Merge tag 'for-5.20/block-2022-07-29' of git://git.kernel.dk/linux-block

[linux-block.git] / Documentation / process / deprecated.rst

.. SPDX-License-Identifier: GPL-2.0

.. _deprecated:

=====================================================================
Deprecated Interfaces, Language Features, Attributes, and Conventions
=====================================================================

In a perfect world, it would be possible to convert all instances of
some deprecated API into the new API and entirely remove the old API in
a single development cycle. However, due to the size of the kernel, the
maintainership hierarchy, and timing, it's not always feasible to do these
kinds of conversions at once. This means that new instances may sneak into
the kernel while old ones are being removed, only making the amount of
work to remove the API grow. In order to educate developers about what
has been deprecated and why, this list has been created as a place to
point when uses of deprecated things are proposed for inclusion in the
kernel.

__deprecated
------------
While this attribute does visually mark an interface as deprecated,
it `does not produce warnings during builds any more
<https://git.kernel.org/linus/771c035372a036f83353eef46dbb829780330234>`_
because one of the standing goals of the kernel is to build without
warnings and no one was actually doing anything to remove these deprecated
interfaces. While using `__deprecated` is nice to note an old API in
a header file, it isn't the full solution. Such interfaces must either
be fully removed from the kernel, or added to this file to discourage
others from using them in the future.

BUG() and BUG_ON()
------------------
Use WARN() and WARN_ON() instead, and handle the "impossible"
error condition as gracefully as possible. While the BUG()-family
of APIs were originally designed to act as an "impossible situation"
assert and to kill a kernel thread "safely", they turn out to just be
too risky. (e.g. "In what order do locks need to be released? Have
various states been restored?") Very commonly, using BUG() will
destabilize a system or entirely break it, which makes it impossible
to debug or even get viable crash reports. Linus has `very strong
<https://lore.kernel.org/lkml/CA+55aFy6jNLsywVYdGp83AMrXBo_P-pkjkphPGrO=82SPKCpLQ@mail.gmail.com/>`_
feelings `about this
<https://lore.kernel.org/lkml/CAHk-=whDHsbK3HTOpTF=ue_o04onRwTEaK_ZoJp_fjbqq4+=Jw@mail.gmail.com/>`_.

Note that the WARN()-family should only be used for "expected to
be unreachable" situations. If you want to warn about "reachable
but undesirable" situations, please use the pr_warn()-family of
functions. System owners may have set the *panic_on_warn* sysctl,
to make sure their systems do not continue running in the face of
"unreachable" conditions. (For example, see commits like `this one
<https://git.kernel.org/linus/d4689846881d160a4d12a514e991a740bcb5d65a>`_.)

open-coded arithmetic in allocator arguments
--------------------------------------------
Dynamic size calculations (especially multiplication) should not be
performed in memory allocator (or similar) function arguments due to the
risk of them overflowing. This could lead to values wrapping around and a
smaller allocation being made than the caller was expecting. Using those
allocations could lead to linear overflows of heap memory and other
misbehaviors. (One exception to this is literal values where the compiler
can warn if they might overflow. However, the preferred way in these
cases is to refactor the code as suggested below to avoid the open-coded
arithmetic.)

For example, do not use ``count * size`` as an argument, as in::

	foo = kmalloc(count * size, GFP_KERNEL);

Instead, the 2-factor form of the allocator should be used::

	foo = kmalloc_array(count, size, GFP_KERNEL);

Specifically, kmalloc() can be replaced with kmalloc_array(), and
kzalloc() can be replaced with kcalloc().

If no 2-factor form is available, the saturate-on-overflow helpers should
be used::

	bar = vmalloc(array_size(count, size));

Another common case to avoid is calculating the size of a structure with
a trailing array of others structures, as in::

	header = kzalloc(sizeof(*header) + count * sizeof(*header->item),
			 GFP_KERNEL);

Instead, use the helper::

	header = kzalloc(struct_size(header, item, count), GFP_KERNEL);

.. note:: If you are using struct_size() on a structure containing a zero-length
        or a one-element array as a trailing array member, please refactor such
        array usage and switch to a `flexible array member
        <#zero-length-and-one-element-arrays>`_ instead.

For other calculations, please compose the use of the size_mul(),
size_add(), and size_sub() helpers. For example, in the case of::

	foo = krealloc(current_size + chunk_size * (count - 3), GFP_KERNEL);

Instead, use the helpers::

	foo = krealloc(size_add(current_size,
				size_mul(chunk_size,
					 size_sub(count, 3))), GFP_KERNEL);

For more details, also see array3_size() and flex_array_size(),
as well as the related check_mul_overflow(), check_add_overflow(),
check_sub_overflow(), and check_shl_overflow() family of functions.

simple_strtol(), simple_strtoll(), simple_strtoul(), simple_strtoull()
----------------------------------------------------------------------
The simple_strtol(), simple_strtoll(),
simple_strtoul(), and simple_strtoull() functions
explicitly ignore overflows, which may lead to unexpected results
in callers. The respective kstrtol(), kstrtoll(),
kstrtoul(), and kstrtoull() functions tend to be the
correct replacements, though note that those require the string to be
NUL or newline terminated.

strcpy()
--------
strcpy() performs no bounds checking on the destination buffer. This
could result in linear overflows beyond the end of the buffer, leading to
all kinds of misbehaviors. While `CONFIG_FORTIFY_SOURCE=y` and various
compiler flags help reduce the risk of using this function, there is
no good reason to add new uses of this function. The safe replacement
is strscpy(), though care must be given to any cases where the return
value of strcpy() was used, since strscpy() does not return a pointer to
the destination, but rather a count of non-NUL bytes copied (or negative
errno when it truncates).

strncpy() on NUL-terminated strings
-----------------------------------
Use of strncpy() does not guarantee that the destination buffer will
be NUL terminated. This can lead to various linear read overflows and
other misbehavior due to the missing termination. It also NUL-pads
the destination buffer if the source contents are shorter than the
destination buffer size, which may be a needless performance penalty
for callers using only NUL-terminated strings. The safe replacement is
strscpy(), though care must be given to any cases where the return value
of strncpy() was used, since strscpy() does not return a pointer to the
destination, but rather a count of non-NUL bytes copied (or negative
errno when it truncates). Any cases still needing NUL-padding should
instead use strscpy_pad().

If a caller is using non-NUL-terminated strings, strncpy() can
still be used, but destinations should be marked with the `__nonstring
<https://gcc.gnu.org/onlinedocs/gcc/Common-Variable-Attributes.html>`_
attribute to avoid future compiler warnings.

strlcpy()
---------
strlcpy() reads the entire source buffer first (since the return value
is meant to match that of strlen()). This read may exceed the destination
size limit. This is both inefficient and can lead to linear read overflows
if a source string is not NUL-terminated. The safe replacement is strscpy(),
though care must be given to any cases where the return value of strlcpy()
is used, since strscpy() will return negative errno values when it truncates.

%p format specifier
-------------------
Traditionally, using "%p" in format strings would lead to regular address
exposure flaws in dmesg, proc, sysfs, etc. Instead of leaving these to
be exploitable, all "%p" uses in the kernel are being printed as a hashed
value, rendering them unusable for addressing. New uses of "%p" should not
be added to the kernel. For text addresses, using "%pS" is likely better,
as it produces the more useful symbol name instead. For nearly everything
else, just do not add "%p" at all.

Paraphrasing Linus's current `guidance <https://lore.kernel.org/lkml/CA+55aFwQEd_d40g4mUCSsVRZzrFPUJt74vc6PPpb675hYNXcKw@mail.gmail.com/>`_:

- If the hashed "%p" value is pointless, ask yourself whether the pointer
  itself is important. Maybe it should be removed entirely?
- If you really think the true pointer value is important, why is some
  system state or user privilege level considered "special"? If you think
  you can justify it (in comments and commit log) well enough to stand
  up to Linus's scrutiny, maybe you can use "%px", along with making sure
  you have sensible permissions.

If you are debugging something where "%p" hashing is causing problems,
you can temporarily boot with the debug flag "`no_hash_pointers
<https://git.kernel.org/linus/5ead723a20e0447bc7db33dc3070b420e5f80aa6>`_".

Variable Length Arrays (VLAs)
-----------------------------
Using stack VLAs produces much worse machine code than statically
sized stack arrays. While these non-trivial `performance issues
<https://git.kernel.org/linus/02361bc77888>`_ are reason enough to
eliminate VLAs, they are also a security risk. Dynamic growth of a stack
array may exceed the remaining memory in the stack segment. This could
lead to a crash, possible overwriting sensitive contents at the end of the
stack (when built without `CONFIG_THREAD_INFO_IN_TASK=y`), or overwriting
memory adjacent to the stack (when built without `CONFIG_VMAP_STACK=y`)

Implicit switch case fall-through
---------------------------------
The C language allows switch cases to fall through to the next case
when a "break" statement is missing at the end of a case. This, however,
introduces ambiguity in the code, as it's not always clear if the missing
break is intentional or a bug. For example, it's not obvious just from
looking at the code if `STATE_ONE` is intentionally designed to fall
through into `STATE_TWO`::

	switch (value) {
	case STATE_ONE:
		do_something();
	case STATE_TWO:
		do_other();
		break;
	default:
		WARN("unknown state");
	}

As there have been a long list of flaws `due to missing "break" statements
<https://cwe.mitre.org/data/definitions/484.html>`_, we no longer allow
implicit fall-through. In order to identify intentional fall-through
cases, we have adopted a pseudo-keyword macro "fallthrough" which
expands to gcc's extension `__attribute__((__fallthrough__))
<https://gcc.gnu.org/onlinedocs/gcc/Statement-Attributes.html>`_.
(When the C17/C18  `[[fallthrough]]` syntax is more commonly supported by
C compilers, static analyzers, and IDEs, we can switch to using that syntax
for the macro pseudo-keyword.)

All switch/case blocks must end in one of:

* break;
* fallthrough;
* continue;
* goto <label>;
* return [expression];

Zero-length and one-element arrays
----------------------------------
There is a regular need in the kernel to provide a way to declare having
a dynamically sized set of trailing elements in a structure. Kernel code
should always use `"flexible array members" <https://en.wikipedia.org/wiki/Flexible_array_member>`_
for these cases. The older style of one-element or zero-length arrays should
no longer be used.

In older C code, dynamically sized trailing elements were done by specifying
a one-element array at the end of a structure::

        struct something {
                size_t count;
                struct foo items[1];
        };

This led to fragile size calculations via sizeof() (which would need to
remove the size of the single trailing element to get a correct size of
the "header"). A `GNU C extension <https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html>`_
was introduced to allow for zero-length arrays, to avoid these kinds of
size problems::

        struct something {
                size_t count;
                struct foo items[0];
        };

But this led to other problems, and didn't solve some problems shared by
both styles, like not being able to detect when such an array is accidentally
being used _not_ at the end of a structure (which could happen directly, or
when such a struct was in unions, structs of structs, etc).

C99 introduced "flexible array members", which lacks a numeric size for
the array declaration entirely::

        struct something {
                size_t count;
                struct foo items[];
        };

This is the way the kernel expects dynamically sized trailing elements
to be declared. It allows the compiler to generate errors when the
flexible array does not occur last in the structure, which helps to prevent
some kind of `undefined behavior
<https://git.kernel.org/linus/76497732932f15e7323dc805e8ea8dc11bb587cf>`_
bugs from being inadvertently introduced to the codebase. It also allows
the compiler to correctly analyze array sizes (via sizeof(),
`CONFIG_FORTIFY_SOURCE`, and `CONFIG_UBSAN_BOUNDS`). For instance,
there is no mechanism that warns us that the following application of the
sizeof() operator to a zero-length array always results in zero::

        struct something {
                size_t count;
                struct foo items[0];
        };

        struct something *instance;

        instance = kmalloc(struct_size(instance, items, count), GFP_KERNEL);
        instance->count = count;

        size = sizeof(instance->items) * instance->count;
        memcpy(instance->items, source, size);

At the last line of code above, ``size`` turns out to be ``zero``, when one might
have thought it represents the total size in bytes of the dynamic memory recently
allocated for the trailing array ``items``. Here are a couple examples of this
issue: `link 1
<https://git.kernel.org/linus/f2cd32a443da694ac4e28fbf4ac6f9d5cc63a539>`_,
`link 2
<https://git.kernel.org/linus/ab91c2a89f86be2898cee208d492816ec238b2cf>`_.
Instead, `flexible array members have incomplete type, and so the sizeof()
operator may not be applied <https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html>`_,
so any misuse of such operators will be immediately noticed at build time.

With respect to one-element arrays, one has to be acutely aware that `such arrays
occupy at least as much space as a single object of the type
<https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html>`_,
hence they contribute to the size of the enclosing structure. This is prone
to error every time people want to calculate the total size of dynamic memory
to allocate for a structure containing an array of this kind as a member::

        struct something {
                size_t count;
                struct foo items[1];
        };

        struct something *instance;

        instance = kmalloc(struct_size(instance, items, count - 1), GFP_KERNEL);
        instance->count = count;

        size = sizeof(instance->items) * instance->count;
        memcpy(instance->items, source, size);

In the example above, we had to remember to calculate ``count - 1`` when using
the struct_size() helper, otherwise we would have --unintentionally-- allocated
memory for one too many ``items`` objects. The cleanest and least error-prone way
to implement this is through the use of a `flexible array member`, together with
struct_size() and flex_array_size() helpers::

        struct something {
                size_t count;
                struct foo items[];
        };

        struct something *instance;

        instance = kmalloc(struct_size(instance, items, count), GFP_KERNEL);
        instance->count = count;

        memcpy(instance->items, source, flex_array_size(instance, items, instance->count));