percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
The parent rb_node needs to be initialized to shut up the compiler, even
though we're unlikely to ever hit this issue at run time.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Now that the DWARF unwinder is being used to provide perf callstacks
unwinding speed is an issue. It is no longer being used in exceptional
circumstances where we don't care about runtime performance, e.g. when
panicing, so it makes sense improve performance is possible.
With this patch I saw a 42% improvement in unwind time when calling
return_address(1). Greater improvements will be seen as the number of
levels unwound increases as each unwind is now cheaper.
Note that insertion time has doubled but that's just the price we pay
for keeping the trees balanced. However, this is a one-time cost for
kernel boot/module load and so the improvements in lookup time dominate
the extra time we spend keeping the trees balanced.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Unfortunately, due to poor DWARF info in current toolchains, unwinding
through interrutps cannot be done reliably. The problem is that the
DWARF info for function epilogues is wrong.
Take this standard epilogue sequence,
80003cc4: e3 6f mov r14,r15
80003cc6: 26 4f lds.l @r15+,pr
80003cc8: f6 6e mov.l @r15+,r14
<---- interrupt here
80003cca: f6 6b mov.l @r15+,r11
80003ccc: f6 6a mov.l @r15+,r10
80003cce: f6 69 mov.l @r15+,r9
80003cd0: 0b 00 rts
If we take an interrupt at the highlighted point, the DWARF info will
bogusly claim that the return address can be found at some offset from
the frame pointer, even though the frame pointer was just restored. The
worst part is if the unwinder finds a text address at the bogus stack
address - unwinding will continue, for a bit, until it finally comes
across an unexpected address on the stack and blows up.
The only solution is to stop unwinding once we've calculated the
function that was executing when the interrupt occurred. This PC can be
easily calculated from pt_regs->pc.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
The dwarf unwinder presently attempts to provide a sane PC value if none
is provided, however the logic is broken and cases where a previous valid
dwarf frame exists along with a bogus PC value can still proceed. This
fixes up the test and prevents the unwinder from blowing up.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
When CONFIG_FUNCTION_GRAPH_TRACER is enabled the function graph tracer
may patch return addresses on the stack with the address of
return_to_handler(). This really confuses the DWARF unwinder because it
will try find the caller of return_to_handler(), not the caller of the
real return address.
So teach the DWARF unwinder how to find the real return address whenever
it encounters return_to_handler().
This patch does not cope very well when multiple return addresses on the
stack have been patched. To make it work properly it would require state
to track how many return_to_handler()'s have been seen so that we'd know
where to look in current->curr_ret_stack[]. So for now, instead of
trying to handle this, just moan if more than one return address on the
stack has been patched.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This enables us to build the dwarf unwinder both with modules enabled and
disabled in addition to reducing code size in the latter case. The
helpers are also consolidated, and modified to resemble the BUG module
helpers.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Originally, dwarf_unwind_stack() was a recursive function and it seems
that some of the old comments were never updated.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
If we broke out of the while (1) loop because the return address of
"frame" was zero, then "frame" needs to be free'd before we return.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Pass a module's .eh_frame section to the DWARF unwinder at module load
time so that the section's FDEs and CIEs can be registered with the
DWARF unwinder. This allows us to unwind the stack through module code
when generating backtraces.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
fix the following 'make includecheck' warning:
arch/sh/kernel/dwarf.c: asm/dwarf.h is included more than once.
Signed-off-by: Jaswinder Singh Rajput <jaswinderrajput@gmail.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
The CIE and FDE structs are big enough and accessed regularly enough in
certain configurations to make cacheline alignment useful.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Allow a DWARF register to have an undefined value. When applied to the
DWARF return address register this lets lets us label a function as
having no direct caller, e.g. kernel_thread_helper().
Signed-off-by: Matt Fleming <matt@console-pimps.org>
The 'end' member of struct dwarf_fde denotes one byte past the end of
the CFA instruction stream for an FDE. The value of 'end' was being
calcualted incorrectly, it was being set too high. This resulted in
dwarf_cfa_execute_insns() interpreting data past the end of valid
instructions, thus causing all sorts of weird crashes.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
We can't assume that if we execute the unwinder code and the unwinder
was already running that it has faulted. Clearly two kernel threads can
invoke the unwinder at the same time and may be running simultaneously.
The previous approach used BUG() and BUG_ON() in the unwinder code to
detect whether the unwinder was incapable of unwinding the stack, and
that the next available unwinder should be used instead. A better
approach is to explicitly invoke a trap handler to switch unwinders when
the current unwinder cannot continue.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
The handling of DW_CFA_val_offset ops was incorrectly using the
DWARF_REG_OFFSET flag but the register's value cannot be calculated
using the DWARF_REG_OFFSET method. Create a new flag to indicate that a
different method must be used to calculate the register's value even
though there is no implementation for DWARF_VAL_OFFSET yet; it's mainly
just a place holder.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Plug a memory leak in dwarf_unwinder_dump() where we didn't free the
memory that we had previously allocated for the DWARF frames and DWARF
registers.
Now is also a opportune time to implement our own mempool and kmem
cache. It's a good idea to have a certain number of frame and register
objects in reserve at all times, so that we are guaranteed to have our
allocation satisfied even when memory is scarce. Since we have pools to
allocate from we can implement the registers for each frame as a linked
list as opposed to a sparsely populated array. Whilst it's true that the
lookup time for a linked list is larger than for arrays, there's only
usually a maximum of 8 registers per frame. So the overhead isn't that
much of a concern.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
This moves the initialization over to an early_initcall(). This fixes up
some lockdep interaction issues. At the same time, kill off some
superfluous locking in the init path.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Also, remove the "fix" to DW_CFA_def_cfa_register where we reset the
frame's cfa_offset to 0. This action is incorrect when handling
DW_CFA_def_cfa_register as the DWARF spec specifically states that the
previous contents of cfa_offset should be used with the new
register. The reason that I thought cfa_offset should be reset to 0 was
because it was being assigned a bogus value prior to executing the
DW_CFA_def_cfa_register op. It turns out that the bogus cfa_offset value
came from interpreting .cfi_escape pseudo-ops (those used by the GNU
extensions) as CFA_DW_def_cfa ops.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
The previous hack for calculating the return address for the first frame
we unwind (dwarf_unwinder_dump) didn't always work. The problem was that
it assumed once it read the rule for calculating the return address,
there would be no new rules for calculating it. This isn't true because
the way in which the CFA is calculated can change as you progress
through a function and the return address is figured out using the
CFA. Therefore, the way to calculate the return address can change.
So, instead of using some offset from the beginning of
dwarf_unwind_stack which is just a flakey approach, and instead of
executing instructions from the FDE until the return address is setup,
we now figure out the pc in dwarf_unwind_stack() just before we call
dwarf_cfa_execute_insns().
Signed-off-by: Matt Fleming <matt@console-pimps.org>
The way that the CFA is calculated can change as we progress through a
function. If we see a DW_CFA_def_cfa_register op we need to reset the
frame's cfa_offset value which may have been previously setup.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
save_stack_trace_tsk() and friends can be called from atomic context (as
triggered by latencytop), and subsequently hit two problematic allocation
points that were using GFP_KERNEL (these were dwarf_unwind_stack() and
dwarf_frame_alloc_regs()). Convert these over to GFP_ATOMIC and get
latencytop working with the DWARF unwinder.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Trying to figure out the best value for DWARF_ARCH_UNWIND_OFFSET is
tricky at best. Various things can change the size (and offset from the
beginning of the function) of the prologue. Notably, turning on ftrace
adds calls to mcount at the beginning of functions, thereby pushing the
prologue further into the function.
So replace DWARF_ARCH_UNWIND_OFFSET with some code that continues to
execute CFA instructions until the value of return address register is
defined. This is safe to do because we know that the return address must
have been pushed onto the frame before our first function call; we just
can't figure out where at compile-time.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
The destination address might be unaligned, so set it with
put_unaligned() for safety. This restores the previous behaviour, albeit
through the proper API.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This was using internal symbols for unaligned accesses, bypassing the
exposed interface for variable sized safe accesses. This converts all of
the __get_unaligned_cpuXX() users over to get_unaligned() directly,
relying on the cast to select the proper internal routine.
Additionally, the __put_unaligned_cpuXX() case is superfluous given that
the destination address is aligned in all of the current cases, so just
drop that outright.
Furthermore, this switches to the asm/unaligned.h header instead of the
asm-generic version, which was silently bypassing the SH-4A optimized
unaligned ops.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This is a first cut at a generic DWARF unwinder for the kernel. It's
still lacking DWARF64 support and the DWARF expression support hasn't
been tested very well but it is generating proper stacktraces on SH for
WARN_ON() and NULL dereferences.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Tejun Heo