syzbot complaints that sometimes the ordering between ctx->mmap_lock and vma->vm_lock->lock aren't the same, which obviously makes lockdep unhappy. We'd normally expect the ctx->mmap_lock to nest inside the vma lock, but vm_flags_set() -> vma_start_write() from the ->mmap() patch can happen in the opposite order. Move the vm_flags_set() to before ctx->mmap_lock is grabbed. This does potentially leak the VM_DONT_EXPAND set for that vma, and while that could get cleared, doesn't look like it's necessary to do so. Hence just keep it simple. Reported-by: syzbot+96c4c7891428e8c9ac1a@xxxxxxxxxxxxxxxxxxxxxxxxx Link: https://lore.kernel.org/io-uring/67d0bee4.050a0220.14e108.001f.GAE@xxxxxxxxxx/ Signed-off-by: Jens Axboe <axboe@xxxxxxxxx> --- diff --git a/io_uring/memmap.c b/io_uring/memmap.c index 361134544427..d325b6ab6b99 100644 --- a/io_uring/memmap.c +++ b/io_uring/memmap.c @@ -309,7 +309,6 @@ static int io_region_mmap(struct io_ring_ctx *ctx, { unsigned long nr_pages = min(mr->nr_pages, max_pages); - vm_flags_set(vma, VM_DONTEXPAND); return vm_insert_pages(vma, vma->vm_start, mr->pages, &nr_pages); } @@ -322,6 +321,8 @@ __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) struct io_mapped_region *region; void *ptr; + vm_flags_set(vma, VM_DONTEXPAND); + guard(mutex)(&ctx->mmap_lock); ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); -- Jens Axboe
