On 2019-07-18, Rasmus Villemoes [off-list ref] wrote:

On 06/07/2019 16.57, Aleksa Sarai wrote:

quoted

--- a/fs/open.c
+++ b/fs/open.c

@@ -928,24 +928,32 @@ struct file *open_with_fake_path(const struct path *path, int flags,
 }
 EXPORT_SYMBOL(open_with_fake_path);
 
-static inline int build_open_flags(int flags, umode_t mode, struct open_flags *op)
+static inline int build_open_flags(struct open_how how, struct open_flags *op)
 {

How does passing such a huge struct by value affect code generation?
Does gcc actually inline the function (and does it even inline the old
one given that it's already non-trivial and has more than one caller).

I'm not sure, but I'll just do what you suggested with passing a const
reference and just copying the few fields that actually are touched by
this function.

quoted

diff --git a/include/linux/fcntl.h b/include/linux/fcntl.h
index 2868ae6c8fc1..e59917292213 100644
--- a/include/linux/fcntl.h
+++ b/include/linux/fcntl.h

@@ -4,13 +4,26 @@
 
 #include <uapi/linux/fcntl.h>
 
-/* list of all valid flags for the open/openat flags argument: */
+/* Should open_how.mode be set for older syscalls wrappers? */
+#define OPENHOW_MODE(flags, mode) \
+	(((flags) | (O_CREAT | __O_TMPFILE)) ? (mode) : 0)
+

Typo: (((flags) & (O_CREAT | __O_TMPFILE)) ? (mode) : 0)

Yup, thanks. I'm not sure why my tests passed on v9 with this bug (they
didn't pass in my v10-draft until I fixed this bug earlier today).

quoted

+/**
+ * Arguments for how openat2(2) should open the target path. If @extra is zero,
+ * then openat2(2) is identical to openat(2).
+ *
+ * @flags: O_* flags (unknown flags ignored).
+ * @mode: O_CREAT file mode (ignored otherwise).

should probably say "O_CREAT/O_TMPFILE file mode".

:+1:

quoted

+ * @upgrade_mask: restrict how the O_PATH may be re-opened (ignored otherwise).
+ * @resolve: RESOLVE_* flags (-EINVAL on unknown flags).
+ * @reserved: reserved for future extensions, must be zeroed.
+ */
+struct open_how {
+	__u32 flags;
+	union {
+		__u16 mode;
+		__u16 upgrade_mask;
+	};
+	__u16 resolve;

So mode and upgrade_mask are naturally u16 aka mode_t. And yes, they
probably never need to be used together, so the union works. That then
makes the next member 2-byte aligned, so using a u16 for the resolve
flags brings us to an 8-byte boundary, and 11 unused flag bits should be
enough for a while. But it seems a bit artificial to cram all this
together and then add 56 bytes of reserved space.

I will happily admit that padding to 64 bytes is probably _very_ extreme
(I picked it purely because it's the size of a cache-line so anything
bigger makes even less sense). I was hoping someone would suggest a
better size once I posted the patchset, since I couldn't think of a good
answer myself.

Do you have any suggestions for a better layout or padding size?

-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
<https://www.cyphar.com/>