Hi, Here's v3 of the ring resizing support. For the v1 posting and details, look here: https://lore.kernel.org/io-uring/20241022021159.820925-1-axboe@xxxxxxxxx/T/#md3a2f049b0527592cc6d8ea25b46bde9fa8e5c68 include/uapi/linux/io_uring.h | 3 + io_uring/io_uring.c | 84 +++++++------ io_uring/io_uring.h | 6 + io_uring/memmap.c | 4 + io_uring/register.c | 214 ++++++++++++++++++++++++++++++++++ 5 files changed, 273 insertions(+), 38 deletions(-) Since v2: - Add patch explicitly checking for valid rings at mmap time. More of a documentation patch than anything, doesn't fix any current or future issues. But it makes it explicit that they have to be valid. - Fix an issue with resizing when the sizes were identical, causing a cleared io_uring_params to be returned and liburing using that to re-mmap the rings. Ensure io_uring_params is filled in before getting copied back, even if nothing was done. - Fix an issue with SQPOLL, it needs to get quiesced before we can proceed with the resize. - Hold the mmap_sem and completion lock across the final part of the operation, where state is copied and the rings swapped. This prevents concurrent IO from accessing rings that are going away, and mmap from seeing any NULL or going-away mappings until the swap has been completed. - Add some more comments. -- Jens Axboe
