fuse: virtio_fs: clamp max_pages_limit by the transport DMA mapping size

[benhillis]

Summary

virtio-fs negotiates max_pages (and therefore max_write) purely from the FUSE_INIT reply, without accounting for the largest single DMA mapping the transport can perform. When the device sits behind a bouncing DMA layer — e.g. swiotlb=force, which WSL enables whenever a virtio-fs device is present — the effective per-mapping limit is IO_TLB_SEGSIZE * IO_TLB_SIZE = 256 KiB. Combined with the FUSE default of 256 pages (1 MiB max_write), the guest builds read/readdir/write requests whose buffers cannot be DMA-mapped, failing with -EIO and wedging the request virtqueue.

virtio-blk already avoids this whole class of bug by clamping its transfer size with virtio_max_dma_size() (see drivers/block/virtio_blk.c, which bounds max_segment_size the same way). This change does the same for virtio-fs: it lowers the connection's max_pages_limit to the transport's per-mapping DMA limit, right next to the existing virtqueue-size clamp in virtio_fs_get_tree(). process_init_reply() then caps the negotiated max_pages (and transitively max_write) to a value the guest can always map. When no DMA limit applies, virtio_max_dma_size() returns SIZE_MAX and the clamp is a no-op — zero change for non-bounce-buffered guests.

The change

/* existing: bound by the request virtqueue size */
fc->max_pages_limit = min_t(unsigned int, fc->max_pages_limit,
                            virtqueue_size - FUSE_HEADER_OVERHEAD);

/* new: also bound by the transport's max single DMA mapping size */
fc->max_pages_limit = min_t(unsigned int, fc->max_pages_limit,
                            virtio_max_dma_size(fs->vqs[VQ_REQUEST].vq->vdev)
                                >> PAGE_SHIFT);

Under swiotlb=force this drops max_pages_limit to 64 → negotiated max_write of 256 KiB, which the guest can always bounce.

Testing

Built linux-msft-wsl-6.18.y + this patch with Microsoft/config-wsl (CONFIG_VIRTIO_FS=y); boots as 6.18.33.1-microsoft-standard-WSL2+.

Validated against a virtio-fs server offering the default 1 MiB max_write, with swiotlb=force active. The only variable between runs is the guest kernel:

Guest kernel	DrvFsTests::WSL2VirtioFs GetDents
stock (no clamp)	-EIO, wedged request virtqueue
this patch	PASS

With the clamp, the negotiated max_pages drops to 64 (256 KiB max_write), so readdir/read/write buffers always fit the swiotlb per-mapping limit.