On Sat, Feb 01, 2020 at 08:27:49PM +0100, Jann Horn wrote: > FWIW, as far as I understand, usercopy doesn't actually have any > effect on drivers that use the modern, proper APIs, since those don't > use the slab allocator at all - as I pointed out in my last mail, the > dma-kmalloc* slabs are used very rarely. (Which is good, because > putting objects from less-than-page-size slabs into iommu entries is a > terrible idea from a security and reliability perspective because it > gives the hardware access to completely unrelated memory.) Instead, > they get pages from the page allocator, and these pages may e.g. be > allocated from the DMA, DMA32 or NORMAL zones depending on the > restrictions imposed by hardware. So I think the usercopy restriction > only affects a few oddball drivers (like this s390 stuff), which is > why you're not seeing more bug reports caused by this. Getting pages from the page allocator is true for dma_alloc_coherent() and friends. But it's not true for streaming DMA mappings (dma_map_*) for which the memory usually comes from kmalloc(). If this is something we want to fix (and I have an awful feeling we're going to regret it if we say "no, we trust the hardware"), we're going to have to come up with a new memory allocation API for these cases. Or bounce bugger the memory for devices we don't trust. The problem with the dma_map_* API is that memory might end up being allocated once and then used multiple times by different drivers. eg if I allocate an NFS packet, it might get sent first to eth0, then (when the route fails) sent to eth1. Similarly in storage, a RAID-5 driver might map the same memory several times to send to different disk controllers.
