Fix compile failure on Linux

[valgrind.git] / README.solaris

Requirements
------------
- You need a recent Solaris-like OS to compile this port. Solaris 11 or
  any illumos-based distribution should work, Solaris 10 is not supported.
  Running `uname -r` has to print '5.11'.
- Recent GCC tools are required, GCC 3 will probably not work. GCC version
  4.5 (or higher) is recommended.
- Solaris ld has to be the first linker in the PATH. GNU ld cannot be used.
  There is currently no linker check in the configure script but the linking
  phase fails if GNU ld is used. Recent Solaris/illumos distributions are ok.
- A working combination of autotools is required: aclocal, autoheader,
  automake and autoconf have to be found in the PATH. You should be able to
  install pkg:/developer/build/automake and pkg:/developer/build/autoconf
  packages to fulfil this requirement.
- System header files are required. On Solaris, these can be installed with:
    # pkg install system/header
- GNU make is also required. On Solaris, this can be quickly achieved with:
    $ PATH=/usr/gnu/bin:$PATH; export PATH
- For remote debugging support, working GDB is required (see below).
- For running regression tests, GNU sed, grep, awk, diff are required.
  This can be quickly achieved on Solaris by prepending /usr/gnu/bin to PATH.


Compilation
-----------
Please follow the generic instructions in the README file,
in the section 'Building and installing it'.

The configure script detects a canonical host to determine which version of
Valgrind should be built. If the system compiler by default produces 32-bit
binaries then only a 32-bit version of Valgrind will be built. To enable
compilation of both 64-bit and 32-bit versions on such a system, issue the
configure script as follows:
./configure CC='gcc -m64' CXX='g++ -m64'


Oracle Solaris and illumos support
----------------------------------
One of the main goal of this port is to support both Oracle Solaris and
illumos kernels. This is a very hard task because Solaris kernel traditionally
does not provide a stable syscall interface and because Valgrind contains
several parts that are closely tied to the underlying kernel. For these
reasons, the port needs to detect which syscall interfaces are present. This
detection cannot be done easily at run time and is currently implemented as
a set of configure tests. This means that a binary version of this port can be
executed only on a kernel that is compatible with a kernel that was used
during the configure and compilation time.

Main currently-known incompatibilities:
- Solaris 11 (released in November 2011) removed a large set of syscalls where
  *at variant of the syscall was also present, for example, open() versus
  openat(AT_FDCWD) [1]
- syscall number for unlinkat() is 76 on Solaris 11, but 65 on illumos [2]
- illumos (in April 2013) changed interface of the accept() and pipe()
  syscalls [3]
- posix_spawn() functionality is backed up by true spawn() syscall on Solaris 11.4
  whereas illumos and Solaris 11.3 leverage vfork()
- illumos and older Solaris use utimesys() syscall whereas newer Solaris
  uses utimensat()

[1] http://docs.oracle.com/cd/E26502_01/html/E28556/gkzlf.html#gkzip
[2] https://www.illumos.org/issues/521
[3] https://github.com/illumos/illumos-gate/commit/5dbfd19ad5fcc2b779f40f80fa05c1bd28fd0b4e


Limitations
-----------
- The port is Work-In-Progress, many things may not work or they can be subtly
  broken.
- Coredumps produced by Valgrind do not contain all information available,
  especially microstate accounting and processor bindings.
- Accessing contents of /proc/self/psinfo is not thread-safe.  That is because
  Valgrind emulates this file on behalf of the client programs.  Entire
  open() - read() - close() sequence on this file needs to be performed
  atomically.
- Fork limitations: vfork() is translated to fork(), forkall() is not
  supported.
- Valgrind does not track definedness of some eflags (OF, SF, ZF, AF, CF, PF)
  individually for each flag. After a syscall is finished, when a carry flag
  is set and defined, all other mentioned flags will be also defined even
  though they might be undefined before making the syscall.
- System call "execve" with a file descriptor which points to a hardlink
  is currently not supported. That is because from the opened file descriptor
  itself it is not possible to reverse map the intended pathname.
  Examples are fexecve(3C) and isaexec(3C).
- Program headers PT_SUNW_SYSSTAT and PT_SUNW_SYSSTAT_ZONE are not supported.
  That is, programs linked with mapfile directive RESERVE_SEGMENT and attribute
  TYPE equal to SYSSTAT or SYSSTAT_ZONE will cause Valgrind exit. It is not
  possible for Valgrind to arrange mapping of a kernel shared page at the
  address specified in the mapfile for the guest application. There is currently
  no such mechanism in Solaris. Hacky workarounds are possible, though.
- When a thread has no stack then all system calls will result in Valgrind
  crash, even though such system calls use just parameters passed in registers.
  This should happen only in pathological situations when a thread is created
  with custom mmap'ed stack and this stack is then unmap'ed during thread
  execution.


Remote debugging support
------------------------
Solaris port of GDB has a major flaw which prevents remote debugging from
working correctly. Fortunately this flaw has an easy fix [4]. Unfortunately
it is not present in the current GDB 7.6.2. This boils down to several
options:
- Use GDB shipped with Solaris 11.2 which has this flaw fixed.
- Wait until GDB 7.7 becomes available (there won't be other 7.6.x releases).
- Build GDB 7.6.2 with the fix by yourself using the following steps:
    # pkg install developer/gnu-binutils
    $ wget http://ftp.gnu.org/gnu/gdb/gdb-7.6.2.tar.gz
    $ gzip -dc gdb-7.6.2.tar.gz | tar xf -
    $ cd gdb-7.6.2
    $ patch -p1 -i /path/to/valgrind-solaris/solaris/gdb-sol-thread.patch
    $ export LIBS="-lncurses"
    $ export CC="gcc -m64"
    $ ./configure --with-x=no --with-curses --with-libexpat-prefix=/usr/lib
    $ gmake && gmake install

[4] https://sourceware.org/ml/gdb-patches/2013-12/msg00573.html


TODO list
---------
- Fix few remaining failing tests.
- Add more Solaris-specific tests (especially for the door and spawn
  syscalls).
- Provide better error reporting for various subsyscalls.
- Implement storing of extra register state in signal frame.
- Performance comparison against other platforms.
- Prevent SIGPIPE when writing to a socket (coregrind/m_libcfile.c).
- Implement ticket locking for fair scheduling (--fair-sched=yes).
- Implement support in DRD and Helgrind tools for thr_join() with thread == 0.
- Add support for accessing thread-local variables via gdb (auxprogs/getoff.c).
  Requires research on internal libc TLS representation.
- VEX supports AVX, BMI and AVX2. Investigate if they can be enabled on
  Solaris/illumos.
- Investigate support for more flags in AT_SUN_AUXFLAGS.
- Fix Valgrind crash when a thread has no stack and syswrap-main.c accesses
  all possible syscall parameters. Enable helgrind/tests/stackteardown.c
  to see this in effect. Would require awareness of syscall parameter semantics.
- Correctly print arguments of DW_CFA_ORCL_arg_loc in show_CF_instruction() when
  it is implemented in libdwarf.
- Handle a situation when guest program sets SC_CANCEL_FLG in schedctl and
  Valgrind needs to invoke a syscall on its own.


Summary of Solaris 11 Kernel Interfaces Used
--------------------------------------------
Valgrind uses directly the following kernel interfaces (not exhaustive list).
Then, of course, it has very intimate knowledge of all syscalls, many ioctls
and some door calls because it has wrappers around them.
- Syscalls:
  . clock_gettime
  . close
  . connect
  . execve
  . exit
  . faccessat
  . fcntl
  . forksys
  . fstatat
  . getcwd
  . getdents
  . geteuid
  . getgid
  . getgroups
  . getpeername
  . getpid
  . getrlimit
  . getsockname
  . getsockopt
  . gettimeofday
  . kill
  . lseek
  . lwp_create
  . lwp_exit
  . lwp_self
  . lwp_sigqueue
  . mknodat
  . mmap
  . mprotect
  . munmap
  . openat
  . pipe
  . pollsys
  . pread
  . prgpsys
  . pwrite
  . read
  . readlinkat
  . renameat
  . rt_sigprocmask
  . send
  . setrlimit
  . setsockopt
  . sigaction
  . sigreturn
  . sigtimedwait
  . so_socket
  . spawn
  . uname
  . unlinkat
  . waitsys
  . write
- Signal frames. Valgrind decomposes and synthetizes signal frames.
- Flag sc_sigblock flag in the schedctl structure by replacing
  function block_all_signals() from libc. The replacement emulates lwp_sigmask
  syscall. More details in coregrind/vg_preloaded.c.
- Initial stack layout for the main thread is synthetized.
- procfs agent thread and other procfs commands for manipulating the process.
- mmapobj syscall is emulated because it gets in the way of the address space
  manager's control.


Contacts
--------
Please send bug reports and any questions about the port to:
Ivo Raisr <ivosh@ivosh.net>
Petr Pavlu <setup@dagobah.cz>