On Thu, Dec 14, 2023 at 12:19:43PM +0300, Arseniy Krasnov wrote:
Hello,
DESCRIPTION
This patchset fixes old problem with hungup of both rx/tx sides and adds
test for it. This happens due to non-default SO_RCVLOWAT value and
deferred credit update in virtio/vsock. Link to previous old patchset:
https://lore.kernel.org/netdev/39b2e9fd-601b-189d-39a9-914e5574524c@xxxxxxxxxxxxxx/
Here is what happens step by step:
TEST
INITIAL CONDITIONS
1) Vsock buffer size is 128KB.
2) Maximum packet size is also 64KB as defined in header (yes it is
hardcoded, just to remind about that value).
3) SO_RCVLOWAT is default, e.g. 1 byte.
STEPS
SENDER RECEIVER
1) sends 128KB + 1 byte in a
single buffer. 128KB will
be sent, but for 1 byte
sender will wait for free
space at peer. Sender goes
to sleep.
2) reads 64KB, credit update not sent
3) sets SO_RCVLOWAT to 64KB + 1
4) poll() -> wait forever, there is
only 64KB available to read.
So in step 4) receiver also goes to sleep, waiting for enough data or
connection shutdown message from the sender. Idea to fix it is that rx
kicks tx side to continue transmission (and may be close connection)
when rx changes number of bytes to be woken up (e.g. SO_RCVLOWAT) and
this value is bigger than number of available bytes to read.
I've added small test for this, but not sure as it uses hardcoded value
for maximum packet length, this value is defined in kernel header and
used to control deferred credit update. And as this is not available to
userspace, I can't control test parameters correctly (if one day this
define will be changed - test may become useless).
Head for this patchset is:
https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/commit/?id=9bab51bd662be4c3ebb18a28879981d69f3ef15a
Link to v1:
https://lore.kernel.org/netdev/20231108072004.1045669-1-avkrasnov@xxxxxxxxxxxxxxxxx/
Link to v2:
https://lore.kernel.org/netdev/20231119204922.2251912-1-avkrasnov@xxxxxxxxxxxxxxxxx/
Link to v3:
https://lore.kernel.org/netdev/20231122180510.2297075-1-avkrasnov@xxxxxxxxxxxxxxxxx/
Link to v4:
https://lore.kernel.org/netdev/20231129212519.2938875-1-avkrasnov@xxxxxxxxxxxxxxxxx/
Link to v5:
https://lore.kernel.org/netdev/20231130130840.253733-1-avkrasnov@xxxxxxxxxxxxxxxxx/
Link to v6:
https://lore.kernel.org/netdev/20231205064806.2851305-1-avkrasnov@xxxxxxxxxxxxxxxxx/
Link to v7:
https://lore.kernel.org/netdev/20231206211849.2707151-1-avkrasnov@xxxxxxxxxxxxxxxxx/
Link to v8:
https://lore.kernel.org/netdev/20231211211658.2904268-1-avkrasnov@xxxxxxxxxxxxxxxxx/
Changelog:
v1 -> v2:
* Patchset rebased and tested on new HEAD of net-next (see hash above).
* New patch is added as 0001 - it removes return from SO_RCVLOWAT set
callback in 'af_vsock.c' when transport callback is set - with that
we can set 'sk_rcvlowat' only once in 'af_vsock.c' and in future do
not copy-paste it to every transport. It was discussed in v1.
* See per-patch changelog after ---.
v2 -> v3:
* See changelog after --- in 0003 only (0001 and 0002 still same).
v3 -> v4:
* Patchset rebased and tested on new HEAD of net-next (see hash above).
* See per-patch changelog after ---.
v4 -> v5:
* Change patchset tag 'RFC' -> 'net-next'.
* See per-patch changelog after ---.
v5 -> v6:
* New patch 0003 which sends credit update during reading bytes from
socket.
* See per-patch changelog after ---.
v6 -> v7:
* Patchset rebased and tested on new HEAD of net-next (see hash above).
* See per-patch changelog after ---.
v7 -> v8:
* See per-patch changelog after ---.
v8 -> v9:
* Patchset rebased and tested on new HEAD of net-next (see hash above).
* Add 'Fixes' tag for the current 0002.
* Reorder patches by moving two fixes first.
Arseniy Krasnov (4):
virtio/vsock: fix logic which reduces credit update messages
virtio/vsock: send credit update during setting SO_RCVLOWAT
vsock: update SO_RCVLOWAT setting callback
vsock/test: two tests to check credit update logic
This order will break the bisectability, since now patch 2 will not
build if patch 3 is not applied.
So you need to implement in patch 2 `set_rcvlowat` and in patch 3
updated it to `notify_set_rcvlowat`, otherwise we always need to
backport patch 3 in stable branches, that should be applied before
patch 2.
You have 2 options:
a. move patch 3 before patch 2 without changing the code
b. change patch 2 to use `set_rcvlowat` and updated that code in patch 3
I don't have a strong opinion, but I slightly prefer option a. BTW that
forces us to backport more patches on stable branches, so I'm fine with
option b as well.
That said:
Nacked-by: Stefano Garzarella <sgarzare@xxxxxxxxxx>
