terminated. Only
synchronize memory when an image of the set stopped. Record the
image set and end the wait when aborted.
(sync_table_terminated): New function.
* caf/shmem/sync.h (sync_t): Add waiting and aborted.
(sync_table): Return bool.
(sync_table_terminated): Declare.
gcc/testsuite/ChangeLog:
* gfortran.dg/coarray/sync_images_failed_1.f90: New test.
* gfortran.dg/coarray/sync_images_stopped_2.f90: New test.
---
-------------- next part --------------
A non-text attachment was scrubbed...
Name: PR127684-caf-shmem-sync-images-terminated-v2.patch
Type: text/x-patch
Size: 17260 bytes
Desc: not available
URL: <https://gcc.gnu.org/pipermail/gcc-patches/attachments/20261005/4deb8844/attachment-0001.bin>
More information about the Gcc-patches
mailing list