diff options
| author | Markus Metzger <markus.t.metzger@intel.com> | 2015-08-11 11:05:58 +0200 |
|---|---|---|
| committer | Markus Metzger <markus.t.metzger@intel.com> | 2015-09-09 10:35:35 +0200 |
| commit | 0568462bbf0f666d5bf9f720e16147da71eec46a (patch) | |
| tree | 59f573e2faef6955c19c23d3ce7ca2c6b9534391 /gdb/nat/linux-btrace.h | |
| parent | c0fa8fbd1c2f02921b90083e595a8452d3d51ba2 (diff) | |
| download | gdb-0568462bbf0f666d5bf9f720e16147da71eec46a.zip gdb-0568462bbf0f666d5bf9f720e16147da71eec46a.tar.gz gdb-0568462bbf0f666d5bf9f720e16147da71eec46a.tar.bz2 | |
btrace: kernel address filtering
For the BTS recording format, we sometimes get a FROM->TO record where the
FROM address lies in the kernel and the TO address lies in user space at
whatever address the user process was resumed.
GDB has a heuristic to filter out such records based on looking at the most
significant bit in the PC. This works fine for 64-bit systems but it doesn't
always work for 32-bit systems. Libraries that are loaded at fairly high
addresses might be mistaken for kernel code and branches inside the library
are filtered out.
Change the heuristic to (again heuristically) try to determine the lowest
address in kernel space. Any PC that is smaller than that should be in
user space.
On today's systems, there should be a symbol "_text" at that address.
Read /proc/kallsyms and search for that symbol.
It is not guaranteed that /proc/kallsyms is readable on all systems. On
64-bit systems, we fall back to check the most significant bit. On 32-bit
systems, we refrain from filtering out addresses.
The filtering should really be done by the kernel. And it soon will be:
https://lkml.org/lkml/2015/8/31/212.
gdb/
* nat/linux-btrace.h (struct btrace_target_info) <ptr_bits>: Remove.
* nat/linux-btrace.c: Include filestuff.h and inttypes.h.
Remove include of sys/utsname.h.
(linux_determine_kernel_ptr_bits): Remove.
(linux_determine_kernel_start): New.
(perf_event_is_kernel_addr): Remove tinfo argument. Update users.
Update check.
(perf_event_skip_bts_record): Remove tinfo argument. Update users.
(linux_enable_bts, linux_enable_pt): Remove tinfo->ptr_bits
initialization.
* x86-linux-nat.c (x86_linux_enable_btrace): Remove ptr_bits
assignment.
gdbserver/
* linux-low.c (linux_low_enable_btrace): Remove.
(linux_target_ops): Replace linux_low_enable_btrace with
linux_enable_btrace.
Diffstat (limited to 'gdb/nat/linux-btrace.h')
| -rw-r--r-- | gdb/nat/linux-btrace.h | 5 |
1 files changed, 0 insertions, 5 deletions
diff --git a/gdb/nat/linux-btrace.h b/gdb/nat/linux-btrace.h index 5ea87a8..042878c 100644 --- a/gdb/nat/linux-btrace.h +++ b/gdb/nat/linux-btrace.h @@ -100,11 +100,6 @@ struct btrace_target_info struct btrace_tinfo_pt pt; } variant; #endif /* HAVE_LINUX_PERF_EVENT_H */ - - /* The size of a pointer in bits for this thread. - The information is used to identify kernel addresses in order to skip - records from/to kernel space. */ - int ptr_bits; }; /* See to_supports_btrace in target.h. */ |