On Tue, Aug 29, 2023 at 05:23:25PM +0200, Andrea Righi wrote:
On Tue, Aug 29, 2023 at 07:47:05AM -0700, Lucas De Marchi wrote:On Tue, Aug 29, 2023 at 02:38:08PM +0200, Andrea Righi wrote: > Fallback to user-space decompression when the kernel cannot allocate > enough memory to perform the decompression. > > This can happen with large modules, such as xfs on linux 6.5 for > example, an ENOMEM would be returned and the module fails to load. > > It seems more reliable to try again with user-space decompression > rather than reporting an error and failing to load the module. > > Fixes: 09c9f8c ("libkmod: Use kernel decompression when available") > Signed-off-by: Andrea Righi <andrea.righi@xxxxxxxxxxxxx> > --- > libkmod/libkmod-module.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/libkmod/libkmod-module.c b/libkmod/libkmod-module.c > index 585da41..d2d4815 100644 > --- a/libkmod/libkmod-module.c > +++ b/libkmod/libkmod-module.c > @@ -978,7 +978,7 @@ KMOD_EXPORT int kmod_module_insert_module(struct kmod_module *mod, > } > > err = do_finit_module(mod, flags, args); > - if (err == -ENOSYS) > + if (err == -ENOSYS || err == -ENOMEM) oh... ENOMEM can be returned in several places. I don't think it's a great interface to just retry in userspace. Luis, can we do better on the kernel side? Andrea, did you track the exact location triggering the ENOMEM? Is it the return of kvmalloc_array() or alloc_page() from module_get_next_page()? Or one of the previous kmalloc we use for the different deflate methods?It was this kmalloc() right here: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/module/decompress.c?h=v6.5#n244 For the records, I've also sent this patch to address the issue from a kernel perspective: https://lore.kernel.org/lkml/20230829120508.317611-1-andrea.righi@xxxxxxxxxxxxx/T/#u
That patch seems better. We will probably have to move other decompress algorithms to vmalloc too. thanks Lucas De Marchi
That seems to solve the issue in my particular case, but I'm not sure if that's a valid solution across all architectures. Thanks, -Andrea
