Libuv: UV_FS_O_DIRECT and N-API addon in C

[jorangreef]

• Version: 1.26.0
• Platform: Linux 4.15.0-54-generic #58-Ubuntu SMP Mon Jun 24 10:55:24 UTC 2019 x86_64 x86_64 x86_64 GNU/Linux

When upgrading a native Node addon from C++ and NAN to C and N-API, I noticed that libuv's UV_FS_O_DIRECT changes from 16384 to 0.

Is there something that needs to be fixed so that C source files including #include <uv.h> have UV_FS_O_DIRECT initialized correctly? Something like the following in libuv's unix.h?

#define _GNU_SOURCE
#include <fcntl.h>

[saghul]

Looks like you need to define _GNU_SOURCE indeed:

The O_DIRECT, O_NOATIME, O_PATH, and O_TMPFILE flags are Linux-
       specific.  One must define _GNU_SOURCE to obtain their definitions.

From: http://man7.org/linux/man-pages/man2/open.2.html

But I wonder why it isn't already working though. Calling in the expert :-) /cc @bnoordhuis

[bnoordhuis]

That's a bug in our headers, you shouldn't have to define _GNU_SOURCE (or _DEFAULT_SOURCE) to get the proper definition.

(And the headers shouldn't define it themselves because that's unhygienic, it affects the libuv consumer.)

Unfortunately the value of O_DIRECT is architecture-dependent so we'll probably end up with code that looks like this:

#if defined(O_DIRECT)
# define UV_FS_O_DIRECT O_DIRECT
#elif defined(__linux__) && defined(__mips__)
# define UV_FS_O_DIRECT 0x8000
#elif defined(__linux__) && defined(__powerpc__)
# define UV_FS_O_DIRECT 0x4000
// etc.
#else
# define UV_FS_O_DIRECT 0
#endif

[jorangreef]

Thanks @saghul and @bnoordhuis

I had a bad feeling about _GNU_SOURCE.

[jorangreef]

@bnoordhuis , would you like me to do a PR exactly as you have it or would you prefer to do it if we need more platforms?

[bnoordhuis]

Linux is the only problematic platform, I think, so it's just a matter of filling out the list.

After thinking about it some more, it's probably best to put the linux+arch combos before the #if defined(O_DIRECT) stanza.

[jorangreef]

Linux is the only problematic platform, I think, so it's just a matter of filling out the list.

If you can give me the values for the various Linux architectures, I don't mind filling out the list.

[saghul]

https://github.com/ifduyue/musl/search?q=O_DIRECT&unscoped_q=O_DIRECT

[jorangreef]

After thinking about it some more, it's probably best to put the linux+arch combos before the #i defined(O_DIRECT) stanza.

@bnoordhuis, just so I share your reasoning, may I ask why? Is it to ensure that C++ and C addons always define O_DIRECT in exactly the same way?

[bnoordhuis]

@jorangreef That's right. Better unconditionally wrong than conditionally right - the former gets detected quicker.