Now that io_uring will be in the upcoming 5.1 kernel, I spent some time writing up how it came to be and how it works. More formats likely coming, and I expect a few revisions too. For now, consider this v1.0 of the doc: http://kernel.dk/io_uring.pdf
-
-
Fixed the various typos and missing '*' that were reported, and added a small blurb on io_uring_get_completion() as well, and pushed it out. Same link. Bottom line will show versioning, this one should read 0.2.
Prikaži ovu nitHvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
It'd be cool if you could include a bit more about f[data]sync (synchronicity, guarantees). Lots of users of the facility are going to need it... Plan to play around with optionally using this in postgres in the next few months.
-
Agree, that would be useful. One thing I've been pondering is a barrier type command. Right now you can queue an fsync after a bunch of writes, but ordering is not guaranteed. It'd be nifty to have explicit ordering be possible for that, fsync not issued before pending writes.
- Još 7 drugih odgovora
Novi razgovor -
-
-
I'd like to use it for gpu drivers.
-
Should be nothing stopping us from doing that, basically just a "simple" matter of defining the appropriate command types for that. Definitely open to pursuing other types of commands for this, I've got a few in mind as well (like KV store).
- Još 4 druga odgovora
Novi razgovor -
-
-
Aren't 4096 entries a little low, if you take socket polls into account? Large servers can be expected to have tens of thousands of connections. Sure they'll be running multiple rings, but still it's too low IMO.
-
It might be, but keep in mind that sqe entries are only used for submission. The real pending count would be twice that, the cqe ring size. If 8192 isn't enough, even over multiple rings, bumping this (somewhat) artificial limit is trivial.
- Još 2 druga odgovora
Novi razgovor -
-
-
I wish it included IORING_OP_POLL (for sockets) so we could have a single API to wait for disk and network events (like IOCB_CMD_POLL)
-
The code does include that, it's just not described in the PDF. But none of the op-codes are, really, I'm deferring to the man pages for that kind of detail. It works almost the same as aio for that, except we have POLL_ADD and POLL_REMOVE, instead of CMD_POLL and then cancel.
- Još 3 druga odgovora
Novi razgovor -
-
-
page 5: "For the submission side, the roles are reserved." Presumably "reversed"?
-
Thanks, will fix!
- Još 5 drugih odgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.