Re: [PATCH v2 4/4] linux-user: fix use of SIGRTMIN

[Peter Maydell]

On Tue, 4 Feb 2020 at 17:11, Laurent Vivier <address@hidden> wrote:
>
> Some RT signals can be in use by glibc,
> it's why SIGRTMIN (34) is generally greater than __SIGRTMIN (32).
>
> So SIGRTMIN cannot be mapped to TARGET_SIGRTMIN.
>
> Instead of swapping only SIGRTMIN and SIGRTMAX, map all the
> range [TARGET_SIGRTMIN ... TARGET_SIGRTMAX - X] to
>       [__SIGRTMIN + X ... SIGRTMAX ]
> (SIGRTMIN is __SIGRTMIN + X).
>
> Signed-off-by: Laurent Vivier <address@hidden>
> ---

In general I think this is a good approach to trying to deal
with this long-standing issue in a pragmatic and not too
complicated way, so thanks for writing this patchset. I have
some fairly minor comments on the code below.

>
> Notes:
>     v2: ignore error when target sig <= TARGET_NSIG but host sig > SIGRTMAX
>         replace i, j by target_sig, host_sig
>         update signal_table_init() trace message
>
>  linux-user/signal.c     | 37 ++++++++++++++++++++++++++++++-------
>  linux-user/trace-events |  3 +++
>  2 files changed, 33 insertions(+), 7 deletions(-)
>
> diff --git a/linux-user/signal.c b/linux-user/signal.c
> index c1e664f97a7c..e7e5581a016f 100644
> --- a/linux-user/signal.c
> +++ b/linux-user/signal.c
> @@ -498,18 +498,23 @@ static int core_dump_signal(int sig)
>
>  static void signal_table_init(void)
>  {
> -    int host_sig, target_sig;
> +    int host_sig, target_sig, count;
>
>      /*
> -     * Nasty hack: Reverse SIGRTMIN and SIGRTMAX to avoid overlap with
> -     * host libpthread signals.  This assumes no one actually uses SIGRTMAX 
> :-/
> -     * To fix this properly we need to do manual signal delivery multiplexed
> -     * over a single host signal.
> +     * some RT signals can be in use by glibc,
> +     * it's why SIGRTMIN (34) is generally greater than __SIGRTMIN (32)
>       */
> -    host_to_target_signal_table[__SIGRTMIN] = __SIGRTMAX;
> -    host_to_target_signal_table[__SIGRTMAX] = __SIGRTMIN;
> +    for (host_sig = SIGRTMIN; host_sig <= SIGRTMAX; host_sig++) {
> +        target_sig = host_sig - SIGRTMIN + TARGET_SIGRTMIN;
> +        if (target_sig <= TARGET_NSIG) {
> +            host_to_target_signal_table[host_sig] = target_sig;
> +        }
> +    }

So the effect of this is that we now support target signals
starting from TARGET_SIGRTMIN and going up until we run out
of host realtime signals that the host libc hasn't reserved ?
That seems reasonable, since glibc at least uses only the
lower 2 rt signals and probably nobody's using the upper ones.
But this would be a good place to have a comment explaining
the limitation (and that if it needed to be fixed we'd have
to multiplex guest signals onto a single host signal). You
could also mention that attempts to configure the "missing"
signals via sigaction will be silently ignored.

>      /* generate signal conversion tables */
> +    for (target_sig = 1; target_sig <= TARGET_NSIG; target_sig++) {
> +        target_to_host_signal_table[target_sig] = _NSIG; /* poison */
> +    }
>      for (host_sig = 1; host_sig < _NSIG; host_sig++) {
>          if (host_to_target_signal_table[host_sig] == 0) {
>              host_to_target_signal_table[host_sig] = host_sig;
> @@ -519,6 +524,15 @@ static void signal_table_init(void)
>              target_to_host_signal_table[target_sig] = host_sig;
>          }
>      }
> +
> +    if (TRACE_SIGNAL_TABLE_INIT_BACKEND_DSTATE()) {

This isn't the right way to conditionalize expensive stuff
that's only used in trace events. You want to use
trace_event_get_state_backends() (see docs/devel/tracing.txt
for details).

> +        for (target_sig = 1, count = 0; target_sig <= TARGET_NSIG; 
> target_sig++) {
> +            if (target_to_host_signal_table[target_sig] == _NSIG) {
> +                count++;
> +            }
> +        }
> +        trace_signal_table_init(count);
> +    }
>  }
>
>  void signal_init(void)
> @@ -817,6 +831,8 @@ int do_sigaction(int sig, const struct target_sigaction 
> *act,
>      int host_sig;
>      int ret = 0;
>
> +    trace_signal_do_sigaction_guest(sig, TARGET_NSIG);
> +
>      if (sig < 1 || sig > TARGET_NSIG || sig == TARGET_SIGKILL || sig == 
> TARGET_SIGSTOP) {
>          return -TARGET_EINVAL;
>      }
> @@ -847,6 +863,13 @@ int do_sigaction(int sig, const struct target_sigaction 
> *act,
>
>          /* we update the host linux signal state */
>          host_sig = target_to_host_signal(sig);
> +        trace_signal_do_sigaction_host(host_sig, TARGET_NSIG);
> +        if (host_sig > SIGRTMAX) {
> +            /* we don't have enough host signals to map all target signals */
> +            qemu_log_mask(LOG_UNIMP, "Unsupported target signal #%d, 
> ignored\n",
> +                          sig);
> +            return 0;

We should have a comment here mentioning why we don't return
an error code here (and explicitly noting that the Go runtime
is the major one which we don't want to upset).

> +        }
>          if (host_sig != SIGSEGV && host_sig != SIGBUS) {
>              sigfillset(&act1.sa_mask);
>              act1.sa_flags = SA_SIGINFO;
> diff --git a/linux-user/trace-events b/linux-user/trace-events
> index f6de1b8befc0..0296133daeb6 100644
> --- a/linux-user/trace-events
> +++ b/linux-user/trace-events
> @@ -1,6 +1,9 @@
>  # See docs/devel/tracing.txt for syntax documentation.
>
>  # signal.c
> +signal_table_init(int i) "number of unavailable signals: %d"
> +signal_do_sigaction_guest(int sig, int max) "target signal %d (MAX %d)"
> +signal_do_sigaction_host(int sig, int max) "host signal %d (MAX %d)"
>  # */signal.c
>  user_setup_frame(void *env, uint64_t frame_addr) "env=%p 
> frame_addr=0x%"PRIx64
>  user_setup_rt_frame(void *env, uint64_t frame_addr) "env=%p 
> frame_addr=0x%"PRIx64

thanks
-- PMM