btrace: bridge gaps

btrace: bridge gaps

Most of the time, the trace should be in one piece.  This case is handled fine
by GDB.  In some cases, however, there may be gaps in the trace.  They result
from trace decode errors or from overflows.

A gap in the trace means we lost an unknown amount of trace.  Gaps can be very
small, such as a few instructions in the same function, or they can be rather
big.  We may, for example, lose a few function calls or returns.  The trace may
continue in a different function and we likely don't know how we got there.

Even though we can't say how the program executed across a gap, higher levels
may not be impacted too much by it.  Let's assume we have functions a-e and a
trace that looks roughly like this:

  a
   \
    b                    b
     \                  /
      c   <gap>        c
                      /
                 d   d
                  \ /
                   e

Even though we can't say for sure, it is likely that b and c are the same
function instance before and after the gap.  This patch is trying to connect
the c and b function segments across the gap.

This will add a to the back trace of b on the right hand side.  The changes are
reflected in GDB's internal representation of the trace and will improve:

  - the output of "record function-call-history /c"
  - the output of "backtrace" in replay mode
  - source stepping in replay mode
    will be improved indirectly via the improved back trace

I don't have an automated test for this patch; decode errors will be fixed and
overflows occur sporadically and are quite rare.  I tested it by hacking GDB to
provoke a decode error and on the expected gap in the gdb.btrace/dlopen.exp
test.

The issue is that we can't predict where we will be able to re-sync in case of
errors.  For the expected decode error in gdb.btrace/dlopen.exp, for example, we
may be able to re-sync somewhere in dlclose, in test, in main, or not at all.

Here's one example run of gdb.btrace/dlopen.exp with and without this patch.

    (gdb) info record
    Active record target: record-btrace
    Recording format: Intel Processor Trace.
    Buffer size: 16kB.
    warning: Non-contiguous trace at instruction 66608 (offset = 0xa83, pc = 0xb7fdcc31).
    warning: Non-contiguous trace at instruction 66652 (offset = 0xa9b, pc = 0xb7fdcc31).
    warning: Non-contiguous trace at instruction 66770 (offset = 0xacb, pc = 0xb7fdcc31).
    warning: Non-contiguous trace at instruction 66966 (offset = 0xb60, pc = 0xb7ff5ee4).
    warning: Non-contiguous trace at instruction 66994 (offset = 0xb74, pc = 0xb7ff5f24).
    warning: Non-contiguous trace at instruction 67334 (offset = 0xbac, pc = 0xb7ff5e6d).
    warning: Non-contiguous trace at instruction 69022 (offset = 0xc04, pc = 0xb7ff60b3).
    warning: Non-contiguous trace at instruction 69116 (offset = 0xc1c, pc = 0xb7ff60b3).
    warning: Non-contiguous trace at instruction 69504 (offset = 0xc74, pc = 0xb7ff605d).
    warning: Non-contiguous trace at instruction 83648 (offset = 0xecc, pc = 0xb7ff6134).
    warning: Decode error (-13) at instruction 83876 (offset = 0xf48, pc = 0xb7fd6380): no memory mapped at this address.
    warning: Non-contiguous trace at instruction 83876 (offset = 0x11b7, pc = 0xb7ff1c70).
    Recorded 83948 instructions in 912 functions (12 gaps) for thread 1 (process 12996).
    (gdb) record instruction-history 83876, +2
    83876   => 0xb7fec46f <call_init.part.0+95>:    call   *%eax
    [decode error (-13): no memory mapped at this address]
    [disabled]
    83877      0xb7ff1c70 <_dl_close_worker.part.0+1584>:   nop

Without the patch, the trace is disconnected and the backtrace is short:

    (gdb) record goto 83876
    #0  0xb7fec46f in call_init.part () from /lib/ld-linux.so.2
    (gdb) backtrace
    #0  0xb7fec46f in call_init.part () from /lib/ld-linux.so.2
    #1  0xb7fec5d0 in _dl_init () from /lib/ld-linux.so.2
    #2  0xb7ff0fe3 in dl_open_worker () from /lib/ld-linux.so.2
    Backtrace stopped: not enough registers or memory available to unwind further
    (gdb) record goto 83877
    #0  0xb7ff1c70 in _dl_close_worker.part.0 () from /lib/ld-linux.so.2
    (gdb) backtrace
    #0  0xb7ff1c70 in _dl_close_worker.part.0 () from /lib/ld-linux.so.2
    #1  0xb7ff287a in _dl_close () from /lib/ld-linux.so.2
    #2  0xb7fc3d5d in dlclose_doit () from /lib/libdl.so.2
    #3  0xb7fec354 in _dl_catch_error () from /lib/ld-linux.so.2
    #4  0xb7fc43dd in _dlerror_run () from /lib/libdl.so.2
    #5  0xb7fc3d98 in dlclose () from /lib/libdl.so.2
    #6  0x0804860a in test ()
    #7  0x08048628 in main ()

With the patch, GDB is able to connect the trace pieces and we get a full
backtrace.

    (gdb) record goto 83876
    #0  0xb7fec46f in call_init.part () from /lib/ld-linux.so.2
    (gdb) backtrace
    #0  0xb7fec46f in call_init.part () from /lib/ld-linux.so.2
    #1  0xb7fec5d0 in _dl_init () from /lib/ld-linux.so.2
    #2  0xb7ff0fe3 in dl_open_worker () from /lib/ld-linux.so.2
    #3  0xb7fec354 in _dl_catch_error () from /lib/ld-linux.so.2
    #4  0xb7ff02e2 in _dl_open () from /lib/ld-linux.so.2
    #5  0xb7fc3c65 in dlopen_doit () from /lib/libdl.so.2
    #6  0xb7fec354 in _dl_catch_error () from /lib/ld-linux.so.2
    #7  0xb7fc43dd in _dlerror_run () from /lib/libdl.so.2
    #8  0xb7fc3d0e in dlopen@@GLIBC_2.1 () from /lib/libdl.so.2
    #9  0xb7ff28ee in _dl_runtime_resolve () from /lib/ld-linux.so.2
    #10 0x0804841c in ?? ()
    #11 0x08048470 in dlopen@plt ()
    #12 0x080485a3 in test ()
    #13 0x08048628 in main ()
    (gdb) record goto 83877
    #0  0xb7ff1c70 in _dl_close_worker.part.0 () from /lib/ld-linux.so.2
    (gdb) backtrace
    #0  0xb7ff1c70 in _dl_close_worker.part.0 () from /lib/ld-linux.so.2
    #1  0xb7ff287a in _dl_close () from /lib/ld-linux.so.2
    #2  0xb7fc3d5d in dlclose_doit () from /lib/libdl.so.2
    #3  0xb7fec354 in _dl_catch_error () from /lib/ld-linux.so.2
    #4  0xb7fc43dd in _dlerror_run () from /lib/libdl.so.2
    #5  0xb7fc3d98 in dlclose () from /lib/libdl.so.2
    #6  0x0804860a in test ()
    #7  0x08048628 in main ()

It worked nicely in this case but it may, of course, also lead to weird
connections; it is a heuristic, after all.

It works best when the gap is small and the trace pieces are long.

gdb/
	* btrace.c (bfun_s): New typedef.
	(ftrace_update_caller): Print caller in debug dump.
	(ftrace_get_caller, ftrace_match_backtrace, ftrace_fixup_level)
	(ftrace_compute_global_level_offset, ftrace_connect_bfun)
	(ftrace_connect_backtrace, ftrace_bridge_gap, btrace_bridge_gaps): New.
	(btrace_compute_ftrace_bts): Pass vector of gaps.  Collect gaps.
	(btrace_compute_ftrace_pt): Likewise.
	(btrace_compute_ftrace): Split into this, ...
	(btrace_compute_ftrace_1): ... this, and ...
	(btrace_finalize_ftrace): ... this.  Call btrace_bridge_gaps.