Thread (61 messages) 61 messages, 8 authors, 2020-07-21

Re: [PATCH v3 4/4] io_uring: add support for zone-append

From: Kanchan Joshi <hidden>
Date: 2020-07-20 16:50:30
Also in: io-uring, linux-fsdevel, lkml

On Fri, Jul 10, 2020 at 7:41 PM Kanchan Joshi [off-list ref] wrote:
On Fri, Jul 10, 2020 at 7:21 PM Matthew Wilcox [off-list ref] wrote:
quoted
On Fri, Jul 10, 2020 at 02:49:32PM +0100, Christoph Hellwig wrote:
quoted
On Fri, Jul 10, 2020 at 02:48:24PM +0100, Matthew Wilcox wrote:
quoted
If we're going to go the route of changing the CQE, how about:

 struct io_uring_cqe {
         __u64   user_data;      /* sqe->data submission passed back */
-        __s32   res;            /* result code for this event */
-        __u32   flags;
+   union {
+           struct {
+                   __s32   res;            /* result code for this event */
+                   __u32   flags;
+           };
+           __s64   res64;
+   };
 };

then we don't need to change the CQE size and it just depends on the SQE
whether the CQE for it uses res+flags or res64.
How do you return a status code or short write when you just have
a u64 that is needed for the offset?
it's an s64 not a u64 so you can return a negative errno.  i didn't
think we allowed short writes for objects-which-have-a-pos.
If we are doing this for zone-append (and not general cases), "__s64
res64" should work -.
64 bits = 1 (sign) + 23 (bytes-copied: cqe->res) + 40
(written-location: chunk_sector bytes limit)
And this is for the scheme when single CQE is used with bits
refactoring into "_s64 res64" instead of res/flags.

41 bits for zone-append completion = in bytes, sufficient to cover
chunk_sectors size zone
1+22 bits for zone-append bytes-copied = can cover 4MB bytes copied
(single I/O is capped at 4MB in NVMe)

+ * zone-append specific flags
+#define APPEND_OFFSET_BITS     (41)
+#define APPEND_RES_BITS                (23)
+
+/*
  * IO completion data structure (Completion Queue Entry)
  */
 struct io_uring_cqe {
-       __u64   user_data;      /* sqe->data submission passed back */
-       __s32   res;            /* result code for this event */
-       __u32   flags;
+       __u64   user_data;      /* sqe->data submission passed back */
+        union {
+                struct {
+                        __s32   res;            /* result code for
this event */
+                        __u32   flags;
+                };
+               /* Alternate for zone-append */
+               struct {
+                       union {
+                               /*
+                                * kernel uses this to store append result
+                                * Most significant 23 bits to return number of
+                                * bytes or error, and least significant 41 bits
+                                * to return zone-relative offset in bytes
+                                * */
+                               __s64 res64;
+                               /*for user-space ease, kernel does not use*/
+                               struct {
+#if defined(__LITTLE_ENDIAN_BITFIELD)
+                                       __u64 append_offset :
APPEND_OFFSET_BITS;
+                                       __s32 append_res : APPEND_RES_BITS;
+#elif defined(__BIG_ENDIAN_BITFIELD)
+                                       __s32 append_res : APPEND_RES_BITS;
+                                       __u64 append_offset :
APPEND_OFFSET_BITS;
+#endif
+                               }__attribute__ ((__packed__));
+                       };
+                };
+        };
 };

-- 
Joshi
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help