On 09/13/2016 07:59 AM, Oleg Nesterov wrote: > On 09/12, Michal Hocko wrote: >> > Considering how this all can be tricky and how partial reads can be >> > confusing and even misleading I am really wondering whether we >> > should simply document that only full reads will provide a sensible >> > results. > I agree. I don't even understand why this was considered as a bug. > Obviously, m_stop() which drops mmap_sep should not be called, or > all the threads should be stopped, if you want to trust the result.
There was a mapping at a given address. That mapping did not change, it was not split, its attributes did not change. But, it didn't show up when reading smaps. Folks _actually_ noticed this in a test suite looking for that address range in smaps. IOW, we had goofy kernel behavior, and it broke a reasonable test program. The test program just used fgets() to read into a fixed-length buffer, which is a completely normal thing to do. To get "sensible results", doesn't userspace have to somehow know in advance how many bytes of data a given VMA will generate in smaps output?
