Getting started with kmemcheck
==============================

Vegard Nossum <vegardno@ifi.uio.no>


Introduction
------------

kmemcheck is a debugging feature for the Linux Kernel. More specifically, it
is a dynamic checker that detects and warns about some uses of uninitialized
memory.

Userspace programmers might be familiar with Valgrind's memcheck. The main
difference between memcheck and kmemcheck is that memcheck works for userspace
programs only, and kmemcheck works for the kernel only. The implementations
are of course vastly different. Because of this, kmemcheck is not as accurate
as memcheck, but it turns out to be good enough in practice to discover real
programmer errors that the compiler is not able to find through static
analysis.

Enabling kmemcheck on a kernel will probably slow it down to the extent that
the machine will not be usable for normal workloads such as e.g. an
interactive desktop. kmemcheck will also cause the kernel to use about twice
as much memory as normal. For this reason, kmemcheck is strictly a debugging
feature.


Downloading
-----------

As of version 2.6.31-rc1, kmemcheck is included in the mainline kernel.


Configuring and compiling
-------------------------

kmemcheck only works for the x86 (both 32- and 64-bit) platform. A number of
configuration variables must have specific settings in order for the kmemcheck
menu to even appear in "menuconfig". These are:

- ``CONFIG_CC_OPTIMIZE_FOR_SIZE=n``
	This option is located under "General setup" / "Optimize for size".

	Without this, gcc will use certain optimizations that usually lead to
	false positive warnings from kmemcheck. An example of this is a 16-bit
	field in a struct, where gcc may load 32 bits, then discard the upper
	16 bits. kmemcheck sees only the 32-bit load, and may trigger a
	warning for the upper 16 bits (if they're uninitialized).

- ``CONFIG_SLAB=y`` or ``CONFIG_SLUB=y``
	This option is located under "General setup" / "Choose SLAB
	allocator".

- ``CONFIG_FUNCTION_TRACER=n``
	This option is located under "Kernel hacking" / "Tracers" / "Kernel
	Function Tracer"

	When function tracing is compiled in, gcc emits a call to another
	function at the beginning of every function. This means that when the
	page fault handler is called, the ftrace framework will be called
	before kmemcheck has had a chance to handle the fault. If ftrace then
	modifies memory that was tracked by kmemcheck, the result is an
	endless recursive page fault.

- ``CONFIG_DEBUG_PAGEALLOC=n``
	This option is located under "Kernel hacking" / "Memory Debugging"
	/ "Debug page memory allocations".

In addition, I highly recommend turning on ``CONFIG_DEBUG_INFO=y``. This is also
located under "Kernel hacking". With this, you will be able to get line number
information from the kmemcheck warnings, which is extremely valuable in
debugging a problem. This option is not mandatory, however, because it slows
down the compilation process and produces a much bigger kernel image.

Now the kmemcheck menu should be visible (under "Kernel hacking" / "Memory
Debugging" / "kmemcheck: trap use of uninitialized memory"). Here follows
a description of the kmemcheck configuration variables:

- ``CONFIG_KMEMCHECK``
	This must be enabled in order to use kmemcheck at all...

- ``CONFIG_KMEMCHECK_``[``DISABLED`` | ``ENABLED`` | ``ONESHOT``]``_BY_DEFAULT``
	This option controls the status of kmemcheck at boot-time. "Enabled"
	will enable kmemcheck right from the start, "disabled" will boot the
	kernel as normal (but with the kmemcheck code compiled in, so it can
	be enabled at run-time after the kernel has booted), and "one-shot" is
	a special mode which will turn kmemcheck off automatically after
	detecting the first use of uninitialized memory.

	If you are using kmemcheck to actively debug a problem, then you
	probably want to choose "enabled" here.

	The one-shot mode is mostly useful in automated test setups because it
	can prevent floods of warnings and increase the chances of the machine
	surviving in case something is really wrong. In other cases, the one-
	shot mode could actually be counter-productive because it would turn
	itself off at the very first error -- in the case of a false positive
	too -- and this would come in the way of debugging the specific
	problem you were interested in.

	If you would like to use your kernel as normal, but with a chance to
	enable kmemcheck in case of some problem, it might be a good idea to
	choose "disabled" here. When kmemcheck is disabled, most of the run-
	time overhead is not incurred, and the kernel will be almost as fast
	as normal.

- ``CONFIG_KMEMCHECK_QUEUE_SIZE``
	Select the maximum number of error reports to store in an internal
	(fixed-size) buffer. Since errors can occur virtually anywhere and in
	any context, we need a temporary storage area which is guaranteed not
	to generate any other page faults when accessed. The queue will be
	emptied as soon as a tasklet may be scheduled. If the queue is full,
	new error reports will be lost.

	The default value of 64 is probably fine. If some code produces more
	than 64 errors within an irqs-off section, then the code is likely to
	produce many, many more, too, and these additional reports seldom give
	any more information (the first report is usually the most valuable
	anyway).

	This number might have to be adjusted if you are not using serial
	console or similar to capture the kernel log. If you are using the
	"dmesg" command to save the log, then getting a lot of kmemcheck
	warnings might overflow the kernel log itself, and the earlier reports
	will get lost in that way instead. Try setting this to 10 or so on
	such a setup.

- ``CONFIG_KMEMCHECK_SHADOW_COPY_SHIFT``
	Select the number of shadow bytes to save along with each entry of the
	error-report queue. These bytes indicate what parts of an allocation
	are initialized, uninitialized, etc. and will be displayed when an
	error is detected to help the debugging of a particular problem.

	The number entered here is actually the logarithm of the number of
	bytes that will be saved. So if you pick for example 5 here, kmemcheck
	will save 2^5 = 32 bytes.

	The default value should be fine for debugging most problems. It also
	fits nicely within 80 columns.

- ``CONFIG_KMEMCHECK_PARTIAL_OK``
	This option (when enabled) works around certain GCC optimizations that
	produce 32-bit reads from 16-bit variables where the upper 16 bits are
	thrown away afterwards.

	The default value (enabled) is recommended. This may of course hide
	some real errors, but disabling it would probably produce a lot of
	false positives.

- ``CONFIG_KMEMCHECK_BITOPS_OK``
	This option silences warnings that would be generated for bit-field
	accesses where not all the bits are initialized at the same time. This
	may also hide some real bugs.

	This option is probably obsolete, or it should be replaced with
	the kmemcheck-/bitfield-annotations for the code in question. The
	default value is therefore fine.

Now compile the kernel as usual.


How to use
----------

Booting
~~~~~~~

First some information about the command-line options. There is only one
option specific to kmemcheck, and this is called "kmemcheck". It can be used
to override the default mode as chosen by the ``CONFIG_KMEMCHECK_*_BY_DEFAULT``
option. Its possible settings are:

- ``kmemcheck=0`` (disabled)
- ``kmemcheck=1`` (enabled)
- ``kmemcheck=2`` (one-shot mode)

If SLUB debugging has been enabled in the kernel, it may take precedence over
kmemcheck in such a way that the slab caches which are under SLUB debugging
will not be tracked by kmemcheck. In order to ensure that this doesn't happen
(even though it shouldn't by default), use SLUB's boot option ``slub_debug``,
like this: ``slub_debug=-``

In fact, this option may also be used for fine-grained control over SLUB vs.
kmemcheck. For example, if the command line includes
``kmemcheck=1 slub_debug=,dentry``, then SLUB debugging will be used only
for the "dentry" slab cache, and with kmemcheck tracking all the other
caches. This is advanced usage, however, and is not generally recommended.


Run-time enable/disable
~~~~~~~~~~~~~~~~~~~~~~~

When the kernel has booted, it is possible to enable or disable kmemcheck at
run-time. WARNING: This feature is still experimental and may cause false
positive warnings to appear. Therefore, try not to use this. If you find that
it doesn't work properly (e.g. you see an unreasonable amount of warnings), I
will be happy to take bug reports.

Use the file ``/proc/sys/kernel/kmemcheck`` for this purpose, e.g.::

	$ echo 0 > /proc/sys/kernel/kmemcheck # disables kmemcheck

The numbers are the same as for the ``kmemcheck=`` command-line option.


Debugging
~~~~~~~~~

A typical report will look something like this::

    WARNING: kmemcheck: Caught 32-bit read from uninitialized memory (ffff88003e4a2024)
    80000000000000000000000000000000000000000088ffff0000000000000000
     i i i i u u u u i i i i i i i i u u u u u u u u u u u u u u u u
             ^

    Pid: 1856, comm: ntpdate Not tainted 2.6.29-rc5 #264 945P-A
    RIP: 0010:[<ffffffff8104ede8>]  [<ffffffff8104ede8>] __dequeue_signal+0xc8/0x190
    RSP: 0018:ffff88003cdf7d98  EFLAGS: 00210002
    RAX: 0000000000000030 RBX: ffff88003d4ea968 RCX: 0000000000000009
    RDX: ffff88003e5d6018 RSI: ffff88003e5d6024 RDI: ffff88003cdf7e84
    RBP: ffff88003cdf7db8 R08: ffff88003e5d6000 R09: 0000000000000000
    R10: 0000000000000080 R11: 0000000000000000 R12: 000000000000000e
    R13: ffff88003cdf7e78 R14: ffff88003d530710 R15: ffff88003d5a98c8
    FS:  0000000000000000(0000) GS:ffff880001982000(0063) knlGS:00000
    CS:  0010 DS: 002b ES: 002b CR0: 0000000080050033
    CR2: ffff88003f806ea0 CR3: 000000003c036000 CR4: 00000000000006a0
    DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
    DR3: 0000000000000000 DR6: 00000000ffff4ff0 DR7: 0000000000000400
     [<ffffffff8104f04e>] dequeue_signal+0x8e/0x170
     [<ffffffff81050bd8>] get_signal_to_deliver+0x98/0x390
     [<ffffffff8100b87d>] do_notify_resume+0xad/0x7d0
     [<ffffffff8100c7b5>] int_signal+0x12/0x17
     [<ffffffffffffffff>] 0xffffffffffffffff

The single most valuable information in this report is the RIP (or EIP on 32-
bit) value. This will help us pinpoint exactly which instruction that caused
the warning.

If your kernel was compiled with ``CONFIG_DEBUG_INFO=y``, then all we have to do
is give this address to the addr2line program, like this::

	$ addr2line -e vmlinux -i ffffffff8104ede8
	arch/x86/include/asm/string_64.h:12
	include/asm-generic/siginfo.h:287
	kernel/signal.c:380
	kernel/signal.c:410

The "``-e vmlinux``" tells addr2line which file to look in. **IMPORTANT:**
This must be the vmlinux of the kernel that produced the warning in the
first place! If not, the line number information will almost certainly be
wrong.

The "``-i``" tells addr2line to also print the line numbers of inlined
functions.  In this case, the flag was very important, because otherwise,
it would only have printed the first line, which is just a call to
``memcpy()``, which could be called from a thousand places in the kernel, and
is therefore not very useful.  These inlined functions would not show up in
the stack trace above, simply because the kernel doesn't load the extra
debugging information. This technique can of course be used with ordinary
kernel oopses as well.

In this case, it's the caller of ``memcpy()`` that is interesting, and it can be
found in ``include/asm-generic/siginfo.h``, line 287::

    281 static inline void copy_siginfo(struct siginfo *to, struct siginfo *from)
    282 {
    283         if (from->si_code < 0)
    284                 memcpy(to, from, sizeof(*to));
    285         else
    286                 /* _sigchld is currently the largest know union member */
    287                 memcpy(to, from, __ARCH_SI_PREAMBLE_SIZE + sizeof(from->_sifields._sigchld));
    288 }

Since this was a read (kmemcheck usually warns about reads only, though it can
warn about writes to unallocated or freed memory as well), it was probably the
"from" argument which contained some uninitialized bytes. Following the chain
of calls, we move upwards to see where "from" was allocated or initialized,
``kernel/signal.c``, line 380::

    359 static void collect_signal(int sig, struct sigpending *list, siginfo_t *info)
    360 {
    ...
    367         list_for_each_entry(q, &list->list, list) {
    368                 if (q->info.si_signo == sig) {
    369                         if (first)
    370                                 goto still_pending;
    371                         first = q;
    ...
    377         if (first) {
    378 still_pending:
    379                 list_del_init(&first->list);
    380                 copy_siginfo(info, &first->info);
    381                 __sigqueue_free(first);
    ...
    392         }
    393 }

Here, it is ``&first->info`` that is being passed on to ``copy_siginfo()``. The
variable ``first`` was found on a list -- passed in as the second argument to
``collect_signal()``. We  continue our journey through the stack, to figure out
where the item on "list" was allocated or initialized. We move to line 410::

    395 static int __dequeue_signal(struct sigpending *pending, sigset_t *mask,
    396                         siginfo_t *info)
    397 {
    ...
    410                 collect_signal(sig, pending, info);
    ...
    414 }

Now we need to follow the ``pending`` pointer, since that is being passed on to
``collect_signal()`` as ``list``. At this point, we've run out of lines from the
"addr2line" output. Not to worry, we just paste the next addresses from the
kmemcheck stack dump, i.e.::

     [<ffffffff8104f04e>] dequeue_signal+0x8e/0x170
     [<ffffffff81050bd8>] get_signal_to_deliver+0x98/0x390
     [<ffffffff8100b87d>] do_notify_resume+0xad/0x7d0
     [<ffffffff8100c7b5>] int_signal+0x12/0x17

	$ addr2line -e vmlinux -i ffffffff8104f04e ffffffff81050bd8 \
		ffffffff8100b87d ffffffff8100c7b5
	kernel/signal.c:446
	kernel/signal.c:1806
	arch/x86/kernel/signal.c:805
	arch/x86/kernel/signal.c:871
	arch/x86/kernel/entry_64.S:694

Remember that since these addresses were found on the stack and not as the
RIP value, they actually point to the _next_ instruction (they are return
addresses). This becomes obvious when we look at the code for line 446::

    422 int dequeue_signal(struct task_struct *tsk, sigset_t *mask, siginfo_t *info)
    423 {
    ...
    431                 signr = __dequeue_signal(&tsk->signal->shared_pending,
    432						 mask, info);
    433			/*
    434			 * itimer signal ?
    435			 *
    436			 * itimers are process shared and we restart periodic
    437			 * itimers in the signal delivery path to prevent DoS
    438			 * attacks in the high resolution timer case. This is
    439			 * compliant with the old way of self restarting
    440			 * itimers, as the SIGALRM is a legacy signal and only
    441			 * queued once. Changing the restart behaviour to
    442			 * restart the timer in the signal dequeue path is
    443			 * reducing the timer noise on heavy loaded !highres
    444			 * systems too.
    445			 */
    446			if (unlikely(signr == SIGALRM)) {
    ...
    489 }

So instead of looking at 446, we should be looking at 431, which is the line
that executes just before 446. Here we see that what we are looking for is
``&tsk->signal->shared_pending``.

Our next task is now to figure out which function that puts items on this
``shared_pending`` list. A crude, but efficient tool, is ``git grep``::

	$ git grep -n 'shared_pending' kernel/
	...
	kernel/signal.c:828:	pending = group ? &t->signal->shared_pending : &t->pending;
	kernel/signal.c:1339:	pending = group ? &t->signal->shared_pending : &t->pending;
	...

There were more results, but none of them were related to list operations,
and these were the only assignments. We inspect the line numbers more closely
and find that this is indeed where items are being added to the list::

    816 static int send_signal(int sig, struct siginfo *info, struct task_struct *t,
    817				int group)
    818 {
    ...
    828		pending = group ? &t->signal->shared_pending : &t->pending;
    ...
    851		q = __sigqueue_alloc(t, GFP_ATOMIC, (sig < SIGRTMIN &&
    852						     (is_si_special(info) ||
    853						      info->si_code >= 0)));
    854		if (q) {
    855			list_add_tail(&q->list, &pending->list);
    ...
    890 }

and::

    1309 int send_sigqueue(struct sigqueue *q, struct task_struct *t, int group)
    1310 {
    ....
    1339	 pending = group ? &t->signal->shared_pending : &t->pending;
    1340	 list_add_tail(&q->list, &pending->list);
    ....
    1347 }

In the first case, the list element we are looking for, ``q``, is being
returned from the function ``__sigqueue_alloc()``, which looks like an
allocation function.  Let's take a look at it::

    187 static struct sigqueue *__sigqueue_alloc(struct task_struct *t, gfp_t flags,
    188						 int override_rlimit)
    189 {
    190		struct sigqueue *q = NULL;
    191		struct user_struct *user;
    192
    193		/*
    194		 * We won't get problems with the target's UID changing under us
    195		 * because changing it requires RCU be used, and if t != current, the
    196		 * caller must be holding the RCU readlock (by way of a spinlock) and
    197		 * we use RCU protection here
    198		 */
    199		user = get_uid(__task_cred(t)->user);
    200		atomic_inc(&user->sigpending);
    201		if (override_rlimit ||
    202		    atomic_read(&user->sigpending) <=
    203				t->signal->rlim[RLIMIT_SIGPENDING].rlim_cur)
    204			q = kmem_cache_alloc(sigqueue_cachep, flags);
    205		if (unlikely(q == NULL)) {
    206			atomic_dec(&user->sigpending);
    207			free_uid(user);
    208		} else {
    209			INIT_LIST_HEAD(&q->list);
    210			q->flags = 0;
    211			q->user = user;
    212		}
    213
    214		return q;
    215 }

We see that this function initializes ``q->list``, ``q->flags``, and
``q->user``. It seems that now is the time to look at the definition of
``struct sigqueue``, e.g.::

    14 struct sigqueue {
    15	       struct list_head list;
    16	       int flags;
    17	       siginfo_t info;
    18	       struct user_struct *user;
    19 };

And, you might remember, it was a ``memcpy()`` on ``&first->info`` that
caused the warning, so this makes perfect sense. It also seems reasonable
to assume that it is the caller of ``__sigqueue_alloc()`` that has the
responsibility of filling out (initializing) this member.

But just which fields of the struct were uninitialized? Let's look at
kmemcheck's report again::

    WARNING: kmemcheck: Caught 32-bit read from uninitialized memory (ffff88003e4a2024)
    80000000000000000000000000000000000000000088ffff0000000000000000
     i i i i u u u u i i i i i i i i u u u u u u u u u u u u u u u u
	     ^

These first two lines are the memory dump of the memory object itself, and
the shadow bytemap, respectively. The memory object itself is in this case
``&first->info``. Just beware that the start of this dump is NOT the start
of the object itself! The position of the caret (^) corresponds with the
address of the read (ffff88003e4a2024).

The shadow bytemap dump legend is as follows:

- i: initialized
- u: uninitialized
- a: unallocated (memory has been allocated by the slab layer, but has not
  yet been handed off to anybody)
- f: freed (memory has been allocated by the slab layer, but has been freed
  by the previous owner)

In order to figure out where (relative to the start of the object) the
uninitialized memory was located, we have to look at the disassembly. For
that, we'll need the RIP address again::

    RIP: 0010:[<ffffffff8104ede8>]  [<ffffffff8104ede8>] __dequeue_signal+0xc8/0x190

	$ objdump -d --no-show-raw-insn vmlinux | grep -C 8 ffffffff8104ede8:
	ffffffff8104edc8:	mov    %r8,0x8(%r8)
	ffffffff8104edcc:	test   %r10d,%r10d
	ffffffff8104edcf:	js     ffffffff8104ee88 <__dequeue_signal+0x168>
	ffffffff8104edd5:	mov    %rax,%rdx
	ffffffff8104edd8:	mov    $0xc,%ecx
	ffffffff8104eddd:	mov    %r13,%rdi
	ffffffff8104ede0:	mov    $0x30,%eax
	ffffffff8104ede5:	mov    %rdx,%rsi
	ffffffff8104ede8:	rep movsl %ds:(%rsi),%es:(%rdi)
	ffffffff8104edea:	test   $0x2,%al
	ffffffff8104edec:	je     ffffffff8104edf0 <__dequeue_signal+0xd0>
	ffffffff8104edee:	movsw  %ds:(%rsi),%es:(%rdi)
	ffffffff8104edf0:	test   $0x1,%al
	ffffffff8104edf2:	je     ffffffff8104edf5 <__dequeue_signal+0xd5>
	ffffffff8104edf4:	movsb  %ds:(%rsi),%es:(%rdi)
	ffffffff8104edf5:	mov    %r8,%rdi
	ffffffff8104edf8:	callq  ffffffff8104de60 <__sigqueue_free>

As expected, it's the "``rep movsl``" instruction from the ``memcpy()``
that causes the warning. We know about ``REP MOVSL`` that it uses the register
``RCX`` to count the number of remaining iterations. By taking a look at the
register dump again (from the kmemcheck report), we can figure out how many
bytes were left to copy::

    RAX: 0000000000000030 RBX: ffff88003d4ea968 RCX: 0000000000000009

By looking at the disassembly, we also see that ``%ecx`` is being loaded
with the value ``$0xc`` just before (ffffffff8104edd8), so we are very
lucky. Keep in mind that this is the number of iterations, not bytes. And
since this is a "long" operation, we need to multiply by 4 to get the
number of bytes. So this means that the uninitialized value was encountered
at 4 * (0xc - 0x9) = 12 bytes from the start of the object.

We can now try to figure out which field of the "``struct siginfo``" that
was not initialized. This is the beginning of the struct::

    40 typedef struct siginfo {
    41	       int si_signo;
    42	       int si_errno;
    43	       int si_code;
    44
    45	       union {
    ..
    92	       } _sifields;
    93 } siginfo_t;

On 64-bit, the int is 4 bytes long, so it must the union member that has
not been initialized. We can verify this using gdb::

	$ gdb vmlinux
	...
	(gdb) p &((struct siginfo *) 0)->_sifields
	$1 = (union {...} *) 0x10

Actually, it seems that the union member is located at offset 0x10 -- which
means that gcc has inserted 4 bytes of padding between the members ``si_code``
and ``_sifields``. We can now get a fuller picture of the memory dump::

		 _----------------------------=> si_code
		/	 _--------------------=> (padding)
	       |	/	 _------------=> _sifields(._kill._pid)
	       |       |	/	 _----=> _sifields(._kill._uid)
	       |       |       |	/
	-------|-------|-------|-------|
	80000000000000000000000000000000000000000088ffff0000000000000000
	 i i i i u u u u i i i i i i i i u u u u u u u u u u u u u u u u

This allows us to realize another important fact: ``si_code`` contains the
value 0x80. Remember that x86 is little endian, so the first 4 bytes
"80000000" are really the number 0x00000080. With a bit of research, we
find that this is actually the constant ``SI_KERNEL`` defined in
``include/asm-generic/siginfo.h``::

    144 #define SI_KERNEL	0x80		/* sent by the kernel from somewhere	 */

This macro is used in exactly one place in the x86 kernel: In ``send_signal()``
in ``kernel/signal.c``::

    816 static int send_signal(int sig, struct siginfo *info, struct task_struct *t,
    817				int group)
    818 {
    ...
    828		pending = group ? &t->signal->shared_pending : &t->pending;
    ...
    851		q = __sigqueue_alloc(t, GFP_ATOMIC, (sig < SIGRTMIN &&
    852						     (is_si_special(info) ||
    853						      info->si_code >= 0)));
    854		if (q) {
    855			list_add_tail(&q->list, &pending->list);
    856			switch ((unsigned long) info) {
    ...
    865			case (unsigned long) SEND_SIG_PRIV:
    866				q->info.si_signo = sig;
    867				q->info.si_errno = 0;
    868				q->info.si_code = SI_KERNEL;
    869				q->info.si_pid = 0;
    870				q->info.si_uid = 0;
    871				break;
    ...
    890 }

Not only does this match with the ``.si_code`` member, it also matches the place
we found earlier when looking for where siginfo_t objects are enqueued on the
``shared_pending`` list.

So to sum up: It seems that it is the padding introduced by the compiler
between two struct fields that is uninitialized, and this gets reported when
we do a ``memcpy()`` on the struct. This means that we have identified a false
positive warning.

Normally, kmemcheck will not report uninitialized accesses in ``memcpy()`` calls
when both the source and destination addresses are tracked. (Instead, we copy
the shadow bytemap as well). In this case, the destination address clearly
was not tracked. We can dig a little deeper into the stack trace from above::

	arch/x86/kernel/signal.c:805
	arch/x86/kernel/signal.c:871
	arch/x86/kernel/entry_64.S:694

And we clearly see that the destination siginfo object is located on the
stack::

    782 static void do_signal(struct pt_regs *regs)
    783 {
    784		struct k_sigaction ka;
    785		siginfo_t info;
    ...
    804		signr = get_signal_to_deliver(&info, &ka, regs, NULL);
    ...
    854 }

And this ``&info`` is what eventually gets passed to ``copy_siginfo()`` as the
destination argument.

Now, even though we didn't find an actual error here, the example is still a
good one, because it shows how one would go about to find out what the report
was all about.


Annotating false positives
~~~~~~~~~~~~~~~~~~~~~~~~~~

There are a few different ways to make annotations in the source code that
will keep kmemcheck from checking and reporting certain allocations. Here
they are:

- ``__GFP_NOTRACK_FALSE_POSITIVE``
	This flag can be passed to ``kmalloc()`` or ``kmem_cache_alloc()``
	(therefore also to other functions that end up calling one of
	these) to indicate that the allocation should not be tracked
	because it would lead to a false positive report. This is a "big
	hammer" way of silencing kmemcheck; after all, even if the false
	positive pertains to particular field in a struct, for example, we
	will now lose the ability to find (real) errors in other parts of
	the same struct.

	Example::

	    /* No warnings will ever trigger on accessing any part of x */
	    x = kmalloc(sizeof *x, GFP_KERNEL | __GFP_NOTRACK_FALSE_POSITIVE);

- ``kmemcheck_bitfield_begin(name)``/``kmemcheck_bitfield_end(name)`` and
	``kmemcheck_annotate_bitfield(ptr, name)``
	The first two of these three macros can be used inside struct
	definitions to signal, respectively, the beginning and end of a
	bitfield. Additionally, this will assign the bitfield a name, which
	is given as an argument to the macros.

	Having used these markers, one can later use
	kmemcheck_annotate_bitfield() at the point of allocation, to indicate
	which parts of the allocation is part of a bitfield.

	Example::

	    struct foo {
		int x;

		kmemcheck_bitfield_begin(flags);
		int flag_a:1;
		int flag_b:1;
		kmemcheck_bitfield_end(flags);

		int y;
	    };

	    struct foo *x = kmalloc(sizeof *x);

	    /* No warnings will trigger on accessing the bitfield of x */
	    kmemcheck_annotate_bitfield(x, flags);

	Note that ``kmemcheck_annotate_bitfield()`` can be used even before the
	return value of ``kmalloc()`` is checked -- in other words, passing NULL
	as the first argument is legal (and will do nothing).


Reporting errors
----------------

As we have seen, kmemcheck will produce false positive reports. Therefore, it
is not very wise to blindly post kmemcheck warnings to mailing lists and
maintainers. Instead, I encourage maintainers and developers to find errors
in their own code. If you get a warning, you can try to work around it, try
to figure out if it's a real error or not, or simply ignore it. Most
developers know their own code and will quickly and efficiently determine the
root cause of a kmemcheck report. This is therefore also the most efficient
way to work with kmemcheck.

That said, we (the kmemcheck maintainers) will always be on the lookout for
false positives that we can annotate and silence. So whatever you find,
please drop us a note privately! Kernel configs and steps to reproduce (if
available) are of course a great help too.

Happy hacking!


Technical description
---------------------

kmemcheck works by marking memory pages non-present. This means that whenever
somebody attempts to access the page, a page fault is generated. The page
fault handler notices that the page was in fact only hidden, and so it calls
on the kmemcheck code to make further investigations.

When the investigations are completed, kmemcheck "shows" the page by marking
it present (as it would be under normal circumstances). This way, the
interrupted code can continue as usual.

But after the instruction has been executed, we should hide the page again, so
that we can catch the next access too! Now kmemcheck makes use of a debugging
feature of the processor, namely single-stepping. When the processor has
finished the one instruction that generated the memory access, a debug
exception is raised. From here, we simply hide the page again and continue
execution, this time with the single-stepping feature turned off.

kmemcheck requires some assistance from the memory allocator in order to work.
The memory allocator needs to

  1. Tell kmemcheck about newly allocated pages and pages that are about to
     be freed. This allows kmemcheck to set up and tear down the shadow memory
     for the pages in question. The shadow memory stores the status of each
     byte in the allocation proper, e.g. whether it is initialized or
     uninitialized.

  2. Tell kmemcheck which parts of memory should be marked uninitialized.
     There are actually a few more states, such as "not yet allocated" and
     "recently freed".

If a slab cache is set up using the SLAB_NOTRACK flag, it will never return
memory that can take page faults because of kmemcheck.

If a slab cache is NOT set up using the SLAB_NOTRACK flag, callers can still
request memory with the __GFP_NOTRACK or __GFP_NOTRACK_FALSE_POSITIVE flags.
This does not prevent the page faults from occurring, however, but marks the
object in question as being initialized so that no warnings will ever be
produced for this object.

Currently, the SLAB and SLUB allocators are supported by kmemcheck.