Skip usage of EIO_O_FSYNC if undefined

[enumag]

See https://pecl.php.net/package-changelog.php?package=eio&release=2.0.2 and reactphp/filesystem#34

[enumag]

Honestly I don't understand why Travis didn't catch this...

[kelunik]

Because Travis uses glibc instead of musl. I think we should conditionally use that flag then.

[enumag]

Oh so the constant still does exist with glibc? Should I add a ternary operator to use it if it exists? I have no idea what it's for...

[trowski]

As I understood, the flag requires writes to be successfully written to disk before returning (synchronous). In context of this library, since the writing is done in a separate thread, we can know the data was actually written to disk when the promise succeeds.

Since we aren't able to use the flag in other drivers (or apparently some platforms), I think let's just remove it.

[kelunik]

In context of this library, since the writing is done in a separate thread, we can know the data was actually written to disk when the promise succeeds.

No, we can't know that without FSYNC, we just know that the OS wrote the data to its internal buffers, but in case of a power failure the data might still be lost, see https://linux.die.net/man/2/fsync.

Please change it to a conditional, yes.

[kelunik]

@enumag See rosmanov/pecl-eio@65e8b54#diff-0cc087e71ac0c866a224fb4846b3314dR932

[trowski]

No, we can't know that without FSYNC, we just know that the OS wrote the data to its internal buffers, but in case of a power failure the data might still be lost, see https://linux.die.net/man/2/fsync.

Reading that, it seems like what I said is true. Am I not understanding something?

[kelunik]

@trowski I understood you said we don't need the flag, because we do writes in another thread?

[trowski]

No, I phrased it poorly I guess. My point was that we use the flag for synchronous writes to ensure the data is written, but the synchronous write happens in another thread, so it's non-blocking.

However, only EIO exposes this flag, so behavior is inconsistent between drivers and apparently between platforms. I suggested removing the flag for consistency. I'm also fine with using it conditionally if we don't care about that.

[kelunik]

We could add fflush to BlockingDriver as well. Should we?

[kelunik]

Thanks!