On 5/24/23 23:36, Luis Chamberlain wrote:
> Add support to use the new kread_uniq_fd() to avoid duplicate kernel
> reads on modules. At the cost of about ~945 bytes to your kernel size,
> enabling this on a 255 CPU x86_64 qemu guest this saves about ~1.8 GiB
> of memory during boot which would otherwise be free'd, and reduces boot
> time by about ~11 seconds.
>
> Userspace loads modules through finit_module(), this in turn will
> use vmalloc space up to 3 times:
>
> a) The kernel_read_file() call
> b) Optional module decompression
> c) Our final copy of the module
>
> Commit 064f4536d139 ("module: avoid allocation if module is already
> present and ready") shows a graph of the amount of vmalloc space
> observed allocated but freed for duplicate module request which end
> up in the trash bin. Since there is a linear relationship with the
> number of CPUs eventually this will bite us and you end up not being
> able to boot. That commit put a stop gap for c) but to avoid the
> vmalloc() space wasted on a) and b) we need to detect duplicates
> earlier.
>
> We could just have userspace fix this, but as reviewed at LSFMM 2023
> this year in Vancouver, fixing this in userspace can be complex and we
> also can't know when userpace is fixed. Fixing this in kernel turned
> out to be easy with the inode and with a simple kconfig option we can
> let users / distros decide if this full stop gap is worthy to enable.
kmod normally uses finit_module() only if a module is not compressed,
otherwise it decompresses it first and then invokes init_module().
Looking at Fedora and openSUSE Tumbleweed, they compress kernel modules
with xz and zstd, respectively. They also have their kernels built
without any CONFIG_MODULE_COMPRESS_{GZIP,XZ,ZSTD} options.
It means that these and similarly organized distributions end up using
init_module(), and adding complexity to optimize finit_module() wouldn't
actually help in their case.
-- Petr
