bfs/exec.c: fix failure with large 64k ppc64le page size

[mksully22]

test_exec_plus test fails because bfs_exec_arg_max() call to sysconf(_SC_ARG_MAX) is not large enough to contain two 64K pages (ppc64le default page size). Since maximum argument size allowed by BFS is only 16K anyway (BFS_EXEC_ARG_MAX) then update arg size check to make sure it is at least 32k. Note that this is occurring on Alpine Linux edge build.

[tavianator]

Thanks for your pull request! A few comments:

test_exec_plus test fails because bfs_exec_arg_max() call to sysconf(_SC_ARG_MAX) is not large enough to contain two 64K pages (ppc64le default page size)

Any idea why it's so low? Note that the build seems to be fine on Debian ppc64, for example, so there might be some weird configuration in Alpine Linux. I'd be curious to see the output of

bfs -D exec -exec echo {} + -quit

and maybe

xargs --show-limits

(with GNU xargs) on an Alpine ppc64 system.

Since maximum argument size allowed by BFS is only 16K anyway (BFS_EXEC_ARG_MAX)

BFS_EXEC_ARG_MAX is 16 MiB (16*1024*1024).

then update arg size check to make sure it is at least 32k. Note that this is occurring on Alpine Linux edge build.

I'm a bit worried this is just papering over some other issue. I'd like to see the debugging output requested above before deciding the best way to fix this.

[tavianator]

Style: this should just be

} else if (page_size > BFS_EXEC_ARG_MAX/1024) {

(but see the other comments)

[mksully22]

~/bfs $ /home/mksully/aports/testing/bfs/src/bfs-1.2.2/bfs -D exec -exec echo {} + -quit
-exec: ARG_MAX: 131072 according to sysconf()
-exec: ARG_MAX: 130772 remaining after environment variables
-exec: ARG_MAX: 130751 remaining after fixed arguments
-exec: ARG_MAX: 97983 remaining after headroom
-exec: ARG_MAX: 97983 final value
-exec: Finishing execution, executing buffered command
-exec: Executing 'echo' ... [1 arguments] (size 10)

On Alpine xargs doesn't support the --show-limits options.

[tavianator]

Ah okay, the limit starts at only 128K, so obviously subtracting two 64K pages from that will leave no space. (That suggests that Alpine ppc64le has only a 1MiB limit by default which seems rather low, but whatever.)

[tavianator]

*a 1MiB stack limit

[tavianator]

Not a big fan of variable names that lie. It's not super obvious what the point of messing with page_size in this way is -- I guess the idea is to avoid subtracting an excessive amount of headroom? Instead, maybe it's better to raise the lower limit at the end, like this:

	if (arg_max < page_size) {
		arg_max = page_size;
	} else if (arg_max > BFS_EXEC_ARG_MAX) {
		arg_max = BFS_EXEC_ARG_MAX;
	}

That's not as dangerous as it would have been in the past, since bfs now detects and recovers from E2BIG if the limit really was super low.

[mksully22]

Yep, that is a much better and more straightforward solution. I tested it and it works fine on Alpine ppc64le. Do you want me to modify the PR?

[tavianator]

Great, thanks for testing! No need for you to update this PR, I ended up taking a slightly different approach in 30b4874. If that doesn't work let me know!

[tavianator]

The issue should be fixed by 30b4874