On 06/14/2017 06:55 PM, Will Deacon wrote: >> >> May be we should relook at pmd PTE udpate interface. We really need an >> interface that can update pmd entries such that we don't clear it in >> between. IMHO, we can avoid the pmdp_invalidate() completely, if we can >> switch from a pmd PTE entry to a pointer to PTE page (pgtable_t). We also >> need this interface to avoid the madvise race fixed by > > There's a good chance I'm not following your suggestion here, but it's > probably worth me pointing out that swizzling a page table entry from a > block mapping (e.g. a huge page mapped at the PMD level) to a table entry > (e.g. a pointer to a page of PTEs) can lead to all sorts of horrible > problems on ARM, including amalgamation of TLB entries and fatal aborts. AFAIK some AMD x86_64 CPU's had the same problem and generated MCE's, and on Intel there are some restrictions when you can do that. See the large comment in __split_huge_pmd_locked(). > So we really need to go via an invalid entry, with appropriate TLB > invalidation before installing the new entry. > > Will >
