Re: [PATCH 12/15] io_uring: add support for pre-mapped user IO buffers

On 1/16/19 2:20 PM, Jens Axboe wrote:

On 1/16/19 1:53 PM, Dave Chinner wrote:

quoted

On Wed, Jan 16, 2019 at 10:50:00AM -0700, Jens Axboe wrote:

quoted

If we have fixed user buffers, we can map them into the kernel when we
setup the io_context. That avoids the need to do get_user_pages() for
each and every IO.

.....

quoted

+			return -ENOMEM;
+	} while (atomic_long_cmpxchg(&ctx->user->locked_vm, cur_pages,
+					new_pages) != cur_pages);
+
+	return 0;
+}
+
+static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
+{
+	int i, j;
+
+	if (!ctx->user_bufs)
+		return -EINVAL;
+
+	for (i = 0; i < ctx->sq_entries; i++) {
+		struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
+
+		for (j = 0; j < imu->nr_bvecs; j++) {
+			set_page_dirty_lock(imu->bvec[j].bv_page);
+			put_page(imu->bvec[j].bv_page);
+		}

Hmmm, so we call set_page_dirty() when the gup reference is dropped...

.....

quoted

+static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
+				  unsigned nr_args)
+{

.....

quoted

+		down_write(&current->mm->mmap_sem);
+		pret = get_user_pages_longterm(ubuf, nr_pages, FOLL_WRITE,
+						pages, NULL);
+		up_write(&current->mm->mmap_sem);

Thought so. This has the same problem as RDMA w.r.t. using
file-backed mappings for the user buffer.  It is not synchronised
against truncate, hole punches, async page writeback cleaning the
page, etc, and so can lead to data corruption and/or kernel panics.

It also can't be used with DAX because the above problems are
actually a user-after-free of storage space, not just a dangling
page reference that can be cleaned up after the gup pin is dropped.

Perhaps, at least until we solve the GUP problems w.r.t. file backed
pages and/or add and require file layout leases for these reference,
we should error out if the  user buffer pages are file-backed
mappings?

Thanks for taking a look at this.

I'd be fine with that restriction, especially since it can get relaxed
down the line. Do we have an appropriate API for this? And why isn't
get_user_pages_longterm() that exact API already? Would seem that most
(all?) callers of this API is currently broken then.

I guess for now I can just pass in a vmas array for
get_user_pages_longeterm() and then iterate those and check for
vma->vm_file. If it's set, then we fail the buffer registration.

And then drop the set_page_dirty() on release, we don't need that.

-- 
Jens Axboe