On 9/18/24 10:32 PM, Linus Torvalds wrote:
> On Thu, 19 Sept 2024 at 05:38, Jens Axboe <axboe@xxxxxxxxx> wrote:
>>
>> I kicked off a quick run with this on 6.9 with my debug patch as well,
>> and it still fails for me... I'll double check everything is sane. For
>> reference, below is the 6.9 filemap patch.
Confirmed with a few more runs, still hits, basically as quickly as it
did before. So no real change observed with the added xas_reset().
> Ok, that's interesting. So it's *not* just about "that code didn't do
> xas_reset() after xas_split_alloc()".
>
> Now, another thing that commit 6758c1128ceb ("mm/filemap: optimize
> filemap folio adding") does is that it now *only* calls xa_get_order()
> under the xa lock, and then it verifies it against the
> xas_split_alloc() that it did earlier.
>
> The old code did "xas_split_alloc()" with one order (all outside the
> lock), and then re-did the xas_get_order() lookup inside the lock. But
> if it changed in between, it ended up doing the "xas_split()" with the
> new order, even though "xas_split_alloc()" was done with the *old*
> order.
>
> That seems dangerous, and maybe the lack of xas_reset() was never the
> *major* issue?
>
> Willy? You know this code much better than I do. Maybe we should just
> back-port 6758c1128ceb in its entirety.
>
> Regardless, I'd want to make sure that we really understand the root
> cause. Because it certainly looks like *just* the lack of xas_reset()
> wasn't it.
Just for sanity's sake, I backported 6758c1128ceb (and the associated
xarray xas_get_order() change) to 6.9 and kicked that off.
--
Jens Axboe
