On Thu, Dec 02, 2021 at 11:08:34AM -0800, Kees Cook wrote: > On Thu, Dec 02, 2021 at 06:08:40PM +0200, Leon Romanovsky wrote: > > On Thu, Dec 02, 2021 at 03:29:47PM +0000, Matthew Wilcox wrote: > > > On Thu, Dec 02, 2021 at 05:23:42PM +0200, Leon Romanovsky wrote: > > > > The problem is that this WARN_ON() is triggered by the users. > > > > > > ... or the problem is that you don't do a sanity check between the user > > > and the MM system. I mean, that's what this conversation is about -- > > > is it a bug to be asking for this much memory in the first place? > > > > We do a lot of checks, and in this case, user provided valid input. > > He asked size that doesn't cross his address space. > > https://elixir.bootlin.com/linux/v5.16-rc3/source/drivers/infiniband/core/umem_odp.c#L67 > > > > start = ALIGN_DOWN(umem_odp->umem.address, page_size); > > if (check_add_overflow(umem_odp->umem.address, > > (unsigned long)umem_odp->umem.length, > > &end)) > > return -EOVERFLOW; > > > > There is a feature called ODP (on-demand-paging) which is supported > > in some RDMA NICs. It allows to the user "export" their whole address > > space to the other RDMA node without pinning the pages. And once the > > other node sends data to not-pinned page, the RDMA NIC will prefetch > > it. > > I think we have two cases: > > - limiting kvmalloc allocations to INT_MAX > - issuing a WARN when that limit is exceeded > > The argument for the having the WARN is "that amount should never be > allocated so we want to find the pathological callers". > > But if the actual issue is that >INT_MAX is _acceptable_, then we have > to do away with the entire check, not just the WARN. First we need to get rid from WARN_ON(), which is completely safe thing to do. Removal of the check can be done in second step as it will require audit of whole kvmalloc* path. Thanks > > -- > Kees Cook
