Hello Paul,

please consider the current patch series for merging into your fixes queue.
The intent is to get them before 6.2, then backported where relevant.

It addresses the following bugs:
  - fd_set was incorrectly defined as arrays of u32 instead of long,
    which breaks BE64. Fix courtesy of Sven Schnelle.

  - S_ISxxx macros were incorrectly testing the bits after applying them
    instead of applying S_ISFMT to the value. Fix from Warner Losh.

  - the mips code was randomly broken due to an unprotected "noreorder"
    directive in the _start code that would prevent the assembler from
    filling delayed slots, and randomly leaving other instructions there

  - since the split of the single include file into multiple files, we're
    implicitly refraining from including some which are not explicitly
    added in the code. It causes build failures when such files contain
    definitions for functions that may be used e.g. by libgcc, such as
    raise() or memset(), which are often called only by a few archs at
    certain optimization levels only.

  - gcc 11.3 in ARM thumb2 mode at -O2 was able to recognize a memset()
    construction inside the memset() definition, and it replaced it with
    a call to... memset(). We cannot impose to userland to build with
    -ffreestanding so the introduction of an empty asm() statement in
    the loop was enough to stop this.

  - most of the O_* macros were wrong on RISCV because their octal value
    was used as a hexadecimal one when the platform was introduced. This
    was revealed by the selftest breaking in getdents64().

The series was tested on x86_64, i386, armv5, armv7, thumb1, thumb2,
mips and riscv, all at -O0, -Os and -O3. This is based on the "nolibc"
branch of your linux-rcu tree. Do not hesitate to let me know if you
prefer that I rebase it on a different one.

Thank you!
Willy

---
Sven Schnelle (1):
  nolibc: fix fd_set type

Warner Losh (1):
  tools/nolibc: Fix S_ISxxx macros

Willy Tarreau (4):
  tools/nolibc: restore mips branch ordering in the _start block
  tools/nolibc: fix missing includes causing build issues at -O0
  tools/nolibc: prevent gcc from making memset() loop over itself
  tools/nolibc: fix the O_* fcntl/open macro definitions for riscv

 tools/include/nolibc/arch-mips.h  |  2 +
 tools/include/nolibc/arch-riscv.h | 14 +++----
 tools/include/nolibc/ctype.h      |  3 ++
 tools/include/nolibc/errno.h      |  3 ++
 tools/include/nolibc/signal.h     |  3 ++
 tools/include/nolibc/stdio.h      |  3 ++
 tools/include/nolibc/stdlib.h     |  3 ++
 tools/include/nolibc/string.h     |  8 +++-
 tools/include/nolibc/sys.h        |  2 +
 tools/include/nolibc/time.h       |  3 ++
 tools/include/nolibc/types.h      | 70 ++++++++++++++++++-------------
 tools/include/nolibc/unistd.h     |  3 ++
 12 files changed, 79 insertions(+), 38 deletions(-)

-- 
2.35.3

On Mon, Jan 09, 2023 at 08:54:36AM +0100, Willy Tarreau wrote:
> Hello Paul,
> 
> please consider the current patch series for merging into your fixes queue.
> The intent is to get them before 6.2, then backported where relevant.
> 
> It addresses the following bugs:
>   - fd_set was incorrectly defined as arrays of u32 instead of long,
>     which breaks BE64. Fix courtesy of Sven Schnelle.
> 
>   - S_ISxxx macros were incorrectly testing the bits after applying them
>     instead of applying S_ISFMT to the value. Fix from Warner Losh.
> 
>   - the mips code was randomly broken due to an unprotected "noreorder"
>     directive in the _start code that would prevent the assembler from
>     filling delayed slots, and randomly leaving other instructions there
> 
>   - since the split of the single include file into multiple files, we're
>     implicitly refraining from including some which are not explicitly
>     added in the code. It causes build failures when such files contain
>     definitions for functions that may be used e.g. by libgcc, such as
>     raise() or memset(), which are often called only by a few archs at
>     certain optimization levels only.
> 
>   - gcc 11.3 in ARM thumb2 mode at -O2 was able to recognize a memset()
>     construction inside the memset() definition, and it replaced it with
>     a call to... memset(). We cannot impose to userland to build with
>     -ffreestanding so the introduction of an empty asm() statement in
>     the loop was enough to stop this.
> 
>   - most of the O_* macros were wrong on RISCV because their octal value
>     was used as a hexadecimal one when the platform was introduced. This
>     was revealed by the selftest breaking in getdents64().
> 
> The series was tested on x86_64, i386, armv5, armv7, thumb1, thumb2,
> mips and riscv, all at -O0, -Os and -O3. This is based on the "nolibc"
> branch of your linux-rcu tree. Do not hesitate to let me know if you
> prefer that I rebase it on a different one.

"81 test(s) passed", so queued at urgent-nolibc.2023.01.09a, thank you!

Also, thank you for the detailed cover letter, which I co-opted into the
signed tag.  But please check to make sure that my wordsmithing didn't
break anything.

If all goes well, I will send the pull request to Linus before the end
of this week.

							Thanx, Paul

> Thank you!
> Willy
> 
> ---
> Sven Schnelle (1):
>   nolibc: fix fd_set type
> 
> Warner Losh (1):
>   tools/nolibc: Fix S_ISxxx macros
> 
> Willy Tarreau (4):
>   tools/nolibc: restore mips branch ordering in the _start block
>   tools/nolibc: fix missing includes causing build issues at -O0
>   tools/nolibc: prevent gcc from making memset() loop over itself
>   tools/nolibc: fix the O_* fcntl/open macro definitions for riscv
> 
>  tools/include/nolibc/arch-mips.h  |  2 +
>  tools/include/nolibc/arch-riscv.h | 14 +++----
>  tools/include/nolibc/ctype.h      |  3 ++
>  tools/include/nolibc/errno.h      |  3 ++
>  tools/include/nolibc/signal.h     |  3 ++
>  tools/include/nolibc/stdio.h      |  3 ++
>  tools/include/nolibc/stdlib.h     |  3 ++
>  tools/include/nolibc/string.h     |  8 +++-
>  tools/include/nolibc/sys.h        |  2 +
>  tools/include/nolibc/time.h       |  3 ++
>  tools/include/nolibc/types.h      | 70 ++++++++++++++++++-------------
>  tools/include/nolibc/unistd.h     |  3 ++
>  12 files changed, 79 insertions(+), 38 deletions(-)
> 
> -- 
> 2.35.3
>

Hello Paul,

On Mon, Jan 09, 2023 at 11:11:41AM -0800, Paul E. McKenney wrote:
> On Mon, Jan 09, 2023 at 08:54:36AM +0100, Willy Tarreau wrote:
> > Hello Paul,
> > 
> > please consider the current patch series for merging into your fixes queue.
> > The intent is to get them before 6.2, then backported where relevant.
> > 
> > It addresses the following bugs:
> >   - fd_set was incorrectly defined as arrays of u32 instead of long,
> >     which breaks BE64. Fix courtesy of Sven Schnelle.
> > 
> >   - S_ISxxx macros were incorrectly testing the bits after applying them
> >     instead of applying S_ISFMT to the value. Fix from Warner Losh.
> > 
> >   - the mips code was randomly broken due to an unprotected "noreorder"
> >     directive in the _start code that would prevent the assembler from
> >     filling delayed slots, and randomly leaving other instructions there
> > 
> >   - since the split of the single include file into multiple files, we're
> >     implicitly refraining from including some which are not explicitly
> >     added in the code. It causes build failures when such files contain
> >     definitions for functions that may be used e.g. by libgcc, such as
> >     raise() or memset(), which are often called only by a few archs at
> >     certain optimization levels only.
> > 
> >   - gcc 11.3 in ARM thumb2 mode at -O2 was able to recognize a memset()
> >     construction inside the memset() definition, and it replaced it with
> >     a call to... memset(). We cannot impose to userland to build with
> >     -ffreestanding so the introduction of an empty asm() statement in
> >     the loop was enough to stop this.
> > 
> >   - most of the O_* macros were wrong on RISCV because their octal value
> >     was used as a hexadecimal one when the platform was introduced. This
> >     was revealed by the selftest breaking in getdents64().
> > 
> > The series was tested on x86_64, i386, armv5, armv7, thumb1, thumb2,
> > mips and riscv, all at -O0, -Os and -O3. This is based on the "nolibc"
> > branch of your linux-rcu tree. Do not hesitate to let me know if you
> > prefer that I rebase it on a different one.
> 
> "81 test(s) passed", so queued at urgent-nolibc.2023.01.09a, thank you!
> 
> Also, thank you for the detailed cover letter, which I co-opted into the
> signed tag.

You're welcome!

> But please check to make sure that my wordsmithing didn't
> break anything.

It all looks perfect to me.

> If all goes well, I will send the pull request to Linus before the end
> of this week.

Great, thank you!
Willy