Messages in this thread | | | Date | Tue, 28 Apr 2015 20:35:36 +0200 | From | Michal Hocko <> | Subject | Re: Should mmap MAP_LOCKED fail if mm_poppulate fails? |
| |
On Tue 28-04-15 09:57:11, Linus Torvalds wrote: > On Tue, Apr 28, 2015 at 9:43 AM, Michal Hocko <mhocko@suse.cz> wrote: > > > > Hmm, no other thread has the address from the current mmap call except > > for MAP_FIXED (more on that below). > > With things like opportunistic SIGSEGV handlers that map/unmap things > as the user takes faults, that's actually not at all guaranteed. > > Yeah, it's unusual, but I've seen it, with threaded applications where > people play games with user-space memory management, and do "demand > allocation" with mmap() in response to signals.
I am still not sure I see the problem here. Let's say we have a userspace page fault handler which would do mmap(fault_addr, MAP_FIXED), right?
If we had a racy mmap(NULL, MAP_LOCKED) that could have mapped fault_addr by the time handler does its work then this is buggy wrt. to MAP_LOCKED semantic because the fault handler would discard the locked part. This wouldn't lead to a data loss but still makes MAP_LOCKED usage buggy IMO.
If the racing thread did mmap(around_fault_addr, MAP_FIXED|MAP_LOCKED) then it would be broken as well, and even worse I would say, because the original fault could have been discarded and data lost.
I would expect that user fault handlers would be synchronized with other mmap activity otherwise I have hard time to see how this can all have a well defined behavior. Especially when MAP_FIXED is involved.
> Admittedly we already do bad things in mmap(MAP_FIXED) for that case, > since we dropped the vm lock. But at least it shouldn't be any worse > than a thread speculatively touching the pages..
Actually we already allow to mmap(MAP_FIXED) to fail after discarding an existing mmaped area (see mmap_region and e.g. security_vm_enough_memory_mm or other failure cases). -- Michal Hocko SUSE Labs
| |