Re: [PATCH] s390/uv: Rework page allocation for guest variable storage area

[Christian Borntraeger]

Am 15.09.26 um 16:43 schrieb Heiko Carstens:
> Alexander Gordeev reported that the allocation of pages of the
> guest variable storage area via the uv_alloc_range_cb() callback
> of apply_to_page_range() will cause problems as soon as s390 fully
> supports lazy mmu mode.
>
> Problem is that uv_alloc_range_cb() allocates pages with GFP_KERNEL while
> preemption would be disabled. This would result in reports like this:
>
>   BUG: sleeping function called from invalid context at ./include/linux/sched/mm.h:322
>   in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 6092, name: qemu-kvm
>   preempt_count: 1, expected: 0
> ...
>   Call Trace:
>    dump_stack_lvl+0xae/0x108
>    __might_resched+0x1de/0x2f0
>    prepare_alloc_pages+0x1ba/0x220
>    __alloc_frozen_pages_noprof+0xc6/0x390
>    alloc_pages_mpol+0xe6/0x220
>    alloc_frozen_pages_noprof+0x5c/0x80
>    alloc_pages_noprof+0x22/0x80
>    uv_alloc_range_cb+0x30/0x300
>    apply_to_pte_range+0x11a/0x3b0
>    apply_to_pmd_range+0x13a/0x250
>    __apply_to_page_range+0x232/0x4d0
>    apply_to_page_range+0x28/0x40
>    uv_alloc_stor_var+0x5e/0x90
>    kvm_s390_pv_alloc_vm+0x118/0x1e0 [kvm]
>    kvm_s390_pv_init_vm+0x84/0x2e0 [kvm]
>    kvm_s390_handle_pv+0x502/0xf70 [kvm]
>    kvm_arch_vm_ioctl+0x234/0xdd0 [kvm]
>    kvm_vm_ioctl+0x33a/0x890 [kvm]
>    __s390x_sys_ioctl+0xfa/0x130
>    __do_syscall+0x1fc/0x700
>    system_call+0x72/0x90
>
> Preallocate the pages and make use of vm_area_map_pages() to address this
> potential future bug.
>
> Reported-by: Alexander Gordeev <agordeev@xxxxxxxxxxxxx>
> Closes: https://lore.kernel.org/all/e9449c63-b8d2-4ed9-a184-7dcab3e8dfce-agordeev@xxxxxxxxxxxxx
> Signed-off-by: Heiko Carstens <hca@xxxxxxxxxxxxx>


Thanks applied. Will queue for the kvms390 tree after some CI runs.