On 22/03/2018 17:13, Matthew Wilcox wrote: > On Thu, Mar 22, 2018 at 09:06:14AM -0700, Yang Shi wrote: >> On 3/22/18 2:10 AM, Michal Hocko wrote: >>> On Wed 21-03-18 15:36:12, Yang Shi wrote: >>>> On 3/21/18 2:23 PM, Michal Hocko wrote: >>>>> On Wed 21-03-18 10:16:41, Yang Shi wrote: >>>>>> proc_pid_cmdline_read(), it calls access_remote_vm() which need acquire >>>>>> mmap_sem too, so the mmap_sem scalability issue will be hit sooner or later. >>>>> Ohh, absolutely. mmap_sem is unfortunatelly abused and it would be great >>>>> to remove that. munmap should perform much better. How to do that safely >>> The full vma will have to be range locked. So there is nothing small or large. >> >> It sounds not helpful to a single large vma case since just one range lock >> for the vma, it sounds equal to mmap_sem. > > But splitting mmap_sem into pieces is beneficial for this case. Imagine > we have a spinlock / rwlock to protect the rbtree Which is more or less what I'm proposing in the speculative page fault series: https://lkml.org/lkml/2018/3/13/1158 This being said, having a per VMA lock could lead to tricky dead lock case, when merging multiple VMA happens in parallel since multiple VMA will have to be locked at the same time, grabbing those lock in a fine order will be required. > ... / arg_start / arg_end > / ... and then each VMA has a rwsem (or equivalent). access_remote_vm() > would walk the tree and grab the VMA's rwsem for read while reading > out the arguments. The munmap code would have a completely different > VMA write-locked. >
