lkml.org 
[lkml]   [2022]   [Jul]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH] maple_tree: Fix sparse reported issues
Date

...

>
> > So the only time I've even seen __vma_adjust() fail is with a fault
> > injector failing mas_preallocate() allocations. If it's safe to not
> > unwind, I'm happy to drop both unwinds but I was concerned in the path
> > of a vma_merge() calling __vma_adjust() and failing out on allocations
> > then OOM recovering, leaving a VMA with a 1/2 merged vma with anon
> > incorrectly set.. which is an even more unlikely scenario.
>
> It's not half-merged, it is correctly set up (just like if a write fault
> had occurred somewhere in that extent before the merge), so no need to
> unwind.
>

I'll drop the incorrect unwinding then.

> ...
>
> > Right, the __split_vma() never adjusts anything but one side of the
> > 'vma' VMA by inserting the 'insert' VMA. This will result in two writes
> > to the tree - but one will exactly fit in an existing range which will
> > be placed without an allocation via the mas_wr_slot_store() function in
> > the maple tree. Exact fits are nice - they are fast.
>
> I'll have to come back and think about this again later on: "Exact fits
> are nice" may answer my concern in the end, but I still have the worry
> that the first store destroys the prealloc, when it might be the second
> store which needs the prealloc.
>
> ...
>
> > > > > Do you have the patch
> > > > > "maple_tree-Fix-stale-data-copy-in-mas_wr_node_store.patch"? It sounds
> > > > > like your issue fits this fix exactly. I was seeing the same issue with
> > > > > gcc 9.3.1 20200408 and this bug doesn't happen for me now. The logs
> > > > > you sent also fit the situation. I went through the same exercise
> > > > > (exorcism?) of debugging the various additions and removals of the VMA
> > > > > only to find the issue in the tree itself. The fix also modified the
> > > > > test code to detect the issue - which was actually hit but not detected
> > > > > in the existing test cases from a live capture of VMA activities. It is
> > > > > difficult to spot in the tree dump as well. I am sure I sent this to
> > > > > Andrew as it is included in v11 and did not show up in his diff, but I
> > > > > cannot find it on lore, perhaps I forgot to CC you? I've attached it
> > > > > here for you in case you missed it.
> > > >
> > > > Thanks! No, I never received that patch, nor can I see it on lore
> > > > or marc.info; but I (still) haven't looked at v11, and don't know
> > > > about Andrew's diff. Anyway, sounds exciting, I'm eager to stop
> > > > writing this mail and get to testing with that in - but please
> > > > let me know whether it's the mas_dead_leaves() or the __vma_adjust()
> > > > mods you attached previously, which you want me to leave out.
>
> The overnight test run ended in an unexpected way, but I believe we can
> count it as a success - a big success for your stale data copy fix.
>
> (If only that fix had got through the mail system on Friday,
> my report on Sunday would have been much more optimistic.)
>
> I said before that I expected the test run to hit the swapops.h
> migration entry !PageLocked BUG, but it did not. It ran for
> nearly 7 hours, and then one of its builds terminated with
>
> {standard input}: Assembler messages:
> {standard input}: Error: open CFI at the end of file;
> missing .cfi_endproc directive
> gcc: fatal error: Killed signal terminated program cc1
> compilation terminated.
>
> which I've never seen before. Usually I'd put something like that down
> to a error in swap, or a TLB flushing error (but I include Nadav's fix
> in my kernel, and wouldn't get very far without it): neither related to
> the maple tree patchset.
>
> But on this occasion, my guess is that it's actually an example of what
> the swapops.h migration entry !PageLocked BUG is trying to alert us to.
>
> Imagine when such a "stale" migration entry is found, but the page it
> points to (now reused for something else) just happens to be PageLocked
> at that instant. Then the BUG won't fire, and we proceed to use the
> page as if it's ours, but it's not. I think that's what happened.
>
> I must get on with the day: more testing, and thinking.


I think this is the same issue seen here:
https://lore.kernel.org/linux-mm/YsQt3IHbJnAhsSWl@casper.infradead.org/

Note that on 20220616, the maple tree was in the next.

I suspect I am doing something wrong in do_brk_munmap(). I am using a
false VMA to munmap a partial vma by setting it up like the part of the
VMA that would have been split, inserted into the tree, then removed and
freed. I must be missing something necessary for this to function
correctly.

Thanks,
Liam



\
 
 \ /
  Last update: 2022-07-18 19:48    [W:0.161 / U:0.184 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site