diff options
| author | bptato <nincsnevem662@gmail.com> | 2025-01-08 18:44:48 +0100 |
|---|---|---|
| committer | bptato <nincsnevem662@gmail.com> | 2025-01-08 21:05:44 +0100 |
| commit | b3e57e3726b3654794057119a3c553312d40426b (patch) | |
| tree | a45790656287934f99da2c6c7be7c0f923ba640c /doc | |
| parent | 9a6512c608dc119d74bc3c57029bf3bf60bad891 (diff) | |
| download | chawan-b3e57e3726b3654794057119a3c553312d40426b.tar.gz | |
loader: use per-process control sockets
Previously, each message (load, resume, passFd, etc.) would open a new
connection to loader's UNIX socket, and assumed the loader was dead
when the loader did not respond (ECONNREFUSED).
As it turns out, this model was hopelessly broken: POSIX does not
specify when a UNIX socket can refuse connections, so while it happened
to work on Linux (which just blocks if it can't accept right now),
FreeBSD would randomly refuse connections whenever its listen queue was
full.
Instead, we now take a socketpair() from the loader in addClient, and
pass on one half to the client (the other half stays in loader); this is
the control stream, and all messages from the specific client are sent
through it. If a message includes a new stream (e.g. lcLoad), then it
sends a new socketpair through the control stream.
Aside from not being completely broken (and being marginally more
efficient), this arrangement has several other benefits: it removes the
need for authentication, almost removes the need for sockdir (now only
buffers use it), and will make it easier to add async message processing
in the future.
Diffstat (limited to 'doc')
| -rw-r--r-- | doc/architecture.md | 10 |
1 files changed, 5 insertions, 5 deletions
diff --git a/doc/architecture.md b/doc/architecture.md index 9153a290..e368addb 100644 --- a/doc/architecture.md +++ b/doc/architecture.md @@ -112,11 +112,11 @@ following steps: obviously, it's more efficient this way.) The loader process distinguishes between clients (i.e processes) through -client keys (session cookies). In theory this should help against rogue -clients, though in practice it is still trivial to crash the loader as -a client. It also helps us block further requests from buffers that have -been discarded by the pager, but still haven't found out yet that their -life time has ended. +their control stream (one end of a socketpair created by loader). In +theory this should help against rogue clients, though in practice it is +still trivial to crash the loader as a client. It also helps us block +further requests from buffers that have been discarded by the pager, but +still haven't found out yet that their life time has ended. ### Buffer |