[Kranzes]

Linux 7.3 lets a FUSE-over-io_uring server give each queue one pool of payload buffers instead of registering a buffer per ring entry. Each buffer has the connection's maximum payload size, 1 MiB with our INIT reply. With 16 entries per queue on a 24-CPU machine, that is 384 buffers. The kernel lends a pool buffer to an entry only while it serves a request that has a payload or a reply body, and returns it when the entry commits.

The first commit adds the protocol 7.46 uAPI to abi::fuse_uring: the ADD_QUEUE and ADD_BUFPOOL commands, the FUSE_HAS_IO_URING_BUFPOOL INIT flag, the offset field of the entry header and the bufpool field of the command. FsOptions doesn't know the INIT flag, so the server records it from the raw INIT flags, along with the max_write and max_pages of its reply. Server::uring_payload_size() uses those to compute the kernel's buffer size the way fuse_uring_create() does.

The second commit uses them. When the kernel offers pools and UringConfig::bufs_per_queue isn't 0, UringFuseServing maps one region for all queues and hands each queue its slice with ADD_QUEUE and ADD_BUFPOOL before any entry registers. Entries then register without a payload buffer, and each request's Reader and Writer point at the buffer whose offset the kernel reports. A request without a payload and without out arguments gets no buffer, and its offset of 0 looks the same as the first buffer, so the opcode table now records which requests have out arguments and the others get an empty reply body. Kernels before 7.3 don't offer pools at INIT and keep a buffer per entry. bufs_per_queue defaults to 4, the default number of entries. UringConfig gains a field, so code that builds it with a struct literal needs ..Default::default().

With the benchmark daemon's --uring mode, 4 workers and 16 entries per queue on a 24-CPU machine, random 1 MiB direct reads from one fio job per CPU ran at the same speed with and without pools. Over three alternating runs, RSS dropped from 393 MiB to 106 MiB. The reserved size dropped from 1492 MiB to 431 MiB, but most of that comes from the two scratch areas every entry allocates at the size of its own payload buffer, which pooled entries don't need. docs/fuse-uring-performance.md has the setup and the table, and the benchmark daemon takes --uring-bufs N to compare.

This sits on #280, because pooled entries rely on the worker giving its entries back only after the kernel does. The mount test now runs with 0, 1 and 4 buffers per queue, writing and reading 3 MiB files from four threads, so requests wait for buffers. The test suite passes in a user namespace on Linux 7.3-rc7 on both async runtimes, apart from the 14 passthrough tests that need real root and also fail on master. Zero-copy, the other 7.3 feature, needs CAP_SYS_ADMIN and comes in a later PR.