From: Michael Ellerman mpe@ellerman.id.au
[ Upstream commit 591b4b268435f00d2f0b81f786c2c7bd5ef66416 ]
Paul reported a warning with DEBUG_ATOMIC_SLEEP=y:
BUG: sleeping function called from invalid context at include/linux/sched/mm.h:256 in_atomic(): 0, irqs_disabled(): 1, non_block: 0, pid: 1, name: swapper/0 preempt_count: 0, expected: 0 ... Call Trace: dump_stack_lvl+0xa0/0xec (unreliable) __might_resched+0x2f4/0x310 kmem_cache_alloc+0x220/0x4b0 __pud_alloc+0x74/0x1d0 hash__map_kernel_page+0x2cc/0x390 do_patch_instruction+0x134/0x4a0 arch_jump_label_transform+0x64/0x78 __jump_label_update+0x148/0x180 static_key_enable_cpuslocked+0xd0/0x120 static_key_enable+0x30/0x50 check_kvm_guest+0x60/0x88 pSeries_smp_probe+0x54/0xb0 smp_prepare_cpus+0x3e0/0x430 kernel_init_freeable+0x20c/0x43c kernel_init+0x30/0x1a0 ret_from_kernel_thread+0x5c/0x64
Peter pointed out that this is because do_patch_instruction() has disabled interrupts, but then map_patch_area() calls map_kernel_page() then hash__map_kernel_page() which does a sleeping memory allocation.
We only see the warning in KVM guests with SMT enabled, which is not particularly common, or on other platforms if CONFIG_KPROBES is disabled, also not common. The reason we don't see it in most configurations is that another path that happens to have interrupts enabled has allocated the required page tables for us, eg. there's a path in kprobes init that does that. That's just pure luck though.
As Christophe suggested, the simplest solution is to do a dummy map/unmap when we initialise the patching, so that any required page table levels are pre-allocated before the first call to do_patch_instruction(). This works because the unmap doesn't free any page tables that were allocated by the map, it just clears the PTE, leaving the page table levels there for the next map.
Reported-by: Paul Menzel pmenzel@molgen.mpg.de Debugged-by: Peter Zijlstra peterz@infradead.org Suggested-by: Christophe Leroy christophe.leroy@csgroup.eu Signed-off-by: Michael Ellerman mpe@ellerman.id.au Link: https://lore.kernel.org/r/20220223015821.473097-1-mpe@ellerman.id.au Signed-off-by: Sasha Levin sashal@kernel.org --- arch/powerpc/lib/code-patching.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)
diff --git a/arch/powerpc/lib/code-patching.c b/arch/powerpc/lib/code-patching.c index a2e4f864b63d..4318aee65a39 100644 --- a/arch/powerpc/lib/code-patching.c +++ b/arch/powerpc/lib/code-patching.c @@ -43,9 +43,14 @@ int raw_patch_instruction(struct ppc_inst *addr, struct ppc_inst instr) #ifdef CONFIG_STRICT_KERNEL_RWX static DEFINE_PER_CPU(struct vm_struct *, text_poke_area);
+static int map_patch_area(void *addr, unsigned long text_poke_addr); +static void unmap_patch_area(unsigned long addr); + static int text_area_cpu_up(unsigned int cpu) { struct vm_struct *area; + unsigned long addr; + int err;
area = get_vm_area(PAGE_SIZE, VM_ALLOC); if (!area) { @@ -53,6 +58,15 @@ static int text_area_cpu_up(unsigned int cpu) cpu); return -1; } + + // Map/unmap the area to ensure all page tables are pre-allocated + addr = (unsigned long)area->addr; + err = map_patch_area(empty_zero_page, addr); + if (err) + return err; + + unmap_patch_area(addr); + this_cpu_write(text_poke_area, area);
return 0;