On Fri, Jan 28, 2022 at 3:23 AM Eric Dumazet [off-list ref] wrote:

On Thu, Jan 27, 2022 at 6:22 PM Jakub Kicinski [off-list ref] wrote:

quoted

On Fri, 28 Jan 2022 03:14:14 +0100 Jann Horn wrote:

quoted

Oh. Whoops. That's what I get for only testing without CONFIG_PCPU_DEV_REFCNT...

I guess a better place to put the new check would be directly after
checking for "dev->reg_state == NETREG_UNREGISTERING"? Like this?

Possibly a very silly suggestion but perhaps we should set
the pointer to NULL for the pcpu case and let it crash?

I like that idea... but this_cpu_dec()/this_cpu_inc() use GS-relative
addressing, at least on X86-64, so NULL might make things worse, I
think? /proc/kallsyms on my machine starts with:

0000000000000000 A fixed_percpu_data
0000000000000000 A __per_cpu_start
0000000000001000 A cpu_debug_store
0000000000002000 A irq_stack_backing_store
0000000000006000 A cpu_tss_rw
000000000000b000 A gdt_page
000000000000c000 A exception_stacks
0000000000010000 A entry_stack_storage
0000000000011000 A espfix_waddr

So we'd probably need some different placeholder instead of NULL to
actually crash...

It is already set to 0

I think he meant do it already in netdev_run_todo(), not just in free_netdev()?