virtiofs: negotiate FUSE_MAX_PAGES for larger I/O requests

[benhillis]

Add FUSE_MAX_PAGES to DEFAULT_FLAGS so the FUSE server advertises support for larger request sizes during FUSE_INIT negotiation.

Background

Without this flag, the Linux kernel uses FUSE_DEFAULT_MAX_PAGES_PER_REQ (32 pages = 128 KiB) as the maximum request size. With the flag, the kernel uses min(fuse_max_pages_limit, virtqueue_size - FUSE_HEADER_OVERHEAD) which — for our default virtqueue size of 256 — is 252 pages (~1008 KiB). This is an 8x increase in max request size, significantly reducing per-request overhead for sequential I/O.

Benchmark results

burette virtio-fs benchmark, 3 iterations each, comparing with and without FUSE_MAX_PAGES flag.

6.1 kernel

Metric	Baseline	With flag	Delta
Sequential read	166 MiB/s	227 MiB/s	+37%
Sequential write	101 MiB/s	161 MiB/s	+59%
Random read IOPS	2231	1953	~noise
Random write IOPS	2312	2376	~noise

6.12 kernel

Metric	Baseline	With flag	Delta
Sequential read	119 MiB/s	219 MiB/s	+83%
Sequential write	93 MiB/s	156 MiB/s	+68%
Random read IOPS	1757	1807	~noise
Random write IOPS	2416	2376	~noise

Random I/O is unaffected as expected — 4K blocks do not benefit from larger request sizes.

Notes

DEFAULT_MAX_PAGES (256) is intentionally left unchanged. The kernel clamps max_pages to min(fuse_max_pages_limit=256, virtqueue_size-4=252), so larger values have no effect with the current virtqueue size of 256. The gain comes entirely from negotiating the flag, which moves the kernel from 32 to 252 pages per request.

On Linux 6.18+, fuse_max_pages_limit becomes a runtime module parameter (previously a compile-time constant). To benefit from values beyond 252 in the future, we would also need to increase the virtqueue size.

[Copilot]

Pull request overview

Updates the FUSE session’s default init negotiation flags so virtio-fs can negotiate larger per-request I/O sizes with Linux guests by advertising FUSE_MAX_PAGES during FUSE_INIT.

Changes:

• Add FUSE_MAX_PAGES to the DEFAULT_FLAGS bitmask used for FUSE init negotiation.

[ptrivedi]

LGTM! why the big difference between 6.1 and 6.12? something else likely changed in the subsystem for newer kernels making this flag more effective there?

[benhillis]

Tested the max-pages PR on Windows with the HDV device against both the baseline and PR builds. The
PR shows broad I/O improvements, sequential throughput +25%, random IOPS +20 - 40%, and metadata
ops about 20%.