Re: null iv parameter passed to combo_init()

From: Zhihong Yu <zyu(at)yugabyte(dot)com>
To: Noah Misch <noah(at)leadboat(dot)com>
Cc: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>, PostgreSQL-development <pgsql-hackers(at)postgresql(dot)org>
Subject: Re: null iv parameter passed to combo_init()
Date: 2022-01-14 05:09:30
Message-ID: CALNJ-vTm2T8a7eioxjum43Teuh-UEeqj38cmR+TtmZFshk1cTw@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

On Thu, Jan 13, 2022 at 8:09 PM Noah Misch <noah(at)leadboat(dot)com> wrote:

> On Sun, Jan 09, 2022 at 06:45:09PM -0800, Zhihong Yu wrote:
> > On Sun, Jan 9, 2022 at 1:27 PM Zhihong Yu <zyu(at)yugabyte(dot)com> wrote:
> > > After installing gcc-11, ./configure passed (with
> 0003-memcpy-null.patch).
> > > In the output of `make check-world`, I don't see `runtime error`.
>
> That's expected. With -fsanitize-undefined-trap-on-error, the program will
> generate SIGILL when UBSan detects undefined behavior. To get "runtime
> error"
> messages in the postmaster log, drop -fsanitize-undefined-trap-on-error.
> Both
> ways of running the tests have uses. -fsanitize-undefined-trap-on-error is
> better when you think the code is clean, because a zero "make check-world"
> exit status confirms the code is clean. Once you know the code is unclean
> in
> some way, -fsanitize-undefined-trap-on-error is better for getting details.
>
> > > Though there was a crash (maybe specific to my machine):
> > >
> > > Core was generated by
> > >
> `/nfusr/dev-server/zyu/postgres/tmp_install/usr/local/pgsql/bin/postgres
> > > --singl'.
> > > Program terminated with signal SIGILL, Illegal instruction.
> > > #0 0x000000000050642d in write_item.cold ()
> > > Missing separate debuginfos, use: debuginfo-install
> > > glibc-2.17-325.el7_9.x86_64 nss-pam-ldapd-0.8.13-25.el7.x86_64
> > > sssd-client-1.16.5-10.el7_9.10.x86_64
> > > (gdb) bt
> > > #0 0x000000000050642d in write_item.cold ()
> > > #1 0x0000000000ba9d1b in write_relcache_init_file ()
> > > #2 0x0000000000bb58f7 in RelationCacheInitializePhase3 ()
> > > #3 0x0000000000bd5cb5 in InitPostgres ()
> > > #4 0x0000000000a0a9ea in PostgresMain ()
>
> That is UBSan detecting undefined behavior. A successful patch version
> will
> fix write_item(), among many other places that are currently making
> check-world fail. I get the same when testing your v5 under "gcc (Debian
> 11.2.0-13) 11.2.0". I used the same host as buildfarm member thorntail,
> and I
> configured like this:
>
> ./configure -C --with-lz4 --prefix=$HOME/sw/nopath/pghead
> --enable-tap-tests --enable-debug --enable-depend --enable-cassert
> CC='ccache gcc-11 -fsanitize=undefined -fsanitize-undefined-trap-on-error'
> CFLAGS='-O2 -funwind-tables'
>
> > Earlier I was using devtoolset-11 which had an `Illegal instruction`
> error.
> >
> > I compiled / installed gcc-11 from source (which took whole afternoon).
> > `make check-world` passed with patch v3.
> > In tmp_install/log/install.log, I saw:
> >
> > gcc -Wall -Wmissing-prototypes -Wpointer-arith
> > -Wdeclaration-after-statement -Werror=vla -Wendif-labels
> > -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type
> > -Wformat-security -fno-strict-aliasing -fwrapv
> -fexcess-precision=standard
> > -Wno-format-truncation -Wno-stringop-truncation -fsanitize=undefined
> > -fsanitize-undefined-trap-on-error -I../../src/port -DFRONTEND
> > -I../../src/include -D_GNU_SOURCE -c -o path.o path.c
> > rm -f libpgport.a
>
> Perhaps this self-compiled gcc-11 is defective, being unable to detect the
> instances of undefined behavior that other builds detect. If so, use the
> "devtoolset-11" gcc instead. You're also building without optimization;
> that
> might be the problem.
>

I tried both locally built gcc-11 and devtoolset-11 with configure command
copied from above.
`make world` failed in both cases with:

performing post-bootstrap initialization ... sh: line 1: 24714 Illegal
instruction (core dumped)
".../postgres/tmp_install/.../postgres/bin/postgres" --single -F -O -j -c
search_path=pg_catalog -c exit_on_error=true -c log_checkpoints=false
template1 > /dev/null
child process exited with exit code 132

#0 0x000000000050a8d6 in write_item (data=<optimized out>, len=<optimized
out>, fp=<optimized out>) at relcache.c:6471
#1 0x0000000000c33273 in write_relcache_init_file (shared=true) at
relcache.c:6368
#2 0x0000000000c33c50 in RelationCacheInitializePhase3 () at
relcache.c:4220
#3 0x0000000000c55825 in InitPostgres (in_dbname=<optimized out>,
dboid=3105442800, username=<optimized out>, useroid=<optimized out>,
out_dbname=0x0, override_allow_connections=<optimized out>) at
postinit.c:1014

FYI

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Zhihong Yu 2022-01-14 05:12:31 Re: null iv parameter passed to combo_init()
Previous Message Robert Treat 2022-01-14 04:27:18 Re: Consistently use the function name CreateCheckPoint instead of CreateCheckpoint in code comments