On Thu, May 07, 2020 at 05:25:55PM -0700, Sean Christopherson wrote: > Ah, fudge. shmem_zero_setup() triggers shmem_acct_size() and thus > __vm_enough_memory(). Which I should have rememered because I've stared > at that code several times when dealing with the enclave's backing store. > I wasn't seeing the issue because I happened to use MAP_PRIVATE. > > So, bad analysis, good conclusion, i.e. the kernel is still doing the > right thing, it's just not ideal for userspace. > > > Jarkko, we should update the docs and selftest to recommend and use > > PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS > > or > > PROT_NONE, MAP_SHARED | MAP_NORESERVE | MAP_ANONYMOUS" > > when carving out ELRANGE, with an explicit comment that all the normal > rules for mapping memory still apply. Ugh, had forgotten this. OK, I guess this comment explains it all: " /* * shmem_file_setup pre-accounts the whole fixed size of a VM object, * for shared memory and for shared anonymous (/dev/zero) mappings * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1), * consistent with the pre-accounting of private mappings ... */ static inline int shmem_acct_size(unsigned long flags, loff_t size) " /Jarkko
