On Tue, Mar 15, 2016 at 1:22 PM, Andrey Ryabinin <ryabinin.a.a@xxxxxxxxx> wrote: > 2016-03-15 12:27 GMT+03:00 Alexander Potapenko <glider@xxxxxxxxxx>: >> On Mon, Mar 14, 2016 at 5:56 PM, Andrey Ryabinin <ryabinin.a.a@xxxxxxxxx> wrote: >>> 2016-03-14 13:43 GMT+03:00 Alexander Potapenko <glider@xxxxxxxxxx>: >>> >>>> + >>>> + rec = this_cpu_ptr(&depot_recursion); >>>> + /* Don't store the stack if we've been called recursively. */ >>>> + if (unlikely(*rec)) >>>> + goto fast_exit; >>>> + *rec = true; >>> >>> >>> This just can't work. As long as preemption enabled, task could >>> migrate on another cpu anytime. >> Ah, you're right. >> Do you think disabling preemption around memory allocation is an option here? > > It's definitely not an option. Flag on current doesn't have any > disadvantage over per-cpu approach > and it doesn't require preemption safe context. > However, making the allocation in a separate context would be a better > way to eliminate recursion. > i.e. instead of allocating memory depot_save_stack() kicks a work > which allocates memory. For the record, I've removed the vmalloc code and reinstated alloc_pages(), so that there's no more recursion. Making the allocation in a separate worker will remove the recursion, but may complicate the synchronization. I'd refrain from that since we don't have problems with recursion right now. > >>> You could use per-task flag, although it's possible to miss some >>> in-irq stacktraces: >>> >>> depot_save_stack() >>> if (current->stackdeport_recursion) >>> goto fast_exit; >>> current->stackdepot_recursion++ >>> <IRQ> >>> .... >>> depot_save_stack() >>> if (current->stackdeport_recursion) >>> goto fast_exit; >>> >>> >>> >>>> + if (unlikely(!smp_load_acquire(&next_slab_inited))) { >>>> + /* Zero out zone modifiers, as we don't have specific zone >>>> + * requirements. Keep the flags related to allocation in atomic >>>> + * contexts and I/O. >>>> + */ >>>> + alloc_flags &= ~GFP_ZONEMASK; >>>> + alloc_flags &= (GFP_ATOMIC | GFP_KERNEL); >>>> + /* When possible, allocate using vmalloc() to reduce physical >>>> + * address space fragmentation. vmalloc() doesn't work if >>>> + * kmalloc caches haven't been initialized or if it's being >>>> + * called from an interrupt handler. >>>> + */ >>>> + if (kmalloc_caches[KMALLOC_SHIFT_HIGH] && !in_interrupt()) { >>> >>> This is clearly a wrong way to check whether is slab available or not. >> Well, I don't think either vmalloc() or kmalloc() provide any >> interface to check if they are available. >> >>> Besides you need to check >>> vmalloc() for availability, not slab. >> The problem was in kmalloc caches being unavailable, although I can >> imagine other problems could have arose. >> Perhaps we can drill a hole to get the value of vmap_initialized? >>> Given that STAC_ALLOC_ORDER is 2 now, I think it should be fine to use >>> alloc_pages() all the time. >>> Or fix condition, up to you. >> Ok, I'm going to drop vmalloc() for now, we can always implement this later. >> Note that this also removes the necessity to check for recursion. >>>> + prealloc = __vmalloc( >>>> + STACK_ALLOC_SIZE, alloc_flags, PAGE_KERNEL); >>>> + } else { >>>> + page = alloc_pages(alloc_flags, STACK_ALLOC_ORDER); >>>> + if (page) >>>> + prealloc = page_address(page); >>>> + } >>>> + } >>>> + >> >> >> >> -- >> Alexander Potapenko >> Software Engineer >> >> Google Germany GmbH >> Erika-Mann-Straße, 33 >> 80636 München >> >> Geschäftsführer: Matthew Scott Sucherman, Paul Terence Manicle >> Registergericht und -nummer: Hamburg, HRB 86891 >> Sitz der Gesellschaft: Hamburg -- Alexander Potapenko Software Engineer Google Germany GmbH Erika-Mann-Straße, 33 80636 München Geschäftsführer: Matthew Scott Sucherman, Paul Terence Manicle Registergericht und -nummer: Hamburg, HRB 86891 Sitz der Gesellschaft: Hamburg -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href