Le 14/02/2022 à 10:35, Peter Zijlstra a écrit : > On Sun, Feb 13, 2022 at 12:05:50AM +0100, Paul Menzel wrote: > >> [ 0.012154][ T1] BUG: sleeping function called from invalid context at > >> [ 0.022443][ T1] [c0000000084837d0] [c000000000961aac] > dump_stack_lvl+0xa0/0xec (unreliable) >> [ 0.023356][ T1] [c000000008483820] [c00000000019b314] > __might_resched+0x2f4/0x310 >> [ 0.024174][ T1] [c0000000084838b0] [c0000000004c0c70] > kmem_cache_alloc+0x220/0x4b0 >> [ 0.025000][ T1] [c000000008483920] [c000000000448af4] > __pud_alloc+0x74/0x1d0 >> [ 0.025772][ T1] [c000000008483970] [c00000000008fe3c] > hash__map_kernel_page+0x2cc/0x390 >> [ 0.026643][ T1] [c000000008483a20] [c0000000000a9944] > do_patch_instruction+0x134/0x4a0 > > do_patch_instruction() rightfully disables IRQs, but then it goes and > tries a memory alloc, which seems a bit daft. > > I'm thinking Christophe might know more... he's recently been poking at > Power text poking.. I don't know all details about PPC64, but here it seems like hash__map_kernel_page() allocates intermediate page directories when there is not one yet. So if that's the first time hash__map_kernel_page() is called for the text_poke_addr() it allocated PUD and PMD if necessary. As it is kernel memory, once PUD and PMD are allocated they will remain forever I think. So maybe the only thing to do it to perform a dummy mapping/unmapping in text_area_cpu_up() to ensure the page directory are created in advance. Christophe
