Milan Broz <[email protected]>:
> Anyway, my understanding is that all device-mapper targets use mempools,
> which should ensure that they can process even under memory pressure.
I used journal mode so far, but, as well as I understand, direct mode is
okay for my use case.
Okay, I spent some time carefully reading dm-integrity source code.
I have read v6.12.48, because this is kernel I use.
And I conclude that dm-integrity code never allocate (not even from mempool)...
...in main code paths (as opposed to initialization code paths)...
...in direct ('D') mode...
...if I/O doesn't fail and checksums match.
(As I said in previous letter, mempools are bad, too, as well as I understand.)
I found exactly one place, where we seem to allocate in main code path:
https://elixir.bootlin.com/linux/v6.12.48/source/drivers/md/dm-integrity.c#L1789
(i. e. these two kmalloc's).
But I think this okay, because:
- we pass GFP_NOIO, so, as well as I understand, this should not lead to
recursion
- we pass __GFP_NORETRY, so, as well as I understand, we will not block in
this kmalloc for too much time
- we gracefully handle possible failure
Other strange place I found is this:
https://elixir.bootlin.com/linux/v6.12.48/source/drivers/md/dm-integrity.c#L1704
.
But I think this is okay, because:
- integrity_recheck is only ever called from here:
https://elixir.bootlin.com/linux/v6.12.48/source/drivers/md/dm-integrity.c#L1857
- that integrity_recheck call is only ever happens if dm_integrity_rw_tag failed
- as well as I understand, dm_integrity_rw_tag can only fail if we got actual
I/O error or checksum mismatch
So, this mempool_alloc call is okay for my use case.
So: in 'D' mode everything should be okay for my use case.
Another note: I used very stupid way to search functions, which allocate:
if function has "alloc" in its name, then I consider it allocating. :)
And final note: there is an elephant in a room: bufio.
As well as I understand, when pages are swapped in my use case, they first
will get to dm-integrity bufio cache, and only after that, they will
actually hit disk.
This, of course, defeats whole purpose of swap.
And possibly can lead to deadlocks.
Is there a way to disable bufio?
Or maybe bufio is used for checksums and metadata only?
--
Askar Safin
