On Mon, May 20, 2019 at 10:16:57AM -0400, Alan Stern wrote: > What if the allocation requires the kernel to swap some old pages out > to the backing store, but the backing store is on the device that the > driver is managing? The swap can't take place until the current I/O > operation is complete (assuming the driver can handle only one I/O > operation at a time), and the current operation can't complete until > the old pages are swapped out. Result: deadlock. > > Isn't that the whole reason for using GFP_NOIO in the first place? It is, or rather was. As it has been incredibly painful to wire up the gfp_t argument through some callstacks, most notably the vmalloc allocator which is used by a lot of the DMA allocators on non-coherent platforms, we now have the memalloc_noio_save and memalloc_nofs_save functions that mark a thread as not beeing to go into I/O / FS reclaim. So even if you use GFP_KERNEL you will not dip into reclaim with those flags set on the thread.
