On 2/13/25 01:58, Keith Busch wrote:
On Thu, Feb 13, 2025 at 01:31:07AM +0000, Pavel Begunkov wrote:
On 2/11/25 00:56, Keith Busch wrote:
@@ -24,6 +24,9 @@ struct io_rsrc_node {
unsigned long file_ptr;
struct io_mapped_ubuf *buf;
};
+
+ void (*release)(void *);
+ void *priv;
Nodes are more generic than buffers. Unless we want to reuse it
for files, and I very much doubt that, it should move into
io_mapped_ubuf.
Good point. Cloning is another reason it should be in the imu.
If we want to save space, the "KBUFFER" type doesn't need ubuf or
folio_shift, so we could even union the new fields to prevent imu from
getting any bigger. The downside would be we can't use "release" to
distinguish kernel vs user buffers, which you suggested in that patch's
thread. Which option do you prefer?
I think we can just grow the struct for now, it's not a big
problem, and follow up on that separately. Should make it
easier for the series as well.
--
Pavel Begunkov
