xref: /illumos-gate/usr/src/lib/libc/README (revision cb159270e9e67dea31dfbd991ec758829de14d35)

#
# CDDL HEADER START
#
# The contents of this file are subject to the terms of the
# Common Development and Distribution License (the "License").
# You may not use this file except in compliance with the License.
#
# You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
# or http://www.opensolaris.org/os/licensing.
# See the License for the specific language governing permissions
# and limitations under the License.
#
# When distributing Covered Code, include this CDDL HEADER in each
# file and include the License file at usr/src/OPENSOLARIS.LICENSE.
# If applicable, add the following below this CDDL HEADER, with the
# fields enclosed by brackets "[]" replaced with your own identifying
# information: Portions Copyright [yyyy] [name of copyright owner]
#
# CDDL HEADER END
#
#
# Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
# Use is subject to license terms.
#

# Process Model Unification

The Solaris Process Model Unification project:
	PSARC/2002/117 Solaris Process Model Unification
	4470917 Solaris Process Model Unification
folded libthread into libc and has led to some fundamental changes
in the rules by which code in libc must be developed and maintained.

All code in libc must be both MT-Safe and Fork-Safe
and where possible (almost everywhere), Async-Signal-Safe.

To this end, the following rules should be followed:

Almost all internal libc locks (mutexes and read-write locks)
should be acquired and released via these interfaces:

	mutex_t some_lock = DEFAULTMUTEX;

	lmutex_lock(&some_lock);
	... do something critical ...
	lmutex_unlock(&some_lock);

	rwlock_t some_rw_lock = DEFAULTRWLOCK;

	lrw_rdlock(&some_rw_lock);
	... multiple threads can do something ...
	lrw_unlock(&some_rw_lock);

	lrw_wrlock(&some_rw_lock);
	... only one thread can do something ...
	lrw_unlock(&some_rw_lock);

The above l* versions of the mutex and rwlock interfaces do more
than the ordinary interfaces:  They define critical regions in
which the calling thread cannot be suspended (making the region
fork-safe) and in which the calling thread has all signals deferred
(making the region async-signal-safe).

However, certain rules apply to the code within these critical regions:

	- The code must be of guaranteed short duration; no calls
	  to interfaces that might block indefinitely are allowed.
	  This means no calls into stdio or syslog() and no calls
	  to cond_wait() unless there is a guarantee of an almost-
	  immediate call to cond_signal() or cond_broadcast()
	  from elsewhere.

	- The code cannot call any non-l* synchronization
	  primitives (mutex_lock(), _private_mutex_lock(),
	  rw_wrlock(), rw_rdlock(), sema_wait(), etc.)

	- The code cannot call any functions outside of libc,
	  including application callbacks and functions from
	  dlopen()ed objects, such as those in the I18N code.

	- Because malloc(), calloc(), realloc(), and free()
	  are designed to be interposed upon, they fall into
	  the previous case of prohibition.  None of these can
	  be called by a thread while in a critical region.

There is a private memory allocator for use internally to libc.
It cannot be interposed upon and it is safe to use while in
a critical region (or for that matter while not in a critical
region; it is async-signal-safe and fork-safe):

	void *lmalloc(size_t);
	void lfree(void *, size_t);

	void *libc_malloc(size_t);
	void *libc_realloc(void *, size_t);
	char *libc_strdup(const char *);
	void libc_free(void *);

lmalloc() and lfree() are the basic interfaces.  The libc_*()
variants are built on top of lmalloc()/lfree() but they have
the same interface signatures as the corresponding functions
without the 'libc_' prefix.  lmalloc() and libc_malloc()
return zeroed memory blocks.  Note that lmalloc()/lfree()
require the caller to remember the size parameter passed
to lmalloc() and to pass the same value to lfree().

Memory allocated by lmalloc() can only be freed by lfree().
Memory allocated by libc_malloc(), libc_realloc(), or libc_strdup()
can only be freed by libc_free().  Never pass such allocated
memory out of libc if the caller of libc is expected to free it.

lmalloc()/lfree() is a small and simple power of two allocator.
Do not use it as a general-purpose allocator.  Be kind to it.

There is a special mutual exclusion interface that exists for
cases, like code in the I18N interfaces, where mutual exclusion
is required but the above rules cannot be followed:

	int fork_lock_enter(const char *);
	void fork_lock_exit(void);

fork_lock_enter() does triple-duty.  Not only does it serialize
calls to fork() and forkall(), but it also serializes calls to
thr_suspend() (fork() and forkall() also suspend other threads),
and furthermore it serializes I18N calls to functions in other
dlopen()ed L10N objects that might be calling malloc()/free().
Use it in general like this:

	(void) fork_lock_enter(NULL);
	... serialized; do something that might call malloc ...
	fork_lock_exit();

The 'const char *' argument to fork_lock_enter() should always
be NULL except for two special cases:
	- When called from fork() or forkall()
	- When called from pthread_atfork()
This enforces the prohibition against calling fork() or pthread_atfork()
from a pthread_atfork()-registered fork handler function while a fork()
prologue or epilogue is in progress.  If _THREAD_ERROR_DETECTION is set
to 1 or 2 in the environment, such cases will draw a nasty message and
will dump core if _THREAD_ERROR_DETECTION=2.  fork_lock_enter() returns
non-zero only if it is called from a fork handler.  This is of interest
only to callers that have to do something about this condition; the
return value should be ignored in all other cases (fork_lock_enter()
never actually fails).

It is an error to call fork_lock_enter() while in a critical region
(that is, while holding any internal libc lock).

On return from fork_lock_enter(), no internal libc locks are held
but a flag has been set to cause other callers of fork_lock_enter()
to delay (via _cond_wait()) until fork_lock_exit() is called.

These are the rules to follow for memory allocation:

  - If a function acquires an internal libc lock or is called while
    an internal libc lock is held:

	* The malloc family cannot be used.

	* lmalloc or libc_malloc should be used.  The memory must
	  be released by lfree or libc_free, respectively.

	* lfree takes an argument to tell the size of the releasing
	  memory.  If the function does not know the size at the
	  releasing point, libc_malloc and libc_free should be used.

	* As the memory allocated by lmalloc or libc_malloc needs
	  to be released by lfree or libc_free and these are internal
	  to libc, they cannot be used to allocate memory that might
	  be released by application code outside libc.

	* If the memory allocation by malloc() cannot be avoided and
	  the scalability of the function does not matter much, the
	  function can be serialized with fork_lock_enter() instead
	  of lmutex_lock().

	* If the memory allocation by malloc() cannot be avoided and
	  the scalability of the function does matter, another
	  implementation of the function will be necessary.

In a DEBUG build of libc:
	make THREAD_DEBUG=-DTHREAD_DEBUG install
many of these rules are enforced by ASSERT() statements scattered about
in the libc sources.  This is the default mode for building libc when
a DEBUG nightly build is performed.

# libaio/librt Implementation In libc

The putback of the project:
	6416832 libaio and librt can and should be folded into libc
introduced several libc-private locking interfaces:
	void	sig_mutex_lock(mutex_t *);
	void	sig_mutex_unlock(mutex_t *);
	int	sig_mutex_trylock(mutex_t *);
	int	sig_cond_wait(cond_t *, mutex_t *);
	int	sig_cond_reltimedwait(cond_t *, mutex_t *, const timespec_t *);
which are declared in both "thr_uberdata.h" and "mtlib.h".

They are used in specialized code in libc, like the asynchronous i/o code.
Unlike the lmutex_lock() and lmutex_unlock() interfaces described above,
these interfaces do not define critical regions, but signals are
deferred while locks acquired by these functions are held, making
their use be async-signal safe.  Calls to malloc(), calloc(), realloc(),
and free() are permissible while holding such locks.

These interfaces were brought over from code in the former libaio
and librt and are necessary because, where they are used, the code
must execute potentially long-term waits and must be cancelable.
sig_cond_wait() and sig_cond_reltimedwait() are cancellation points.

These interfaces are available for other uses inside libc, as
the need arises.  (There is no need if the code does not perform
long-term waits.)  Just follow a few rules to be self-consistent:
 - Don't mix calls to mutex_[un]lock(), lmutex_[un]lock() and
   sig_mutex_[un]lock() on the same mutex.
 - Don't call cond_wait() with a mutex acquired by sig_mutex_lock();
   call sig_cond_wait() or sig_cond_reltimedwait().
 - Use pthread_cleanup_push() and pthread_cleanup_pop() to make
   your code cancellation-safe.
 - The sig_*() interfaces are not in themselves fork-safe.
   You have to employ other logic to make your code fork-safe.
   See the tail of postfork1_child() for examples.

# Removal Of Synonym Symbols

The following project:
    PSARC 2008/309 expunge synonyms.h
    6700179 expunge synonyms.h

Removed the historical "synonym" symbols from the C Library.

Historically, for every public function symbol in the C library a second,
private, symbol of the same value was defined to be used internally by libc
(generally, one or the other will be a weak symbol, precisely which is
inconsistent).

These synonym symbols existed such that an application which provided
otherwise conflicting symbols could interpose on the version in libc without
compromising libc itself, that is if libc defines fopen() which needs open() it
would call _open() and an application defined open() would not cause fopen()
to break.  This was made transparent to code within libc via a header,
synonyms.h, which would #define open _open, for all such symbols.

Since ON now uses direct bindings extensively all symbols not explicitly
marked "NODIRECT" are directly bound within libc anyway, and none of this is
actually necessary.  Thus synonyms.h was removed, and no new synonym symbols
need be added.  However, unfortunately, certain of the private symbols were
inadvertently exposed to applications, and several are known to have been
used, thus these existing synonyms must continue to exist to maintain
compatibility.  A preloadable library, /lib/c_synonyms.so.1 is provided which
also provides the historical names with underscore prefixes to allow any other
incorrect application to continue to function.

It should never be necessary to add additional synonym symbols to libc nor to
add underscore prefixed aliases to c_synonyms.so.1.

# libc Internals Scoped Protected

The integration of the fix for:
	6689238 libc needs global protection against ld.so.1

Scopes all function symbols within libc protected, excepting those that need
to be accepting of interposition and to have a consistent version called both
within and without libc (basically, the malloc() family).

This causes references by libc to itself to be permanently and unavoidably
bound, and thus to never enter ld.so (and potentially from there audit
libraries or other such support libraries).  This maintains an otherwise
complicated to verify invariant: within critical sections (with any internal
lock held, etc) execution can never leave the context of libc.  Previously
this was done with a selection of known-to-be-problematic functions having
weak synonyms scoped private, but this was both difficult to verify, difficult
to remember, and thus always at least somewhat incomplete.

In summary, any new function symbol in libc must be scoped protected unless it
is one of a very small group of functions that must allow interposed versions
to be bound to from the C library itself -- it is grossly unlikely that more
of these will occur.