On MIPS calls to _mcount in modules generate 2 instructions to load
the _mcount address (and therefore 2 relocations). The mcount_loc
table should only reference the first of these, so the second is
filtered out by checking the relocation offset and ignoring ones that
immediately follow the previous one seen.
However if a module has an _mcount call at offset 0, the second
relocation would not be filtered out due to old_r_offset == 0
being taken to mean that the current relocation is the first one
seen, and both would end up in the mcount_loc table.
This results in ftrace_make_nop() patching both (adjacent)
instructions to branches over the _mcount call sequence like so:
0xffffffffc08a8000: 04 00 00 10 b 0xffffffffc08a8014
0xffffffffc08a8004: 04 00 00 10 b 0xffffffffc08a8018
0xffffffffc08a8008: 2d 08 e0 03 move at,ra
...
The second branch is in the delay slot of the first, which is
defined to be unpredictable - on the platform on which this bug was
encountered, it triggers a reserved instruction exception.
Fix by initializing old_r_offset to ~0 and using that instead of 0
to determine whether the current relocation is the first seen.
Signed-off-by: Alex Smith <alex.smith@imgtec.com>
Cc: linux-kernel@vger.kernel.org
Cc: stable@vger.kernel.org
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/7098/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
With gcc 4.6.0 the -mfentry feature places the function profiling
call at the start of the function. When this is used, the call is
to __fentry__ and not mcount.
Change recordmcount.c to record both callers to __fentry__ and
mcount.
Link: http://lkml.kernel.org/r/20120807194058.990674363@goodmis.org
Acked-by: H. Peter Anvin <hpa@linux.intel.com>
Acked-by: John Reiser <jreiser@bitwagon.com>
Acked-by: Ingo Molnar <mingo@kernel.org>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
In ELF64, the sh_flags field is 64-bits wide. recordmcount was
erroneously treating it as a 32-bit wide field. For little endian
objects this works because the flags of interest (SHF_EXECINSTR)
reside in the lower 32 bits of the word, and you get the same result
with either a 32-bit or 64-bit read. Big endian objects on the
other hand do not work at all with this error.
The fix: Correctly treat sh_flags as 64-bits wide in elf64 objects.
The symptom I observed was that my
__start_mcount_loc..__stop_mcount_loc was empty even though ftrace
function tracing was enabled.
Link: http://lkml.kernel.org/r/1324345362-12230-1-git-send-email-ddaney.cavm@gmail.com
Cc: stable@kernel.org # 3.0+
Signed-off-by: David Daney <david.daney@cavium.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
While find_secsym_ndx often finds the unamed local STT_SECTION, if a
section has only one function in it, the ARM toolchain generates the
STT_FUNC symbol before the STT_SECTION, and recordmcount finds this
instead.
This is problematic on ARM because in ARM ELFs, "if a [STT_FUNC] symbol
addresses a Thumb instruction, its value is the address of the
instruction with bit zero set (in a relocatable object, the section
offset with bit zero set)". This leads to incorrect mcount addresses
being recorded.
Fix this by not using STT_FUNC symbols as the base on ARM.
Signed-off-by: Rabin Vincent <rabin@rab.in>
Link: http://lkml.kernel.org/r/1305134631-31617-1-git-send-email-rabin@rab.in
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Introduce mcount_adjust{,_32,_64} to the C implementation of
recordmcount analog to $mcount_adjust in the perl script.
The adjustment is added to the address of the relocations
against the mcount symbol. If this adjustment is done by
recordmcount at compile time the ftrace_call_adjust function
can be turned into a nop.
Cc: John Reiser <jreiser@bitwagon.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The code to get the symbol, string, and relp pointers in the two functions
sift_rel_mcount() and nop_mcount() are identical and also non-trivial.
Moving this duplicate code into a single helper function makes the code
easier to read and more maintainable.
Cc: John Reiser <jreiser@bitwagon.com>
Link: http://lkml.kernel.org/r/20110421023739.723658553@goodmis.org
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The code in sift_rel_mcount() and nop_mcount() to get the mcount symbol
number is identical. Replace the two locations with a call to a function
that does the work.
Cc: John Reiser <jreiser@bitwagon.com>
Link: http://lkml.kernel.org/r/20110421023739.488093407@goodmis.org
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
There's some sections that should not have mcount recorded and should not have
modifications to the that code. But currently they waste some time by calling
mcount anyway (which simply returns). As the real answer should be to
either whitelist the section or have gcc ignore it fully.
This change adds a option to recordmcount to warn when it finds a section
that is ignored by ftrace but still contains mcount callers. This is not on
by default as developers may not know if the section should be completely
ignored or added to the whitelist.
Cc: John Reiser <jreiser@bitwagon.com>
Link: http://lkml.kernel.org/r/20110421023738.476989377@goodmis.org
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
There are sections that are ignored by ftrace for the function tracing because
the text is in a section that can be removed without notice. The mcount calls
in these sections are ignored and ftrace never sees them. The downside of this
is that the functions in these sections still call mcount. Although the mcount
function is defined in assembly simply as a return, this added overhead is
unnecessary.
The solution is to convert these callers into nops at compile time.
A better solution is to add 'notrace' to the section markers, but as new sections
come up all the time, it would be nice that they are delt with when they
are created.
Later patches will deal with finding these sections and doing the proper solution.
Thanks to H. Peter Anvin for giving me the right nops to use for x86.
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: John Reiser <jreiser@bitwagon.com>
Link: http://lkml.kernel.org/r/20110421023738.237101176@goodmis.org
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
PROGBITS is not enough to determine if the section should be modified
or not. Only process sections that are marked as executable.
Cc: John Reiser <jreiser@bitwagon.com>
Link: http://lkml.kernel.org/r/20110421023737.991485123@goodmis.org
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The Linux ftrace subsystem style for comparing is:
var == 1
var > 0
and not:
1 == var
0 < var
It is considered that Linux developers are smart enough not to do the
if (var = 1)
mistake.
Cc: John Reiser <jreiser@bitwagon.com>
Link: http://lkml.kernel.org/r/20110421023737.290712238@goodmis.org
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Depending on the compiler version, ARM GCC calls the mcount function
either __gnu_mcount_nc or mcount.
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Rabin Vincent <rabin@rab.in>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
It looks to me like the change which introduced "virtual functions"
forgot about cross-platform endianness.
Thank you to Arnaud for supplying before+after data files do_mounts*.o.
This fixes a MIPS build failure triggered by recordmcount.
Reported-by: Arnaud Lacombe <lacombar@gmail.com>
Tested-by: Arnaud Lacombe <lacombar@gmail.com>
Acked-by: Wu Zhangjin <wuzhangjin@gmail.com>
Acked-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: John Reiser <jreiser@BitWagon.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Since MIPS modules' address space differs from the core kernel space, to access
the _mcount in the core kernel, the kernel functions in modules must use long
call (-mlong-calls): load the _mcount address into one register and jump to the
address stored by the register:
c: 3c030000 lui v1,0x0 <--------> b label
c: R_MIPS_HI16 _mcount
c: R_MIPS_NONE *ABS*
c: R_MIPS_NONE *ABS*
10: 64630000 daddiu v1,v1,0
10: R_MIPS_LO16 _mcount
10: R_MIPS_NONE *ABS*
10: R_MIPS_NONE *ABS*
14: 03e0082d move at,ra
18: 0060f809 jalr v1
label:
In the old Perl version of recordmcount, we only need to record the position of
the 1st R_MIPS_HI16 type of _mcount, and later, in ftrace_make_nop(), replace
the instruction in this position by a "b label" and in ftrace_make_call(),
replace it back.
But, the default C version of recordmcount records all of the _mcount symbols,
so, we must filter the 2nd _mcount like the Perl version of recordmcount does.
The C version of recordmcount copes with the symbols before they are linked, So
It doesn't know the type of the symbols and therefore can not filter the
symbols as the Perl version of recordmcount does. But as we can see above, the
2nd _mcount symbols of the long call alawys follows the 1st _mcount symbol of
the same long call, which means the offset from the 1st to the 2nd is fixed, it
is 0x10-0xc = 4 here, 4 is the length of the 1st load instruciton, for MIPS has
fixed length of instructions, this offset is always 4.
And as we know, the _mcount is inserted into the entry of every kernel
function, the offset between the other _mcount's is expected to be always
bigger than 4. So, to filter the 2ns _mcount symbol of the long call, we can
simply check the offset between two _mcount symbols, If it is 4, then, filter
the 2nd _mcount symbol.
To avoid touching too much code, an 'empty' function fn_is_fake_mcount() is
added for all of the archs, and the specific archs can override it via chaning
the function pointer: is_fake_mcount in do_file() with the e_machine. e.g. This
patch adds MIPS_is_fake_mcount() to override the default fn_is_fake_mcount()
pointed by is_fake_mcount.
This fn_is_fake_mcount() checks if the _mcount symbol is fake, e.g. the 2nd
_mcount symbol of the long call is fake, for there are 2 _mcount symbols mapped
to one real mcount call, so, one of them is fake and must be filtered.
This fn_is_fake_mcount() is called in sift_rel_mcount() after finding the
_mcount symbols and before adding the _mcount symbol into mrelp, so, it can
prevent the fake mcount symbol going into the last __mcount_loc table.
Signed-off-by: Wu Zhangjin <wuzhangjin@gmail.com>
LKML-Reference: <b866f0138224340a132d31861fa3f9300dee30ac.1288176026.git.wuzhangjin@gmail.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
MIPS64 has 'weird' Elf64_Rel.r_info[1,2], which must be used instead of
the generic Elf64_Rel.r_info, otherwise, the C version of recordmcount
will not work for "segmentation fault".
Usage of "union mips_r_info" and the functions MIPS64_r_sym() and
MIPS64_r_info() written by Maciej W. Rozycki <macro@linux-mips.org>
----
[1] http://techpubs.sgi.com/library/manuals/4000/007-4658-001/pdf/007-4658-001.pdf
[2] arch/mips/include/asm/module.h
Tested-by: Wu Zhangjin <wuzhangjin@gmail.com>
Signed-off-by: John Reiser <jreiser@BitWagon.com>
Signed-off-by: Maciej W. Rozycki <macro@linux-mips.org>
LKML-Reference: <AANLkTinwXjLAYACUfhLYaocHD_vBbiErLN3NjwN8JqSy@mail.gmail.com>
LKML-Reference: <910dc2d5ae1ed042df4f96815fe4a433078d1c2a.1288176026.git.wuzhangjin@gmail.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
The elf reader for recordmcount.c had duplicate functions for both
32 bit and 64 bit elf handling. This was due to the need of using
the 32 and 64 bit elf structures.
This patch consolidates the two by using macros to define the 32
and 64 bit names in a recordmcount.h file, and then by just defining
a RECORD_MCOUNT_64 macro and including recordmcount.h twice we
create the funtions for both the 32 bit version as well as the
64 bit version using one code source.
Cc: John Reiser <jreiser@bitwagon.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
