The guest (our Gramine) should use tx_cnt, peer_fwd_cnt and peer_buf_alloc to calculate the number of bytes currently available in the host buffer, and thus currently possible for Gramine to send on a particular connection (particular virtual socket). Previously, we didn't implement this guest-side buffer space management which could lead to app's select/poll/epoll incorrectly returning POLLOUT events.
This commit fixes this by implementing guest-side buffer space management as outlined by the virtio spec.
How to test this PR?
Actually, Redis and Memcached seem to not use this (i.e., don't perform poll(..., POLLOUT)). So I couldn't really test it, only in my head.
Description of the changes
The guest (our Gramine) should use
tx_cnt,peer_fwd_cntandpeer_buf_allocto calculate the number of bytes currently available in the host buffer, and thus currently possible for Gramine to send on a particular connection (particular virtual socket). Previously, we didn't implement this guest-side buffer space management which could lead to app's select/poll/epoll incorrectly returning POLLOUT events.This commit fixes this by implementing guest-side buffer space management as outlined by the virtio spec.
How to test this PR?
Actually, Redis and Memcached seem to not use this (i.e., don't perform
poll(..., POLLOUT)). So I couldn't really test it, only in my head.This change is