Skip to contentSkip to conversation

Add IORING_FEAT_FILES_SKIP feature flag

PATCH In review Drew DeVault 3 participants 5 messages May 18, 2021 – May 20, 2021 io_uring

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Add IORING_FEAT_FILES_SKIP feature flag

This is aliased to IORING_FEAT_NATIVE_WORKERS, which was shipped in the same kernel release. A separate flag is useful because the features are unrelated, so user code testing for IORING_FEAT_NATIVE_WORKERS before using FILES_SKIP is not very obvious in intent.

---
 man/io_uring_register.2         | 10 ++++++++--
 src/include/liburing/io_uring.h |  1 +
 2 files changed, 9 insertions(+), 2 deletions(-)
man/io_uring_register.2 +8 -2
index 5326a87..c71ce40 100644
@@ -156,8 +156,14 @@ since 5.5.
 File descriptors can be skipped if they are set to
 .B IORING_REGISTER_FILES_SKIP.
 Skipping an fd will not touch the file associated with the previous
-fd at that index. Available since 5.12.
-
+fd at that index. Available since 5.12. Availability of this feature is
+indicated by the presence of the
+.B IORING_FEAT_FILES_SKIP
+bit in the
+.I features
+field of the
+.I io_uring_params
+structure.
 
 .TP
 .B IORING_UNREGISTER_FILES
src/include/liburing/io_uring.h +1 -0
index 5a3cb90..091dcf7 100644
@@ -285,6 +285,7 @@ struct io_uring_params {
 #define IORING_FEAT_SQPOLL_NONFIXED	(1U << 7)
 #define IORING_FEAT_EXT_ARG		(1U << 8)
 #define IORING_FEAT_NATIVE_WORKERS	(1U << 9)
+#define IORING_FEAT_FILES_SKIP		IORING_FEAT_NATIVE_WORKERS
 
 /*
  * io_uring_register(2) opcodes and arguments
Signature · 1 line
2.31.1
maintainer
#1 lore raw
Quoted 37 lines · Drew DeVault

This is aliased to IORING_FEAT_NATIVE_WORKERS, which was shipped in the same kernel release. A separate flag is useful because the features are unrelated, so user code testing for IORING_FEAT_NATIVE_WORKERS before using FILES_SKIP is not very obvious in intent.

---
 man/io_uring_register.2         | 10 ++++++++--
 src/include/liburing/io_uring.h |  1 +
 2 files changed, 9 insertions(+), 2 deletions(-)
man/io_uring_register.2 +8 -2
index 5326a87..c71ce40 100644
@@ -156,8 +156,14 @@ since 5.5.
 File descriptors can be skipped if they are set to
 .B IORING_REGISTER_FILES_SKIP.
 Skipping an fd will not touch the file associated with the previous
-fd at that index. Available since 5.12.
-
+fd at that index. Available since 5.12. Availability of this feature is
+indicated by the presence of the
+.B IORING_FEAT_FILES_SKIP
+bit in the
+.I features
+field of the
+.I io_uring_params
+structure.
 
 .TP
 .B IORING_UNREGISTER_FILES
src/include/liburing/io_uring.h +0 -0
index 5a3cb90..091dcf7 100644

src/include/liburing/io_uring.h struct io_uring_params { · line 288

Quoted 4 lines
src/include/liburing/io_uring.h +1 -0
@@ -285,6 +285,7 @@ struct io_uring_params {
 #define IORING_FEAT_SQPOLL_NONFIXED	(1U << 7)
 #define IORING_FEAT_EXT_ARG		(1U << 8)
 #define IORING_FEAT_NATIVE_WORKERS	(1U << 9)
+#define IORING_FEAT_FILES_SKIP		IORING_FEAT_NATIVE_WORKERS

I don't think this is a great idea. It can be used as a "probably we have this feature" in userspace, but I don't like aliasing on the kernel side.

Signature · 1 line
Jens Axboe
#2 lore raw
Quoted 5 lines · Jens Axboe
Quoted 5 lines
src/include/liburing/io_uring.h +0 -0
index 5a3cb90..091dcf7 100644
@@ -285,6 +285,7 @@ struct io_uring_params {
Quoted 8 lines

src/include/liburing/io_uring.h struct io_uring_params { · line 288

Quoted 4 lines
src/include/liburing/io_uring.h +1 -0
 #define IORING_FEAT_SQPOLL_NONFIXED	(1U << 7)
 #define IORING_FEAT_EXT_ARG		(1U << 8)
 #define IORING_FEAT_NATIVE_WORKERS	(1U << 9)
+#define IORING_FEAT_FILES_SKIP		IORING_FEAT_NATIVE_WORKERS

I don't think this is a great idea. It can be used as a "probably we have this feature" in userspace, but I don't like aliasing on the kernel side.

This patch is for liburing, following the feedback on the kernel patch (which didn't alias, but regardless).

#3 lore raw
Quoted 15 lines · Drew DeVault
Quoted 5 lines · Jens Axboe
Quoted 5 lines
src/include/liburing/io_uring.h +0 -0
index 5a3cb90..091dcf7 100644
@@ -285,6 +285,7 @@ struct io_uring_params {
Quoted 8 lines

src/include/liburing/io_uring.h struct io_uring_params { · line 288

Quoted 4 lines
src/include/liburing/io_uring.h +1 -0
 #define IORING_FEAT_SQPOLL_NONFIXED	(1U << 7)
 #define IORING_FEAT_EXT_ARG		(1U << 8)
 #define IORING_FEAT_NATIVE_WORKERS	(1U << 9)
+#define IORING_FEAT_FILES_SKIP		IORING_FEAT_NATIVE_WORKERS

I don't think this is a great idea. It can be used as a "probably we have this feature" in userspace, but I don't like aliasing on the kernel side.

Quoted 2 lines

This patch is for liburing, following the feedback on the kernel patch (which didn't alias, but regardless).

This file is a copy (almost) of the kernel's uapi header, so better be off this file and have naming that wouldn't alias with names in this header.

I think that's the problem Jens mean. Jens, is it? And I do still believe it's a better way to not have an itchy one release gap between actual feature introduction and new feat flag.

Signature · 1 line
Pavel Begunkov
#4 lore raw
Quoted 19 lines · Pavel Begunkov
Quoted 15 lines · Drew DeVault
Quoted 5 lines · Jens Axboe
Quoted 5 lines
src/include/liburing/io_uring.h +0 -0
index 5a3cb90..091dcf7 100644
@@ -285,6 +285,7 @@ struct io_uring_params {
Quoted 8 lines

src/include/liburing/io_uring.h struct io_uring_params { · line 288

Quoted 4 lines
src/include/liburing/io_uring.h +1 -0
 #define IORING_FEAT_SQPOLL_NONFIXED	(1U << 7)
 #define IORING_FEAT_EXT_ARG		(1U << 8)
 #define IORING_FEAT_NATIVE_WORKERS	(1U << 9)
+#define IORING_FEAT_FILES_SKIP		IORING_FEAT_NATIVE_WORKERS

I don't think this is a great idea. It can be used as a "probably we have this feature" in userspace, but I don't like aliasing on the kernel side.

Quoted 2 lines

This patch is for liburing, following the feedback on the kernel patch (which didn't alias, but regardless).

Quoted 7 lines

This file is a copy (almost) of the kernel's uapi header, so better be off this file and have naming that wouldn't alias with names in this header.

I think that's the problem Jens mean. Jens, is it? And I do still believe it's a better way to not have an itchy one release gap between actual feature introduction and new feat flag.

It does itch, but I don't think it's actually wrong, per-se. Unless we expect features to be removed or optionally available, it might make sense to have used an incrementing number rather than a bitfield.

What should userspace programs do if they're not sure if they can use FILES_SKIP? What was the previous behavior if -2 is used? I'm guessing it was EINVAL, EBADF, something like that.