Re: [PATCH v3 0/4] Implement IOCTL to get and clear soft dirty PTE

[Danylo Mocherniuk]

On Tue, Oct 11, 2022 at 6:52 AM Andrei Vagin <avagin@xxxxxxxxx> wrote:
>
> On Mon, Oct 03, 2022 at 04:21:22PM +0500, Muhammad Usama Anjum wrote:
> > On 9/28/22 10:24 PM, Andrei Vagin wrote:
> > > On Wed, Sep 21, 2022 at 11:26 AM Muhammad Usama Anjum
> > > <usama.anjum@xxxxxxxxxxxxx> wrote:
> > >>
> > >> Hi,
> > >>
> > >> Thank you for reviewing.
> > >>
> > >> On 9/19/22 7:58 PM, Andrei Vagin wrote:
> > >>>> This ioctl can be used by the CRIU project and other applications which
> > >>>> require soft-dirty PTE bit information. The following operations are
> > >>>> supported in this ioctl:
> > >>>> - Get the pages that are soft-dirty.
> > >>>
> > >>> I think this interface doesn't have to be limited by the soft-dirty
> > >>> bits only. For example, CRIU needs to know whether file, present and swap bits
> > >>> are set or not.
> > >> These operations can be performed by pagemap procfs file. Definitely
> > >> performing them through IOCTL will be faster. But I'm trying to add a
> > >> simple IOCTL by which some specific PTE bit can be read and cleared
> > >> atomically. This IOCTL can be extended to include other bits like file,
> > >> present and swap bits by keeping the interface simple. The following
> > >> mask advice is nice. But if we add that kind of masking, it'll start to
> > >> look like a filter on top of pagemap. My intention is to not duplicate
> > >> the functionality already provided by the pagemap. One may ask, then why
> > >> am I adding "get the soft-dirty pages" functionality? I'm adding it to
> > >> complement the get and clear operation. The "get" and "get and clear"
> > >> operations with special flag (PAGEMAP_SD_NO_REUSED_REGIONS) can give
> > >> results quicker by not splitting the VMAs.
> > >
> > > This simple interface is good only for a limited number of use-cases.
> > > The interface
> > > that I suggest doesn't duplicate more code than this one, but it is much more
> > > universal. It will be a big mess if you add a separate API for each
> > > specific use-case.
> > >
> > >
> > >>> I mean we should be able to specify for what pages we need to get info
> > >>> for. An ioctl argument can have these four fields:
> > >>> * required bits (rmask & mask == mask) - all bits from this mask have to be set.
> > >>> * any of these bits (amask & mask != 0) - any of these bits is set.
> > >>> * exclude masks (emask & mask == 0) = none of these bits are set.
> > >>> * return mask - bits that have to be reported to user.
> > The required mask (rmask) makes sense to me. At the moment, I only know
> > about the practical use case for the required mask. Can you share how
> > can any and exclude masks help for the CRIU?
> >
>
> I looked at should_dump_page in the CRIU code:
> https://github.com/checkpoint-restore/criu/blob/45641ab26d7bb78706a6215fdef8f9133abf8d10/criu/mem.c#L102
>
> When CRIU dumps file private mappings, it needs to get pages that have
> PME_PRESENT or PME_SWAP but don't have PME_FILE.

I would really like to see the mask discussed will be adopted. With it CRIU will
be able to migrate huge sparse VMAs assuming that a single hole is processed in 
O(1) time. 

Use cases for migrating sparse VMAs are binaries sanitized with ASAN, MSAN or
TSAN [1]. All of these sanitizers produce sparse mappings of shadow memory [2].
Being able to migrate such binaries allows to highly reduce the amount of work
needed to identify and fix post-migration crashes, which happen constantly.

>
> > >>>> - Clear the pages which are soft-dirty.
> > >>>> - The optional flag to ignore the VM_SOFTDIRTY and only track per page
> > >>>> soft-dirty PTE bit
> > >>>>
> > >>>> There are two decisions which have been taken about how to get the output
> > >>>> from the syscall.
> > >>>> - Return offsets of the pages from the start in the vec
> > >>>
> > >>> We can conside to return regions that contains pages with the same set
> > >>> of bits.
> > >>>
> > >>> struct page_region {
> > >>>       void *start;
> > >>>       long size;
> > >>>       u64 bitmap;
> > >>> }
> > >>>
> > >>> And ioctl returns arrays of page_region-s. I believe it will be more
> > >>> compact form for many cases.
> > >> Thank you for mentioning this. I'd considered this while development.
> > >> But I gave up and used the simple array to return the offsets of the
> > >> pages as in the problem I'm trying to solve, the dirty pages may be
> > >> present amid non-dirty pages. The range may not be useful in that case.
> > >
> > > This is a good example. If we expect more than two consequent pages
> > > on average, the "region" interface looks more prefered. I don't know your
> > > use-case, but in the case of CRIU, this assumption looks reasonable.

Plus one for page_region data structure. It will make ASAN shadow memory
representation much more compact as well as any other classical use-case. 

[1] https://github.com/google/sanitizers	
[2] https://github.com/google/sanitizers/wiki/AddressSanitizerAlgorithm#64-bit	

Best,
Danylo