Re: [PATCH] fix madvise vma merging

[Hugh Dickins <hugh@veritas.com>]

On Fri, 5 Aug 2005, Prasanna Meda wrote:
> Hugh Dickins wrote:
> 
> > 2. Correct initial value of prev when starting part way into a vma: as
> >    in sys_mprotect and do_mlock, it needs to be set to vma in this case
> >    (vma_merge handles only that minimum of cases shown in its comments).
> 
> Acknowledge corrections 1 and  3 readily.  Treated vma_merge
> as block box that can handle all cases.  Motivation for pointless
> case 3 is to skip holes and did not notice that has been covered.
> Thanks for corrections.

And thanks for the confirmations.

> Correction 2 is tricky.  Sometimes it merges similar to case 3,
> misses a needed split,  where after the fix  we can get case 4
> merge. If that is what you are saying, we are in agreement.
> Otherwise, can you explain the real problem?

I probably am saying what you are saying there,
but it's hard for me to understand it that way.

Missing out the "start > vma->vm_start" adjustment of prev introduces
additional (but redundant: non-canonical) cases not considered at all
by vma_merge, now entered with a "prev" which is remote and surely
irrelevant to merging.  "misses a needed split", yes, I saw that;
indeed my test ended up taking the "cases 3, 8" path, when, given
the right prev, it should have been handled as a "case 4".

mmap(0x80000000, 0x3000, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED, fd, 0);
madvise(0x80000000, 0x2000, MADV_RANDOM);
madvise(0x80001000, 0x1000, MADV_NORMAL);

ended up (if I'm remembering it aright) as one single VM_RAND_READ vma,
when it should have been a one page VM_RAND_READ vma followed by a two
page normal vma.  Merged into one because of the incorrect prev, and
VM_RAND_READ rather than normal because of the misplaced success label.

Hugh