Re: [PATCH 1/2] tools/nolibc: Add fread() to stdio.h

[Willy Tarreau <w@1wt.eu>]

On Tue, Jan 06, 2026 at 12:02:18PM +0100, Thomas Weißschuh wrote:
> On 2026-01-05 11:01:42+0000, David Laight wrote:
> > On Mon, 5 Jan 2026 18:43:03 +0900
> > Daniel Palmer <daniel@thingy.jp> wrote:
> > > On Mon, 5 Jan 2026 at 18:27, David Laight <david.laight.linux@gmail.com> wrote:
> > > > But you've deleted the partial bytes from the input.
> > > > I'm sure that isn't right.
> > > > Normally a FILE is buffered and the bytes are saved for the next read.
> > > > Remember you can be reading from a pipe that is being written using
> > > > 'block buffering' - so it is valid for only a partial 'item' be read.
> > > > (I'm sure non-blocking IO is also valid...)  
> > > 
> > > I see now. If a partial read happens, the next call to fread() will
> > > read from after the end of the partial read that happened and it'll be
> > > broken.
> > > Since in nolibc the FILE pointer that gets used isn't really a pointer
> > > but the file descriptor I'm not sure where we'd stash the partial part
> > > so we need to avoid doing the partial read entirely.
> > 
> > Except you can't really avoid the partial read.
> > Doing multiple read() system calls doesn't help.
> > The situation where it can happen probably doesn't happen for nolibc.
> 
> > Is there support for ferror() and/or feof() ?
> > (a global 'u8 fstate[64]' indexed by fd number would probably suffice.)
> > If so you could set the 'error' bit and then error any further fread()s.
> 
> No, neither of them is currently supported.
> Given that the support for 'FILE *' is obviously somewhat limited in
> general and the existing fwrite() shares the same issues, I am leaning
> towards picking up these patches. We can then find a proper solution.
> 
> Users trying to use these APIs really correctly will quite fast find out
> that ferror() and feof() are missing.

I agree with you. The lib's called "nolibc" for a reason: it's not a
full-featured libc. Users have to accept some limitations.

> The only real solution for partial writes that I see with the current
> architecture is to limit the 'size' argument to '1'.

I thought about this as well but I don't know if it will be more
annoying than helpful...

> Maybe even with a static assertion. But there still won't be any way to
> properly signal the issue to the caller. Not that ferror() could
> distinguish the types of error in any case...

Let's see with the current patch how it goes. The first reported issues
(if any) will help us lean in the most desirable direction.

Willy