On 25 Aug 2021, at 15:59, Peter Xu <peterx@redhat.com<mailto:peterx@redhat.com>> wrote:

On Wed, Aug 25, 2021 at 01:40:13PM +0000, Tiberiu Georgescu wrote:
Hello Peter, sorry for the late reply,

Hi, Tiberiu,

No worries on that.


On Fri, Aug 20, 2021 at 04:49:58PM +0000, Tiberiu Georgescu wrote:
Firstly, I am worried lseek with the SEEK_HOLE flag would page in pages from
swap, so using it would be a direct factor on its own output. If people are working
on Live Migration, this would not be ideal. I am not 100% sure this is how lseek
works, so please feel free to contradict me, but I think it would swap in some
of the pages that it seeks through, if not all, to figure out when to stop. Unless it
leverages the page cache somehow, or an internal bitmap.

It shouldn't.  Man page is clear on that:

     SEEK_DATA
            Adjust the file offset to the next location in the file greater
            than or equal to offset containing data.  If offset points to
            data, then the file offset is set to offset.

Ok, I got to test it out and you are right. lseek does not swap in pages. That is
great news.

Again, I think your requirement is different from CRIU, so I think mincore() is
the right thing for you.


Secondly, mincore() could return some "false positives" for this particular use
case. That is because it returns flag=1 for pages which are still in the swap
cache, so the output becomes ambiguous.

I don't think so; mincore() should return flag=0 if it's either in swap cache
or even got dropped from it.  I think its name/doc also shows that in the fact
that "as long as it's not in RAM, the flag is cleared".  That's why I think
that should indeed be what you're looking for, if swp entry can be ignored.
More below on that.

By saying there are "false positives", I do not mean that the mincore() would
not work as expected, only that its definition is a little more subtle than that. And
that it does not suit our needs entirely by itself.

I tested mincore() compared to the pagemap, and I discovered that there are
more flags set to 1 (not necessarily contiguous) compared to the pages pagemap
was reporting as present. By also looking through the code, I could only conclude
that pages in the swap cache were considered "still in RAM", so were set to 1 as
well. When looking into what the swap cache does, it makes sense.

Please see mincore_page():

/*
 * When tmpfs swaps out a page from a file, any process mapping that
 * file will not get a swp_entry_t in its pte, but rather it is like
 * any other file mapping (ie. marked !present and faulted in with
 * tmpfs's .fault). So swapped out tmpfs mappings are tested here.
 */
page = find_get_incore_page(mapping, index);
if (page) {
present = PageUptodate(page);
put_page(page);
}

I think the "testing" means when swapped out, the page will be NULL. If it's
just the pte got zapped, the page will be returned.  The call stack should look
like:

 find_get_incore_page -> find_get_page -> pagecache_get_page(fgp_flags==0).

I think the fgp_flags guaranteed it, with no FGP_ENTRY.

Did you test mincore() without my patch (as my current patch will indeed cause
more 1's returned than it should)?  My guess is there's something else that
made your test read more 1's with mincore() than pagemap, but I have no solid
idea on that.

I made sure to avoid any of our patches while testing mincore and counted the
flags to make sure. There are more set than they are present. I will look into the
code again and test on a non-rc kernel version, just to be safe.

I still think it makes sense for mincore to consider pages in the swap cache
to be "in RAM". I start seeing how useful it can be to differentiate between
present pages and in-swap-cache pages.


We could use mincore() and pagemap to find the pages in the swap cache.

In short, mincore() is not enough because it does not differentiate between
present pages and swap-cache entries, as both are in RAM, but the latter
is also in swap. It can be used with other tools to get more specific information
though, so it is useful.

Note that my series is as you mentioned missing the changes to support
mincore() (otherwise I'll know the existance of it!).  It'll be trivial to add
that, but let's see whether mincore() will satisfy your need.

We are currently trying to make use of all tools that we have learned of so far
during our discussions (lseek, map_files, even mincore) to get the information
that we need about swap pages. In theory, for many of our use cases, a
combination of 2 or 3 should be enough.

It is a little more convoluted than a simple pagemap call, but it can be more
versatile (using lseek to skip multiple unallocated pages). As to whether the swap
bit (and more) should be eventually added on the pagemap, maybe this topic
makes more sense to continue on the Documentation thread.

[...]

It is possible for the swap device to be network attached and shared, so multiple
hosts would need to understand its content. Then it is no longer internal to one
kernel only.

By being swap-aware, we can skip swapped-out pages during migration (to prevent IO and potential thrashing), and transfer those pages in another way that
is zero-copy.

That sounds reasonable, but I'm not aware of any user-API that exposes swap
entries to userspace, or is there one?

Good question. AFAIK, the swap device can be retrieved by using the swap type,
which is part of the swap entry. During our discussions, I was always assuming
that, if the pagemap entry kept track of the swap offset, it would keep track of the
swap type and, conversely, the swap device as well. Sorry if I haven't made this
assumption clear until now.

So we were relying on the pagemap to expose swap entry information. Seeing it
works for private pages, we thought it made sense to have worked on shared pages as well.

I.e., how do you know which swap device is which?  How do you guarantee the
kernel swp entry information won't change along with time?

I don't think we can guarantee it unless we halt the guest.

Yes, halting the guest helps, though then performance start to matter because
all time consumed in either pagemap or mincore() will be counted in as downtime
of the VM live migration, and it's not "live" at all during this period.

That's true. That is why this "halting" is supposed to only happen once, and
its duration needs to be minimised by pre-copy. Live migration AFAIK is intended
to have a single imperceptible pause at some point in order to converge more
quickly and cleanly. Without this halt, the whole migration procedure could run
indefinitely, or for days instead of hours/minutes.

I'm not sure how it was done with private mappings before, because I thought
that's a pre-requisite knowledge to decide whether we should migrate a page or
not, but I might have missed something.  We can stop vm, sample, start vm, but
it could become hiccups in the guest too, or otherwise contribute to downtime
when src/dst vm switches.

I see your point. Safe to say the vm should never be stopped during pre-copy.
See my comment above.

But QEMU does most
migration work in pre-copy using a best-effort approach anyway.

So, having a way to retrieve temporary, but accurate information about swap
entries (i.e. post-patch pagemap) should be enough to guarantee a smoother
migration process. It is intended to be repeated, unless there is no change
between iterations.

The kernel will allocate swap device index which will be assigned as swp_type,
right?  If there're multiple swap devices, how do you know which swp_entry is
located on which device?  I wanted to look for that info in "swapon -s" but I
didn't.  Or maybe that solution only works if there's only one swap device?

Besides, I also still have a question on the accuracy. If there's no formal way
for userspace to interact with the kernel, I'm wondering how to guarantee the
page will be kept swapped out, and data located on the swap device will always
be the latest?  Because IMHO the kernel can swap in pages as wish, even if it's
not accessed from the userspace.  After all, all these operations should be
transparent to userspace.

One example in my mind is we do have page fault-around enabled for shmem, so
that even if the VM is stopped, its pages can be faulted in if (unluckily, in
this case, though) some page near the swapped out page got faulted in - it
could be some qemu malloc()ed region that (again, unluckily) was allocated to
have a virtual address very close to the mmap()ed guest memories.

I am not sure whether it's a real problem, e.g., even if some page swapped in
during guest halted for some reason, if no one is writting to that page, at
least the data on the page will still be identical to the one located on the
swap device.  However I think that still sounds too tricky, and maybe fragile.

Great example, and good point. Thank you for raising that up. This looks like a
very real problem that we need to take into consideration in our prototypes.

--
Kind regards,

Tibi