On 14.10.22 08:09, Miroslav Benes wrote:

Hi,

On Thu, 13 Oct 2022, David Hildenbrand wrote:

quoted

We already make sure to allocate percpu data only after we verified that
the module we're loading hasn't already been loaded and isn't
concurrently getting loaded -- that it's unique.

On big systems (> 400 CPUs and many devices) with KASAN enabled, we're now
phasing a similar issue with the module vmap space.

When KASAN_INLINE is enabled (resulting in large module size), plenty
of devices that udev wants to probe and plenty (> 400) of CPUs that can
carry out that probing concurrently, we can actually run out of module
vmap space and trigger vmap allocation errors:

[  165.818200] vmap allocation for size 2498560 failed: use vmalloc=<size> to increase size
[  165.836622] vmap allocation for size 315392 failed: use vmalloc=<size> to increase size
[  165.837461] vmap allocation for size 315392 failed: use vmalloc=<size> to increase size
[  165.840573] vmap allocation for size 2498560 failed: use vmalloc=<size> to increase size
[  165.841059] vmap allocation for size 2498560 failed: use vmalloc=<size> to increase size
[  165.841428] vmap allocation for size 2498560 failed: use vmalloc=<size> to increase size
[  165.841819] vmap allocation for size 2498560 failed: use vmalloc=<size> to increase size
[  165.842123] vmap allocation for size 2498560 failed: use vmalloc=<size> to increase size
[  165.843359] vmap allocation for size 2498560 failed: use vmalloc=<size> to increase size
[  165.844894] vmap allocation for size 2498560 failed: use vmalloc=<size> to increase size
[  165.847028] CPU: 253 PID: 4995 Comm: systemd-udevd Not tainted 5.19.0 #2
[  165.935689] Hardware name: Lenovo ThinkSystem SR950 -[7X12ABC1WW]-/-[7X12ABC1WW]-, BIOS -[PSE130O-1.81]- 05/20/2020
[  165.947343] Call Trace:
[  165.950075]  <TASK>
[  165.952425]  dump_stack_lvl+0x57/0x81
[  165.956532]  warn_alloc.cold+0x95/0x18a
[  165.960836]  ? zone_watermark_ok_safe+0x240/0x240
[  165.966100]  ? slab_free_freelist_hook+0x11d/0x1d0
[  165.971461]  ? __get_vm_area_node+0x2af/0x360
[  165.976341]  ? __get_vm_area_node+0x2af/0x360
[  165.981219]  __vmalloc_node_range+0x291/0x560
[  165.986087]  ? __mutex_unlock_slowpath+0x161/0x5e0
[  165.991447]  ? move_module+0x4c/0x630
[  165.995547]  ? vfree_atomic+0xa0/0xa0
[  165.999647]  ? move_module+0x4c/0x630
[  166.003741]  module_alloc+0xe7/0x170
[  166.007747]  ? move_module+0x4c/0x630
[  166.011840]  move_module+0x4c/0x630
[  166.015751]  layout_and_allocate+0x32c/0x560
[  166.020519]  load_module+0x8e0/0x25c0
[  166.024623]  ? layout_and_allocate+0x560/0x560
[  166.029586]  ? kernel_read_file+0x286/0x6b0
[  166.034269]  ? __x64_sys_fspick+0x290/0x290
[  166.038946]  ? userfaultfd_unmap_prep+0x430/0x430
[  166.044203]  ? lock_downgrade+0x130/0x130
[  166.048698]  ? __do_sys_finit_module+0x11a/0x1c0
[  166.053854]  __do_sys_finit_module+0x11a/0x1c0
[  166.058818]  ? __ia32_sys_init_module+0xa0/0xa0
[  166.063882]  ? __seccomp_filter+0x92/0x930
[  166.068494]  do_syscall_64+0x59/0x90
[  166.072492]  ? do_syscall_64+0x69/0x90
[  166.076679]  ? do_syscall_64+0x69/0x90
[  166.080864]  ? do_syscall_64+0x69/0x90
[  166.085047]  ? asm_sysvec_apic_timer_interrupt+0x16/0x20
[  166.090984]  ? lockdep_hardirqs_on+0x79/0x100
[  166.095855]  entry_SYSCALL_64_after_hwframe+0x63/0xcd[  165.818200] vmap allocation for size 2498560 failed: use vmalloc=<size> to increase size

Interestingly, when reducing the number of CPUs (nosmt), it works as
expected.

The underlying issue is that we first allocate memory (including module
vmap space) in layout_and_allocate(), and then verify whether the module
is unique in add_unformed_module(). So we end up allocating module vmap
space even though we might not need it -- which is a problem when modules
are big and we can have a lot of concurrent probing of the same set of
modules as on the big system at hand.

Unfortunately, we cannot simply add the module earlier, because
move_module() -- that allocates the module vmap space -- essentially
brings the module to life from a temporary one. Adding the temporary one
and replacing it is also sub-optimal (because replacing it would require
to synchronize against RCU) and feels kind of dangerous judging that we
end up copying it.

So instead, add a second list (pending_load_infos) that tracks the modules
(via their load_info) that are unique and are still getting loaded
("pending"), but haven't made it to the actual module list yet. This
shouldn't have a notable runtime overhead when concurrently loading
modules: the new list is expected to usually either be empty or contain
very few entries for a short time.

Thanks to Uladzislau for his help to verify that it's not actually a
vmap code issue.

this seems to be related to what
https://lore.kernel.org/all/20220919123233.8538-1-petr.pavlu@suse.com/ (local)
tries to solve. Just your symptoms are different. Does the patch set fix
your issue too?

Hi Miroslav,

the underlying approach with a load_info list is similar (which is nice 
to see), so I assume it will similarly fix the issue.

I'm not sure if merging the requests (adding the refcount logic and the 
-EBUSY change is really required/wanted), though. Looks like some of 
these changes that might have been factored out into separate patches.

Not my call to make. I'll give the set a churn on the machine where I 
can reproduce the issue.

-- 
Thanks,

David / dhildenb