Rewrite virtio-net IO to be more efficient #530
Rewrite virtio-net IO to be more efficient #530mtjhrc wants to merge 22 commits intocontainers:mainfrom
Conversation
5220656 to
db2af02
Compare
bccf5ac to
3e444c0
Compare
Signed-off-by: Matej Hrica <mhrica@redhat.com>
Use correct platform-specific library directory (lib vs lib64) and library path env variable (DYLD_LIBRARY_PATH vs LD_LIBRARY_PATH). Signed-off-by: Matej Hrica <mhrica@redhat.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
Make TcpTester::run_server also leak the file descriptor for the stream socket. Closing the fd caused the tests to fail on macOS (they randomly worked on Linux I suppose). Signed-off-by: Matej Hrica <mhrica@redhat.com>
Explicitly specify a /tmp directory, this fixes an issue on macOS where the default tmp path that gets used could be very long, causing unix domain socket tests to fail due to path length. Signed-off-by: Matej Hrica <mhrica@redhat.com>
This flag should be used to indicate to libkrun that downstream network backend wants to receive and transmit the virtio-net header along with Ethernet frames. Network backends using this flag can then forward unmodified headers to another VM or build a sensible virtio_net_hdr (e.g. with GSO fields correctly set) such that receiving VM handles GSO'd frames properly. Signed-off-by: Albin Kerouanton <albin.kerouanton@docker.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
Introduce TxQueueConsumer and RxQueueProducer utilities, which allow consuming virtio queues as a bunch of iovec vectors. Notably these utilities are different than the preexisiting descriptor_utilis. The Reader and Writer in descriptor utilis operate on the order of single descriptor chains and don't allow the multiple descriptor chains to be processed at once due to borrowing issues, wheras these the TxQueueConsumer/RxQueueProducer operate on the order of all descriptors of all descriptor chains at once allowing for batch processing of the whole queue at once. Signed-off-by: Matej Hrica <mhrica@redhat.com>
Rewrite the all of the backend (unixstream, unixgram, tap) in terms of the new RxQueueProducer/TxQueueConsumer abstractions. Signed-off-by: Matej Hrica <mhrica@redhat.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
3e444c0 to
398016b
Compare
Restore accidentally deleted guest-agent crate. Signed-off-by: Matej Hrica <mhrica@redhat.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
6e14f63 to
d09264b
Compare
Signed-off-by: Matej Hrica <mhrica@redhat.com>
1621b2c to
567b33d
Compare
Signed-off-by: Matej Hrica <mhrica@redhat.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
Signed-off-by: Matej Hrica <mhrica@redhat.com>
nirs
left a comment
There was a problem hiding this comment.
Exciting change! I did not review everything but I added comments related to testing vmnet-helper.
| config.mac.as_mut_ptr(), | ||
| COMPAT_NET_FEATURES, | ||
| 0, // no VFKIT flag - vmnet-helper uses raw datagrams | ||
| 0, // no offloading - vmnet-helper uses raw ethernet frames |
There was a problem hiding this comment.
you can use offloading if you add vmnet-helper options --enable-tso --enable-check-offload.
https://github.com/nirs/vmnet-helper/blob/455d17220e07e0301206050438e1f02620bf54aa/vmnet/helper.py#L158
|
|
||
| // Increase socket buffer sizes so libkrun's Unixgram backend (which uses | ||
| // the fd path and does NOT set these) can batch frames without drops. | ||
| let buf_size: libc::c_int = 7 * 1024 * 1024; |
There was a problem hiding this comment.
This is not needed, 4m is large enough for recv buffer.
| libc::setsockopt( | ||
| our_fd, | ||
| libc::SOL_SOCKET, | ||
| libc::SO_SNDBUF, |
There was a problem hiding this comment.
Send buffer is not used for datagram socket on macOS. The size should be larger than the maximum packet size. See https://github.com/nirs/vmnet-helper/blob/455d17220e07e0301206050438e1f02620bf54aa/config.h.in#L9
tests/test_cases/src/test_net/mod.rs
Outdated
| #[cfg(feature = "guest")] | ||
| gateway: Some([192, 168, 105, 1]), | ||
| // HACK: hardcoded host LAN IP for testing; guest needs a default | ||
| // route via the vmnet gateway (192.168.105.1) to reach it. |
There was a problem hiding this comment.
No need to hard code the IPs. If you test with a real vm you can use cloud-init to report the vm ip address in the serial log. Hacky works well for testing.
ip is written here:
https://github.com/nirs/vmnet-helper/blob/455d17220e07e0301206050438e1f02620bf54aa/vmnet/cidata.py#L72
and read from serial log here:
https://github.com/nirs/vmnet-helper/blob/455d17220e07e0301206050438e1f02620bf54aa/vmnet/vm.py#L114
This PR
RxQueueProducer,TxQueueConsumer) in an efficient mannerTODO:
Adresses: #385, #405
supersedes: #493