fsapi: migrates PollRead to Poll with Pflag #1599
Conversation
This migrates PollRead to Poll so that it can support read, write or a combination via flags. This should allow `poll_oneoff` to be implemented even if internally it may need to use a deadline decrementing approach to cycle through each file instead of via multiple at the same time. The benefit of doing things this way is we don't need to top-level a concept of file descriptor, which complicates the design of `fsapi.FS` which is far easier to implement if there are no "multi-file" apis on it. This intentionally doesn't implement the approach at the same time, as the goal is to create the api to implement. Doing so also puts a lot less pressure on the task to complete migration off syscall types. Signed-off-by: Adrian Cole <[email protected]>
| @@ -1,3 +1,5 @@ | |||
| //go:build windows || linux || darwin | |||
There was a problem hiding this comment.
this was missing. we really need to finish #1578 because it is very easy to break it more, with each change to things that can use syscalls.
Signed-off-by: Edoardo Vacchi <[email protected]>
|
fixed the Windows build. Without a lower-level concept of fd/handle, we improve encapsulation/portability but we won't be able to multiplex I/O (e.g. in this specific case, we won't be able to |
|
@evacchi I think by "why we encapsulate FDs" what you are asking is why do we expose a replacement for The primary use case of this abstraction is to be able to write delete, etc files. Currently, we continually break portability and if we don't stop doing that we can never expose it. I'll make notes on this. |
|
Also, it would be helpful when discussing what won't work, is to be very concrete. For example, some existing case (real not forged, like test we are failing) that cannot be solved as a poll loop or as a goroutine externally. We have a tendency to over-solve a lot and by making failing tests it helps conserve energy. |
Signed-off-by: Adrian Cole <[email protected]>
|
@mathetake can you take a look? I had planned to do code and still have time today. I spent the last 3 hours explaining why our abstraction doesn't need to up front interfaces for every niche features, using multi-poll as an example of that. |
| While some use cases are solved with real files, not all are. Regardless, an | ||
| interface approach is necessary to ensure users can intercept I/O operations. | ||
|
|
||
| ### Why doesn't `sys.File` have a `Fd()` method? |
There was a problem hiding this comment.
note: I'm using sys.File as even though we may have forgotten, we did promise to expose our filesystem abstraction in the next release. Even if that ends up experimentalsys.File, the end name would be this.
Signed-off-by: Adrian Cole <[email protected]>
Signed-off-by: Adrian Cole <[email protected]>
This migrates PollRead to Poll so that it can support read, write or a combination via flags. This should allow
poll_oneoffto be implemented even if internally it may need to use a deadline decrementing approach to cycle through each file instead of via multiple at the same time.The benefit of doing things this way is we don't need to top-level a concept of file descriptor, which complicates the design of
fsapi.FSwhich is far easier to implement if there are no "multi-file" apis on it.This intentionally doesn't implement the approach at the same time, as the goal is to create the api to implement. Doing so also puts a lot less pressure on the task to complete migration off syscall types.