newlib-nano ignores setvbuf?

Hi Peter,
This is an intentional behavior to save code size of streamio in newlib. This feature is also upstreamed to newlib and is available with configuration option "--disable-newlib-fvwrite-in-streamio". Below is description about this feature I checked in newlib's README.

+`--disable-newlib-fvwrite-in-streamio'
+ NEWLIB implements the vector buffer mechanism to support stream IO
+ buffering required by C standard. This feature is possibly
+ unnecessary for embedded systems which won't change file buffering
+ with functions like `setbuf' or `setvbuf'. The buffering mechanism
+ still acts as default for STDIN/STDOUT/STDERR even if this option
+ is specified.
+ Enabled by default.

As for how to disable the buffering in nano, I think it can be done by hacking streamio (and maybe easily), though I don't remember where to hack exactly.

OK; I had seen that in the documentation, but didn't interpret it correctly. Unfortunately this behavior makes nano unusable for stdio for me, but if the feature is already upstream maybe I can provide a solution there, or just use http://pabigot.github.io/embtextf/.

Hi,
I think we may be able to introduce another configuration option to enable/disable buffering for STDIO/STDOUT/STDERR (or all streamio file?) when "--disable-newlib-fvwrite-in-streamio" is specified. I think the code size can be further reduced in this case. I think I can look into it later when I revisit newlib-nano.

Thanks; I think that'd be a big help. I did look at the newlib source but was unable to understand what it was doing within the time I was willing to devote to the question.

Hi Peter,
Would you mind describe your use case a little bit more please? Why the buffering streamio doesn't work for you?

Thanks.
bin

Sure. I'm doing a basic BSP in which my use involves a serial console that is primarily used for debugging output. I want printf/putchar to be usable to display progress, e.g. single characters emitted to tell me what point in the execution I've reached. I've implemented _write() to do blocking output to the UART.

Because printf/putchar are by default line-buffered, I use setvbuf() to disable buffering so that the characters appear as the code executes, rather than all at once whenever a newline is finally generated. This works with standard newlib, but takes 25 kiB or so because of the library overhead.

With newlib-nano, the code size drops by about 20 kiB, but the output is buffered internally with no way to turn it off. Where I misunderstood the documentation you added was in its statement that "The buffering mechanism still acts as default for STDIN/STDOUT/STDERR"; this is misleading because it's not a default that can be changed, it's hard-coded to be enabled.

So I can't use newlib-nano because the functionality is missing, and I don't want to use newlib because it's too big.

Hope this helps.

Thanks for pointing it out. I will change that part of words for now, then get back to the buffering/non-buffering issue later.

I discovered that the CMSIS standard gcc-arm.ld script I was using inappropriately (IMO) included a GROUP command that searched in libc.a even when nano (libc_s.a) was being used. Once that command is removed, it appears that setvbuf() does work as it should, and newlib-nano is small enough to meet my needs.