Merge branch 'perf-core-for-linus' of...

What:		/sys/kernel/debug/kmemtrace/

Date:		July 2008

Contact:	Eduard - Gabriel Munteanu <eduard.munteanu@linux360.ro>

Description:

In kmemtrace-enabled kernels, the following files are created:

/sys/kernel/debug/kmemtrace/

	cpu<n>		(0400)	Per-CPU tracing data, see below. (binary)

	total_overruns	(0400)	Total number of bytes which were dropped from

				cpu<n> files because of full buffer condition,

				non-binary. (text)

	abi_version	(0400)	Kernel's kmemtrace ABI version. (text)

Each per-CPU file should be read according to the relay interface. That is,

the reader should set affinity to that specific CPU and, as currently done by

the userspace application (though there are other methods), use poll() with

an infinite timeout before every read(). Otherwise, erroneous data may be

read. The binary data has the following _core_ format:

	Event ID	(1 byte)	Unsigned integer, one of:

		0 - represents an allocation (KMEMTRACE_EVENT_ALLOC)

		1 - represents a freeing of previously allocated memory

		    (KMEMTRACE_EVENT_FREE)

	Type ID		(1 byte)	Unsigned integer, one of:

		0 - this is a kmalloc() / kfree()

		1 - this is a kmem_cache_alloc() / kmem_cache_free()

		2 - this is a __get_free_pages() et al.

	Event size	(2 bytes)	Unsigned integer representing the

					size of this event. Used to extend

					kmemtrace. Discard the bytes you

					don't know about.

	Sequence number	(4 bytes)	Signed integer used to reorder data

					logged on SMP machines. Wraparound

					must be taken into account, although

					it is unlikely.

	Caller address	(8 bytes)	Return address to the caller.

	Pointer to mem	(8 bytes)	Pointer to target memory area. Can be

					NULL, but not all such calls might be

					recorded.

In case of KMEMTRACE_EVENT_ALLOC events, the next fields follow:

	Requested bytes	(8 bytes)	Total number of requested bytes,

					unsigned, must not be zero.

	Allocated bytes (8 bytes)	Total number of actually allocated

					bytes, unsigned, must not be lower

					than requested bytes.

	Requested flags	(4 bytes)	GFP flags supplied by the caller.

	Target CPU	(4 bytes)	Signed integer, valid for event id 1.

					If equal to -1, target CPU is the same

					as origin CPU, but the reverse might

					not be true.

The data is made available in the same endianness the machine has.

Other event ids and type ids may be defined and added. Other fields may be

added by increasing event size, but see below for details.

Every modification to the ABI, including new id definitions, are followed

by bumping the ABI version by one.

Adding new data to the packet (features) is done at the end of the mandatory

data:

	Feature size	(2 byte)

	Feature ID	(1 byte)

	Feature data	(Feature size - 3 bytes)

Users:

	kmemtrace-user - git://repo.or.cz/kmemtrace-user.git

	nousb		[USB] Disable the USB subsystem

	nowatchdog	[KNL] Disable the lockup detector.

	nowb		[ARM]

	nox2apic	[X86-64,APIC] Do not enable x2APIC mode.

want more explanation of a feature in terms of common code, review the common

ftrace.txt file.

Ideally, everyone who wishes to retain performance while supporting tracing in

their kernel should make it all the way to dynamic ftrace support.

Prerequisites

-------------

exiting of a function.  On exit, the value is compared and if it does not

match, then it will panic the kernel.  This is largely a sanity check for bad

code generation with gcc.  If gcc for your port sanely updates the frame

pointer under different opitmization levels, then ignore this option.

pointer under different optimization levels, then ignore this option.

However, adding support for it isn't terribly difficult.  In your assembly code

that calls prepare_ftrace_return(), pass the frame pointer as the 3rd argument.

HAVE_SYSCALL_TRACEPOINTS

---------------------

------------------------

You need very few things to get the syscalls tracing in an arch.

HAVE_FTRACE_MCOUNT_RECORD

-------------------------

See scripts/recordmcount.pl for more info.

See scripts/recordmcount.pl for more info.  Just fill in the arch-specific

details for how to locate the addresses of mcount call sites via objdump.

This option doesn't make much sense without also implementing dynamic ftrace.

HAVE_DYNAMIC_FTRACE

-------------------

You will first need HAVE_FTRACE_MCOUNT_RECORD and HAVE_FUNCTION_TRACER, so

scroll your reader back up if you got over eager.

Once those are out of the way, you will need to implement:

	- asm/ftrace.h:

		- MCOUNT_ADDR

		- ftrace_call_adjust()

		- struct dyn_arch_ftrace{}

	- asm code:

		- mcount() (new stub)

		- ftrace_caller()

		- ftrace_call()

		- ftrace_stub()

	- C code:

		- ftrace_dyn_arch_init()

		- ftrace_make_nop()

		- ftrace_make_call()

		- ftrace_update_ftrace_func()

First you will need to fill out some arch details in your asm/ftrace.h.

Define MCOUNT_ADDR as the address of your mcount symbol similar to:

	#define MCOUNT_ADDR ((unsigned long)mcount)

Since no one else will have a decl for that function, you will need to:

	extern void mcount(void);

You will also need the helper function ftrace_call_adjust().  Most people

will be able to stub it out like so:

	static inline unsigned long ftrace_call_adjust(unsigned long addr)

	{

		return addr;

	}

<details to be filled>

Lastly you will need the custom dyn_arch_ftrace structure.  If you need

some extra state when runtime patching arbitrary call sites, this is the

place.  For now though, create an empty struct:

	struct dyn_arch_ftrace {

		/* No extra data needed */

	};

With the header out of the way, we can fill out the assembly code.  While we

did already create a mcount() function earlier, dynamic ftrace only wants a

stub function.  This is because the mcount() will only be used during boot

and then all references to it will be patched out never to return.  Instead,

the guts of the old mcount() will be used to create a new ftrace_caller()

function.  Because the two are hard to merge, it will most likely be a lot

easier to have two separate definitions split up by #ifdefs.  Same goes for

the ftrace_stub() as that will now be inlined in ftrace_caller().

Before we get confused anymore, let's check out some pseudo code so you can

implement your own stuff in assembly:

HAVE_DYNAMIC_FTRACE

---------------------

void mcount(void)

{

	return;

}

void ftrace_caller(void)

{

	/* implement HAVE_FUNCTION_TRACE_MCOUNT_TEST if you desire */

	/* save all state needed by the ABI (see paragraph above) */

	unsigned long frompc = ...;

	unsigned long selfpc = <return address> - MCOUNT_INSN_SIZE;

ftrace_call:

	ftrace_stub(frompc, selfpc);

	/* restore all state needed by the ABI */

ftrace_stub:

	return;

}

This might look a little odd at first, but keep in mind that we will be runtime

patching multiple things.  First, only functions that we actually want to trace

will be patched to call ftrace_caller().  Second, since we only have one tracer

active at a time, we will patch the ftrace_caller() function itself to call the

specific tracer in question.  That is the point of the ftrace_call label.

With that in mind, let's move on to the C code that will actually be doing the

runtime patching.  You'll need a little knowledge of your arch's opcodes in

order to make it through the next section.

Every arch has an init callback function.  If you need to do something early on

to initialize some state, this is the time to do that.  Otherwise, this simple

function below should be sufficient for most people:

int __init ftrace_dyn_arch_init(void *data)

{

	/* return value is done indirectly via data */

	*(unsigned long *)data = 0;

	return 0;

}

There are two functions that are used to do runtime patching of arbitrary

functions.  The first is used to turn the mcount call site into a nop (which

is what helps us retain runtime performance when not tracing).  The second is

used to turn the mcount call site into a call to an arbitrary location (but

typically that is ftracer_caller()).  See the general function definition in

linux/ftrace.h for the functions:

	ftrace_make_nop()

	ftrace_make_call()

The rec->ip value is the address of the mcount call site that was collected

by the scripts/recordmcount.pl during build time.

The last function is used to do runtime patching of the active tracer.  This

will be modifying the assembly code at the location of the ftrace_call symbol

inside of the ftrace_caller() function.  So you should have sufficient padding

at that location to support the new function calls you'll be inserting.  Some

people will be using a "call" type instruction while others will be using a

"branch" type instruction.  Specifically, the function is:

	ftrace_update_ftrace_func()

HAVE_DYNAMIC_FTRACE + HAVE_FUNCTION_GRAPH_TRACER

------------------------------------------------

The function grapher needs a few tweaks in order to work with dynamic ftrace.

Basically, you will need to:

	- update:

		- ftrace_caller()

		- ftrace_graph_call()

		- ftrace_graph_caller()

	- implement:

		- ftrace_enable_ftrace_graph_caller()

		- ftrace_disable_ftrace_graph_caller()

<details to be filled>

Quick notes:

	- add a nop stub after the ftrace_call location named ftrace_graph_call;

	  stub needs to be large enough to support a call to ftrace_graph_caller()

	- update ftrace_graph_caller() to work with being called by the new

	  ftrace_caller() since some semantics may have changed

	- ftrace_enable_ftrace_graph_caller() will runtime patch the

	  ftrace_graph_call location with a call to ftrace_graph_caller()

	- ftrace_disable_ftrace_graph_caller() will runtime patch the

	  ftrace_graph_call location with nops

			kmemtrace - Kernel Memory Tracer

			  by Eduard - Gabriel Munteanu

			     <eduard.munteanu@linux360.ro>

I. Introduction

===============

kmemtrace helps kernel developers figure out two things:

1) how different allocators (SLAB, SLUB etc.) perform

2) how kernel code allocates memory and how much

To do this, we trace every allocation and export information to the userspace

through the relay interface. We export things such as the number of requested

bytes, the number of bytes actually allocated (i.e. including internal

fragmentation), whether this is a slab allocation or a plain kmalloc() and so

on.

The actual analysis is performed by a userspace tool (see section III for

details on where to get it from). It logs the data exported by the kernel,

processes it and (as of writing this) can provide the following information:

- the total amount of memory allocated and fragmentation per call-site

- the amount of memory allocated and fragmentation per allocation

- total memory allocated and fragmentation in the collected dataset

- number of cross-CPU allocation and frees (makes sense in NUMA environments)

Moreover, it can potentially find inconsistent and erroneous behavior in

kernel code, such as using slab free functions on kmalloc'ed memory or

allocating less memory than requested (but not truly failed allocations).

kmemtrace also makes provisions for tracing on some arch and analysing the

data on another.

II. Design and goals

====================

kmemtrace was designed to handle rather large amounts of data. Thus, it uses

the relay interface to export whatever is logged to userspace, which then

stores it. Analysis and reporting is done asynchronously, that is, after the

data is collected and stored. By design, it allows one to log and analyse

on different machines and different arches.

As of writing this, the ABI is not considered stable, though it might not

change much. However, no guarantees are made about compatibility yet. When

deemed stable, the ABI should still allow easy extension while maintaining

backward compatibility. This is described further in Documentation/ABI.

Summary of design goals:

	- allow logging and analysis to be done across different machines

	- be fast and anticipate usage in high-load environments (*)

	- be reasonably extensible

	- make it possible for GNU/Linux distributions to have kmemtrace

	included in their repositories

(*) - one of the reasons Pekka Enberg's original userspace data analysis

    tool's code was rewritten from Perl to C (although this is more than a

    simple conversion)

III. Quick usage guide

======================

1) Get a kernel that supports kmemtrace and build it accordingly (i.e. enable

CONFIG_KMEMTRACE).

2) Get the userspace tool and build it:

$ git clone git://repo.or.cz/kmemtrace-user.git		# current repository

$ cd kmemtrace-user/

$ ./autogen.sh

$ ./configure

$ make

3) Boot the kmemtrace-enabled kernel if you haven't, preferably in the

'single' runlevel (so that relay buffers don't fill up easily), and run

kmemtrace:

# '$' does not mean user, but root here.

$ mount -t debugfs none /sys/kernel/debug

$ mount -t proc none /proc

$ cd path/to/kmemtrace-user/

$ ./kmemtraced

Wait a bit, then stop it with CTRL+C.

$ cat /sys/kernel/debug/kmemtrace/total_overruns	# Check if we didn't

							# overrun, should

							# be zero.

$ (Optionally) [Run kmemtrace_check separately on each cpu[0-9]*.out file to

		check its correctness]

$ ./kmemtrace-report

Now you should have a nice and short summary of how the allocator performs.

IV. FAQ and known issues

========================

Q: 'cat /sys/kernel/debug/kmemtrace/total_overruns' is non-zero, how do I fix

this? Should I worry?

A: If it's non-zero, this affects kmemtrace's accuracy, depending on how

large the number is. You can fix it by supplying a higher

'kmemtrace.subbufs=N' kernel parameter.

---

Q: kmemtrace_check reports errors, how do I fix this? Should I worry?

A: This is a bug and should be reported. It can occur for a variety of

reasons:

	- possible bugs in relay code

	- possible misuse of relay by kmemtrace

	- timestamps being collected unorderly

Or you may fix it yourself and send us a patch.

---

Q: kmemtrace_report shows many errors, how do I fix this? Should I worry?

A: This is a known issue and I'm working on it. These might be true errors

in kernel code, which may have inconsistent behavior (e.g. allocating memory

with kmem_cache_alloc() and freeing it with kfree()). Pekka Enberg pointed

out this behavior may work with SLAB, but may fail with other allocators.

It may also be due to lack of tracing in some unusual allocator functions.

We don't want bug reports regarding this issue yet.

---

V. See also

===========

Documentation/kernel-parameters.txt

Documentation/ABI/testing/debugfs-kmemtrace

  +|-offs(FETCHARG) : Fetch memory at FETCHARG +|- offs address.(**)

  NAME=FETCHARG : Set NAME as the argument name of FETCHARG.

  FETCHARG:TYPE : Set TYPE as the type of FETCHARG. Currently, basic types

		  (u8/u16/u32/u64/s8/s16/s32/s64) are supported.

		  (u8/u16/u32/u64/s8/s16/s32/s64) and string are supported.

  (*) only for return probe.

  (**) this is useful for fetching a field of data structures.