struct dentry is one of the most critical structures in the kernel. So it's
sad to see it going neglected.
With CONFIG_PROFILING turned on (which is probably the common case at least
for distros and kernel developers), sizeof(struct dcache) == 208 here
(64-bit). This gives 19 objects per slab.
I packed d_mounted into a hole, and took another 4 bytes off the inline
name length to take the padding out from the end of the structure. This
shinks it to 200 bytes. I could have gone the other way and increased the
length to 40, but I'm aiming for a magic number, read on...
I then got rid of the d_cookie pointer. This shrinks it to 192 bytes. Rant:
why was this ever a good idea? The cookie system should increase its hash
size or use a tree or something if lookups are a problem. Also the "fast
dcookie lookups" in oprofile should be moved into the dcookie code -- how
can oprofile possibly care about the dcookie_mutex? It gets dropped after
get_dcookie() returns so it can't be providing any sort of protection.
At 192 bytes, 21 objects fit into a 4K page, saving about 3MB on my system
with ~140 000 entries allocated. 192 is also a multiple of 64, so we get
nice cacheline alignment on 64 and 32 byte line systems -- any given dentry
will now require 3 cachelines to touch all fields wheras previously it
would require 4.
I know the inline name size was chosen quite carefully, however with the
reduction in cacheline footprint, it should actually be just about as fast
to do a name lookup for a 36 character name as it was before the patch (and
faster for other sizes). The memory footprint savings for names which are
<= 32 or > 36 bytes long should more than make up for the memory cost for
33-36 byte names.
Performance is a feature...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The number of lost samples could be greater than the number of
received samples. This patches fixes this. The implementation
introduces return values for add_sample() and add_code().
Signed-off-by: Robert Richter <robert.richter@amd.com>
This function is no longer available after the port to the new ring
buffer. Its removal can lead to incomplete sampling sequences since
IBS samples and backtraces are transfered in multiple samples. Due to
a full buffer, samples could be lost any time. The userspace daemon
has to live with such incomplete sampling sequences as long as the
data within one sample is consistent.
This will be fixed by changing the internal buffer data there all data
of one IBS sample or a backtrace is packed in a single ring buffer
entry. This is possible since the new ring buffer supports variable
data size.
Signed-off-by: Robert Richter <robert.richter@amd.com>
This patch replaces the current oprofile cpu buffer implementation
with the ring buffer provided by the tracing framework. The motivation
here is to leave the pain of implementing ring buffers to others. Oh,
no, there are more advantages. Main reason is the support of different
sample sizes that could be stored in the buffer. Use cases for this
are IBS and Cell spu profiling. Using the new ring buffer ensures
valid and complete samples and allows copying the cpu buffer stateless
without knowing its content. Second it will use generic kernel API and
also reduce code size. And hopefully, there are less bugs.
Since the new tracing ring buffer implementation uses spin locks to
protect the buffer during read/write access, it is difficult to use
the buffer in an NMI handler. In this case, writing to the buffer by
the NMI handler (x86) could occur also during critical sections when
reading the buffer. To avoid this, there are 2 buffers for independent
read and write access. Read access is in process context only, write
access only in the NMI handler. If the read buffer runs empty, both
buffers are swapped atomically. There is potentially a small window
during swapping where the buffers are disabled and samples could be
lost.
Using 2 buffers is a little bit overhead, but the solution is clear
and does not require changes in the ring buffer implementation. It can
be changed to a single buffer solution when the ring buffer access is
implemented as non-locking atomic code.
The new buffer requires more size to store the same amount of samples
because each sample includes an u32 header. Also, there is more code
to execute for buffer access. Nonetheless, the buffer implementation
is proven in the ftrace environment and worth to use also in oprofile.
Patches that changes the internal IBS buffer usage will follow.
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Robert Richter <robert.richter@amd.com>
This patch restores default values for:
/dev/oprofile/cpu_buffer_size
/dev/oprofile/buffer_watershed
/dev/oprofile/buffer_size
when creating the oprofilefs:
# opcontrol --deinit
# opcontrol --init
# cat /dev/oprofile/cpu_buffer_size
8192
# echo 5123 > /dev/oprofile/cpu_buffer_size
# cat /dev/oprofile/cpu_buffer_size
5123
# opcontrol --deinit
# opcontrol --init
# cat /dev/oprofile/cpu_buffer_size
8192
# opcontrol --deinit
This sets the values in a defined state. Before, there was no way to
restore the defaults without rebooting the system or reloading the
module.
Signed-off-by: Robert Richter <robert.richter@amd.com>
Regular bitops don't work as locks on all architectures.
Also: can use non-atomic unlock as no concurrent stores to the word.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Robert Richter <robert.richter@amd.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rric/oprofile: (21 commits)
OProfile: Fix buffer synchronization for IBS
oprofile: hotplug cpu fix
oprofile: fixing whitespaces in arch/x86/oprofile/*
oprofile: fixing whitespaces in arch/x86/oprofile/*
oprofile: fixing whitespaces in drivers/oprofile/*
x86/oprofile: add the logic for enabling additional IBS bits
x86/oprofile: reordering functions in nmi_int.c
x86/oprofile: removing unused function parameter in add_ibs_begin()
oprofile: more whitespace fixes
oprofile: whitespace fixes
OProfile: Rename IBS sysfs dir into "ibs_op"
OProfile: Rework string handling in setup_ibs_files()
OProfile: Rework oprofile_add_ibs_sample() function
oprofile: discover counters for op ppro too
oprofile: Implement Intel architectural perfmon support
oprofile: Don't report Nehalem as core_2
oprofile: drop const in num counters field
Revert "Oprofile Multiplexing Patch"
x86, oprofile: BUG: using smp_processor_id() in preemptible code
x86/oprofile: fix on_each_cpu build error
...
Manually fixed trivial conflicts in
drivers/oprofile/{cpu_buffer.c,event_buffer.h}
The issue is the SPU code is not holding the kernel mutex lock while
adding samples to the kernel buffer.
This patch creates per SPU buffers to hold the data. Data
is added to the buffers from in interrupt context. The data
is periodically pushed to the kernel buffer via a new Oprofile
function oprofile_put_buff(). The oprofile_put_buff() function
is called via a work queue enabling the funtion to acquire the
mutex lock.
The existing user controls for adjusting the per CPU buffer
size is used to control the size of the per SPU buffers.
Similarly, overflows of the SPU buffers are reported by
incrementing the per CPU buffer stats. This eliminates the
need to have architecture specific controls for the per SPU
buffers which is not acceptable to the OProfile user tool
maintainer.
The export of the oprofile add_event_entry() is removed as it
is no longer needed given this patch.
Note, this patch has not addressed the issue of indexing arrays
by the spu number. This still needs to be fixed as the spu
numbering is not guarenteed to be 0 to max_num_spus-1.
Signed-off-by: Carl Love <carll@us.ibm.com>
Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Acked-by: Robert Richter <robert.richter@amd.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The patch is needed since there is some IBS code in add_ibs_begin()
that handles more than one sample per iteration. This requires calling
get_slots() during each loop.
This fixes the current problem, but a proper solution that reworks the
cpu buffer synchronization is needed here in the future.
Signed-off-by: Barry Kasindorf <barry.kasindorf@amd.com>
Signed-off-by: Robert Richter <robert.richter@amd.com>
This patch addresses problems when hotplugging cpus while
profiling. Instead of allocating only online cpus, all possible cpu
buffers are allocated, which allows cpus to be onlined during
operation. If a cpu is offlined before profiling is shutdown
wq_sync_buffer checks for this condition then cancels this work and
does not sync this buffer.
Signed-off-by: Chris J Arges <arges@linux.vnet.ibm.com>
Signed-off-by: Robert Richter <robert.richter@amd.com>
Nick Piggin
Robert Richter
Linus Torvalds
Carl Love
Barry Kasindorf
Chris J Arges
Ingo Molnar