On Tue, Sep 29, 2020 at 5:06 PM Mark Rutland <mark.rutland@xxxxxxx> wrote: > > On Tue, Sep 29, 2020 at 04:51:29PM +0200, Marco Elver wrote: > > On Tue, 29 Sep 2020 at 16:24, Mark Rutland <mark.rutland@xxxxxxx> wrote: > > [...] > > > > > > From other sub-threads it sounds like these addresses are not part of > > > the linear/direct map. Having kmalloc return addresses outside of the > > > linear map is going to break anything that relies on virt<->phys > > > conversions, and is liable to make DMA corrupt memory. There were > > > problems of that sort with VMAP_STACK, and this is why kvmalloc() is > > > separate from kmalloc(). > > > > > > Have you tested with CONFIG_DEBUG_VIRTUAL? I'd expect that to scream. > > > > > > I strongly suspect this isn't going to be safe unless you always use an > > > in-place carevout from the linear map (which could be the linear alias > > > of a static carevout). > > > > That's an excellent point, thank you! Indeed, on arm64, a version with > > naive static-pool screams with CONFIG_DEBUG_VIRTUAL. > > > > We'll try to put together an arm64 version using a carveout as you suggest. > > Great, thanks! > > Just to be clear, the concerns for DMA and virt<->phys conversions also > apply to x86 (the x86 virt<->phys conversion behaviour is more forgiving > in the common case, but still has cases that can go wrong). To clarify, shouldn't kmalloc/kmem_cache allocations used with DMA be allocated with explicit GFP_DMA? If so, how practical would it be to just skip such allocations in KFENCE allocator?
