lib: pure: selectors: for Nuttx, change ioselectors to use "select".

[centurysys]

When I first ported NuttX, used "EPOLL" because "select" seemed to leak memory.
However, the number of fd allocated by default in newSelector() is 1024,
which is a lot for a small MCU, so I changed the implementation to select.

This implementation tested with Jester, on Cortex-A5 platform.

[Araq]

Isn't epoll with a limit smaller than 1024 the better solution?

[centurysys]

@Araq
NuttX config has a LIBC_OPEN_MAX entry, which is now returned by getrlimit(RLIMIT_NOFILE) as in other POSIX OS (merged into apache/nuttx master branch today).

apache/nuttx#8555

I guess it depends on the RAM capacity of the board, etc., but it seems to be 255 by default.

config LIBC_OPEN_MAX
        int "OPEN_MAX for this device"
        default 255
        ---help---
                The maximum number of files that a process can have open
                at any time.  Must not be less than _POSIX_OPEN_MAX.

If using this value instead of 1024 with "when defined(nuttx):", it may not interfere even with EPOLL.
BTW, is it recommended to use EPOLL over "select" in the Nim implementation for performance?

[centurysys]

Would the following changes be acceptable?

--- a/lib/pure/ioselects/ioselectors_epoll.nim
+++ b/lib/pure/ioselects/ioselectors_epoll.nim
@@ -76,7 +76,7 @@ proc newSelector*[T](): Selector[T] =
   var maxFD = maxDescriptors()
   doAssert(maxFD > 0)
   # Start with a reasonable size, checkFd() will grow this on demand
-  const numFD = 1024
+  let numFD = min(maxFD, 1024)
 
   var epollFD = epoll_create1(O_CLOEXEC)

[centurysys]

Even for MCUs as shown below, LIBC_OPEN_MAX was set to 255 unless config was explicitly changed.

https://www.st.com/en/microcontrollers-microprocessors/stm32l476rg.html

Up to 128 KB of SRAM including 32 KB with hardware parity check

It may be safer to set it to "select".

[Araq]

BTW, is it recommended to use EPOLL over "select" in the Nim implementation for performance?

Yes, exactly. Assuming that it is more efficient for the OS.

[centurysys]

Yes, exactly. Assuming that it is more efficient for the OS.

I see.
Since NuttX targets range from MCUs with several hundred KiB of RAM to MPUs with several hundred MiB of RAM, it would be good if the internal implementation of select/EPOLL could be switched according to the memory capacity of the operating environment.

[centurysys]

I would like to get some input from the NuttX OS internal implementation perspective, so I will discuss this with the NuttX community.

[centurysys]

@Araq
I received an answer from the perspective of NuttX's internal implementation.

apache/nuttx#8604

They said that a patch was applied last year and that EPOLL is now the fastest.
It would be preferable to start with a smaller number of initial fd arrays and use EPOLL, but what do you think is the most preferable way to change the internals of the EPOLL implementation?

[Araq]

I don't understand the question, tbh. Add a when section to ioselectors_epoll.nim?

[centurysys]

I found this value in NuttX config.

config FS_NEPOLL_DESCRIPTORS
	int "Maximum number of default epoll descriptors for epoll_create1(2)"
	default 8

How about the following changes instead of this pull request?

diff --git a/lib/posix/posix_other_consts.nim b/lib/posix/posix_other_consts.nim
index 506c92158..08069fe9a 100644
--- a/lib/posix/posix_other_consts.nim
+++ b/lib/posix/posix_other_consts.nim
@@ -750,3 +750,6 @@ var SEEK_SET* {.importc: "SEEK_SET", header: "<unistd.h>".}: cint
 var SEEK_CUR* {.importc: "SEEK_CUR", header: "<unistd.h>".}: cint
 var SEEK_END* {.importc: "SEEK_END", header: "<unistd.h>".}: cint
 
+# <nuttx/config.h>
+when defined(nuttx):
+  var NEPOLL_MAX* {.importc: "CONFIG_FS_NEPOLL_DESCRIPTORS", header: "<nuttx/config.h>".}: cint
diff --git a/lib/pure/ioselects/ioselectors_epoll.nim b/lib/pure/ioselects/ioselectors_epoll.nim
index 8526eb8a3..621fa9a9e 100644
--- a/lib/pure/ioselects/ioselectors_epoll.nim
+++ b/lib/pure/ioselects/ioselectors_epoll.nim
@@ -72,11 +72,16 @@ type
   SelectEvent* = ptr SelectEventImpl
 
 proc newSelector*[T](): Selector[T] =
+  proc initialNumFD(): int {.inline.} =
+    when defined(nuttx):
+      result = NEPOLL_MAX
+    else:
+      result = 1024
   # Retrieve the maximum fd count (for current OS) via getrlimit()
   var maxFD = maxDescriptors()
   doAssert(maxFD > 0)
   # Start with a reasonable size, checkFd() will grow this on demand
-  const numFD = 1024
+  let numFD = initialNumFD()
 
   var epollFD = epoll_create1(O_CLOEXEC)
   if epollFD < 0:

I am concerned that the initial fd numbers other than NuttX will change from const to let.

[Araq]

I am concerned that the initial fd numbers other than NuttX will change from const to let.

That should be fine.

[centurysys]

@Araq
Thank you for your very kind review !
I will be submitting a new pull request.