Re: [RFC] a question about mlockall() and mprotect()

[Michal Hocko]

On Tue 26-09-17 16:39:56, Xishi Qiu wrote:
> On 2017/9/26 16:17, Michal Hocko wrote:
> 
> > On Tue 26-09-17 15:56:55, Xishi Qiu wrote:
> >> When we call mlockall(), we will add VM_LOCKED to the vma,
> >> if the vma prot is ---p,
> > 
> > not sure what you mean here. apply_mlockall_flags will set the flag on
> > all vmas except for special mappings (mlock_fixup). This phase will
> > cause that memory reclaim will not free already mapped pages in those
> > vmas (see page_check_references and the lazy mlock pages move to
> > unevictable LRUs).
> > 
> >> then mm_populate -> get_user_pages will not alloc memory.
> > 
> > mm_populate all the vmas with pages. Well there are certainly some
> > constrains - e.g. memory cgroup hard limit might be hit and so the
> > faulting might fail.
> > 
> >> I find it said "ignore errors" in mm_populate()
> >> static inline void mm_populate(unsigned long addr, unsigned long len)
> >> {
> >> 	/* Ignore errors */
> >> 	(void) __mm_populate(addr, len, 1);
> >> }
> > 
> > But we do not report the failure because any failure past
> > apply_mlockall_flags would be tricky to handle. We have already dropped
> > the mmap_sem lock so some other address space operations could have
> > interfered.
> >  
> >> And later we call mprotect() to change the prot, then it is
> >> still not alloc memory for the mlocked vma.
> >>
> >> My question is that, shall we alloc memory if the prot changed,
> >> and who(kernel, glibc, user) should alloc the memory?
> > 
> > I do not understand your question but if you are asking how to get pages
> > to map your vmas then touching that area will fault the memory in.
> 
> Hi Michal,
> 
> syscall mlockall() will first apply the VM_LOCKED to the vma, then
> call mm_populate() to map the vmas.
> 
> mm_populate
> 	populate_vma_page_range
> 		__get_user_pages
> 			check_vma_flags
> And the above path maybe return -EFAULT in some case, right?
> 
> If we call mprotect() to change the prot of vma, just let
> check_vma_flags() return 0, then we will get the mlocked pages
> in following page-fault, right?

Any future page fault to the existing vma will result in the mlocked
page. That is what VM_LOCKED guarantess.

> My question is that, shall we map the vmas immediately when
> the prot changed? If we should map it immediately, who(kernel, glibc, user)
> do this step?

This is still very fuzzy. What are you actually trying to achieve?
-- 
Michal Hocko
SUSE Labs