On Sun, Dec 16, 2018 at 05:38:13PM +0800, Hou Tao wrote: > Hi, > > On 2018/12/15 22:38, Matthew Wilcox wrote: > > On Tue, Dec 04, 2018 at 10:08:40AM +0800, Hou Tao wrote: > >> There is no need to disable __GFP_FS in ->readpage: > >> * It's a read-only fs, so there will be no dirty/writeback page and > >> there will be no deadlock against the caller's locked page > >> * It just allocates one page, so compaction will not be invoked > >> * It doesn't take any inode lock, so the reclamation of inode will be fine > >> > >> And no __GFP_FS may lead to hang in __alloc_pages_slowpath() if a > >> squashfs page fault occurs in the context of a memory hogger, because > >> the hogger will not be killed due to the logic in __alloc_pages_may_oom(). > > > > I don't understand your argument here. There's a comment in > > __alloc_pages_may_oom() saying that we _should_ treat GFP_NOFS > > specially, but we currently don't. > I am trying to say that if __GFP_FS is used in pagecache_get_page() when it tries > to allocate a new page for squashfs, that will be no possibility of dead-lock for > squashfs. > > We do treat GFP_NOFS specially in out_of_memory(): > > /* > * The OOM killer does not compensate for IO-less reclaim. > * pagefault_out_of_memory lost its gfp context so we have to > * make sure exclude 0 mask - all other users should have at least > * ___GFP_DIRECT_RECLAIM to get here. > */ > if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS)) > return true; > > So if GFP_FS is used, no task will be killed because we will return from > out_of_memory() prematurely. And that will lead to an infinite loop in > __alloc_pages_slowpath() as we have observed: > > * a squashfs page fault occurred in the context of a memory hogger > * the page used for page fault allocated successfully > * in squashfs_readpage() squashfs will try to allocate other pages > in the same 128KB block, and __GFP_NOFS is used (actually GFP_HIGHUSER_MOVABLE & ~__GFP_FS) > * in __alloc_pages_slowpath() we can not get any pages through reclamation > (because most of memory is used by the current task) and we also can not kill > the current task (due to __GFP_NOFS), and it will loop forever until it's killed. Ah, yes, that makes perfect sense. Thank you for the explanation. I wonder if the correct fix, however, is not to move the check for GFP_NOFS in out_of_memory() down to below the check whether to kill the current task. That would solve your problem, and I don't _think_ it would cause any new ones. Michal, you touched this code last, what do you think?