summaryrefslogtreecommitdiff
path: root/Documentation/locking
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2020-06-02 01:45:27 +0300
committerLinus Torvalds <torvalds@linux-foundation.org>2020-06-02 01:45:27 +0300
commitb23c4771ff62de8ca9b5e4a2d64491b2fb6f8f69 (patch)
tree3ff6b2bdfec161fbc383bba06bab6329e81b02f7 /Documentation/locking
parentc2b0fc847f3122e5a4176c3772626a7a8facced0 (diff)
parente35b5a4c494a75a683ddf4901a43e0a128d5bfe3 (diff)
downloadlinux-b23c4771ff62de8ca9b5e4a2d64491b2fb6f8f69.tar.xz
Merge tag 'docs-5.8' of git://git.lwn.net/linux
Pull documentation updates from Jonathan Corbet: "A fair amount of stuff this time around, dominated by yet another massive set from Mauro toward the completion of the RST conversion. I *really* hope we are getting close to the end of this. Meanwhile, those patches reach pretty far afield to update document references around the tree; there should be no actual code changes there. There will be, alas, more of the usual trivial merge conflicts. Beyond that we have more translations, improvements to the sphinx scripting, a number of additions to the sysctl documentation, and lots of fixes" * tag 'docs-5.8' of git://git.lwn.net/linux: (130 commits) Documentation: fixes to the maintainer-entry-profile template zswap: docs/vm: Fix typo accept_threshold_percent in zswap.rst tracing: Fix events.rst section numbering docs: acpi: fix old http link and improve document format docs: filesystems: add info about efivars content Documentation: LSM: Correct the basic LSM description mailmap: change email for Ricardo Ribalda docs: sysctl/kernel: document unaligned controls Documentation: admin-guide: update bug-hunting.rst docs: sysctl/kernel: document ngroups_max nvdimm: fixes to maintainter-entry-profile Documentation/features: Correct RISC-V kprobes support entry Documentation/features: Refresh the arch support status files Revert "docs: sysctl/kernel: document ngroups_max" docs: move locking-specific documents to locking/ docs: move digsig docs to the security book docs: move the kref doc into the core-api book docs: add IRQ documentation at the core-api book docs: debugging-via-ohci1394.txt: add it to the core-api book docs: fix references for ipmi.rst file ...
Diffstat (limited to 'Documentation/locking')
-rw-r--r--Documentation/locking/futex-requeue-pi.rst132
-rw-r--r--Documentation/locking/hwspinlock.rst485
-rw-r--r--Documentation/locking/index.rst7
-rw-r--r--Documentation/locking/locktorture.rst2
-rw-r--r--Documentation/locking/percpu-rw-semaphore.rst28
-rw-r--r--Documentation/locking/pi-futex.rst122
-rw-r--r--Documentation/locking/preempt-locking.rst144
-rw-r--r--Documentation/locking/robust-futex-ABI.rst184
-rw-r--r--Documentation/locking/robust-futexes.rst221
-rw-r--r--Documentation/locking/rt-mutex.rst2
10 files changed, 1325 insertions, 2 deletions
diff --git a/Documentation/locking/futex-requeue-pi.rst b/Documentation/locking/futex-requeue-pi.rst
new file mode 100644
index 000000000000..14ab5787b9a7
--- /dev/null
+++ b/Documentation/locking/futex-requeue-pi.rst
@@ -0,0 +1,132 @@
+================
+Futex Requeue PI
+================
+
+Requeueing of tasks from a non-PI futex to a PI futex requires
+special handling in order to ensure the underlying rt_mutex is never
+left without an owner if it has waiters; doing so would break the PI
+boosting logic [see rt-mutex-desgin.txt] For the purposes of
+brevity, this action will be referred to as "requeue_pi" throughout
+this document. Priority inheritance is abbreviated throughout as
+"PI".
+
+Motivation
+----------
+
+Without requeue_pi, the glibc implementation of
+pthread_cond_broadcast() must resort to waking all the tasks waiting
+on a pthread_condvar and letting them try to sort out which task
+gets to run first in classic thundering-herd formation. An ideal
+implementation would wake the highest-priority waiter, and leave the
+rest to the natural wakeup inherent in unlocking the mutex
+associated with the condvar.
+
+Consider the simplified glibc calls::
+
+ /* caller must lock mutex */
+ pthread_cond_wait(cond, mutex)
+ {
+ lock(cond->__data.__lock);
+ unlock(mutex);
+ do {
+ unlock(cond->__data.__lock);
+ futex_wait(cond->__data.__futex);
+ lock(cond->__data.__lock);
+ } while(...)
+ unlock(cond->__data.__lock);
+ lock(mutex);
+ }
+
+ pthread_cond_broadcast(cond)
+ {
+ lock(cond->__data.__lock);
+ unlock(cond->__data.__lock);
+ futex_requeue(cond->data.__futex, cond->mutex);
+ }
+
+Once pthread_cond_broadcast() requeues the tasks, the cond->mutex
+has waiters. Note that pthread_cond_wait() attempts to lock the
+mutex only after it has returned to user space. This will leave the
+underlying rt_mutex with waiters, and no owner, breaking the
+previously mentioned PI-boosting algorithms.
+
+In order to support PI-aware pthread_condvar's, the kernel needs to
+be able to requeue tasks to PI futexes. This support implies that
+upon a successful futex_wait system call, the caller would return to
+user space already holding the PI futex. The glibc implementation
+would be modified as follows::
+
+
+ /* caller must lock mutex */
+ pthread_cond_wait_pi(cond, mutex)
+ {
+ lock(cond->__data.__lock);
+ unlock(mutex);
+ do {
+ unlock(cond->__data.__lock);
+ futex_wait_requeue_pi(cond->__data.__futex);
+ lock(cond->__data.__lock);
+ } while(...)
+ unlock(cond->__data.__lock);
+ /* the kernel acquired the mutex for us */
+ }
+
+ pthread_cond_broadcast_pi(cond)
+ {
+ lock(cond->__data.__lock);
+ unlock(cond->__data.__lock);
+ futex_requeue_pi(cond->data.__futex, cond->mutex);
+ }
+
+The actual glibc implementation will likely test for PI and make the
+necessary changes inside the existing calls rather than creating new
+calls for the PI cases. Similar changes are needed for
+pthread_cond_timedwait() and pthread_cond_signal().
+
+Implementation
+--------------
+
+In order to ensure the rt_mutex has an owner if it has waiters, it
+is necessary for both the requeue code, as well as the waiting code,
+to be able to acquire the rt_mutex before returning to user space.
+The requeue code cannot simply wake the waiter and leave it to
+acquire the rt_mutex as it would open a race window between the
+requeue call returning to user space and the waiter waking and
+starting to run. This is especially true in the uncontended case.
+
+The solution involves two new rt_mutex helper routines,
+rt_mutex_start_proxy_lock() and rt_mutex_finish_proxy_lock(), which
+allow the requeue code to acquire an uncontended rt_mutex on behalf
+of the waiter and to enqueue the waiter on a contended rt_mutex.
+Two new system calls provide the kernel<->user interface to
+requeue_pi: FUTEX_WAIT_REQUEUE_PI and FUTEX_CMP_REQUEUE_PI.
+
+FUTEX_WAIT_REQUEUE_PI is called by the waiter (pthread_cond_wait()
+and pthread_cond_timedwait()) to block on the initial futex and wait
+to be requeued to a PI-aware futex. The implementation is the
+result of a high-speed collision between futex_wait() and
+futex_lock_pi(), with some extra logic to check for the additional
+wake-up scenarios.
+
+FUTEX_CMP_REQUEUE_PI is called by the waker
+(pthread_cond_broadcast() and pthread_cond_signal()) to requeue and
+possibly wake the waiting tasks. Internally, this system call is
+still handled by futex_requeue (by passing requeue_pi=1). Before
+requeueing, futex_requeue() attempts to acquire the requeue target
+PI futex on behalf of the top waiter. If it can, this waiter is
+woken. futex_requeue() then proceeds to requeue the remaining
+nr_wake+nr_requeue tasks to the PI futex, calling
+rt_mutex_start_proxy_lock() prior to each requeue to prepare the
+task as a waiter on the underlying rt_mutex. It is possible that
+the lock can be acquired at this stage as well, if so, the next
+waiter is woken to finish the acquisition of the lock.
+
+FUTEX_CMP_REQUEUE_PI accepts nr_wake and nr_requeue as arguments, but
+their sum is all that really matters. futex_requeue() will wake or
+requeue up to nr_wake + nr_requeue tasks. It will wake only as many
+tasks as it can acquire the lock for, which in the majority of cases
+should be 0 as good programming practice dictates that the caller of
+either pthread_cond_broadcast() or pthread_cond_signal() acquire the
+mutex prior to making the call. FUTEX_CMP_REQUEUE_PI requires that
+nr_wake=1. nr_requeue should be INT_MAX for broadcast and 0 for
+signal.
diff --git a/Documentation/locking/hwspinlock.rst b/Documentation/locking/hwspinlock.rst
new file mode 100644
index 000000000000..6f03713b7003
--- /dev/null
+++ b/Documentation/locking/hwspinlock.rst
@@ -0,0 +1,485 @@
+===========================
+Hardware Spinlock Framework
+===========================
+
+Introduction
+============
+
+Hardware spinlock modules provide hardware assistance for synchronization
+and mutual exclusion between heterogeneous processors and those not operating
+under a single, shared operating system.
+
+For example, OMAP4 has dual Cortex-A9, dual Cortex-M3 and a C64x+ DSP,
+each of which is running a different Operating System (the master, A9,
+is usually running Linux and the slave processors, the M3 and the DSP,
+are running some flavor of RTOS).
+
+A generic hwspinlock framework allows platform-independent drivers to use
+the hwspinlock device in order to access data structures that are shared
+between remote processors, that otherwise have no alternative mechanism
+to accomplish synchronization and mutual exclusion operations.
+
+This is necessary, for example, for Inter-processor communications:
+on OMAP4, cpu-intensive multimedia tasks are offloaded by the host to the
+remote M3 and/or C64x+ slave processors (by an IPC subsystem called Syslink).
+
+To achieve fast message-based communications, a minimal kernel support
+is needed to deliver messages arriving from a remote processor to the
+appropriate user process.
+
+This communication is based on simple data structures that is shared between
+the remote processors, and access to it is synchronized using the hwspinlock
+module (remote processor directly places new messages in this shared data
+structure).
+
+A common hwspinlock interface makes it possible to have generic, platform-
+independent, drivers.
+
+User API
+========
+
+::
+
+ struct hwspinlock *hwspin_lock_request(void);
+
+Dynamically assign an hwspinlock and return its address, or NULL
+in case an unused hwspinlock isn't available. Users of this
+API will usually want to communicate the lock's id to the remote core
+before it can be used to achieve synchronization.
+
+Should be called from a process context (might sleep).
+
+::
+
+ struct hwspinlock *hwspin_lock_request_specific(unsigned int id);
+
+Assign a specific hwspinlock id and return its address, or NULL
+if that hwspinlock is already in use. Usually board code will
+be calling this function in order to reserve specific hwspinlock
+ids for predefined purposes.
+
+Should be called from a process context (might sleep).
+
+::
+
+ int of_hwspin_lock_get_id(struct device_node *np, int index);
+
+Retrieve the global lock id for an OF phandle-based specific lock.
+This function provides a means for DT users of a hwspinlock module
+to get the global lock id of a specific hwspinlock, so that it can
+be requested using the normal hwspin_lock_request_specific() API.
+
+The function returns a lock id number on success, -EPROBE_DEFER if
+the hwspinlock device is not yet registered with the core, or other
+error values.
+
+Should be called from a process context (might sleep).
+
+::
+
+ int hwspin_lock_free(struct hwspinlock *hwlock);
+
+Free a previously-assigned hwspinlock; returns 0 on success, or an
+appropriate error code on failure (e.g. -EINVAL if the hwspinlock
+is already free).
+
+Should be called from a process context (might sleep).
+
+::
+
+ int hwspin_lock_timeout(struct hwspinlock *hwlock, unsigned int timeout);
+
+Lock a previously-assigned hwspinlock with a timeout limit (specified in
+msecs). If the hwspinlock is already taken, the function will busy loop
+waiting for it to be released, but give up when the timeout elapses.
+Upon a successful return from this function, preemption is disabled so
+the caller must not sleep, and is advised to release the hwspinlock as
+soon as possible, in order to minimize remote cores polling on the
+hardware interconnect.
+
+Returns 0 when successful and an appropriate error code otherwise (most
+notably -ETIMEDOUT if the hwspinlock is still busy after timeout msecs).
+The function will never sleep.
+
+::
+
+ int hwspin_lock_timeout_irq(struct hwspinlock *hwlock, unsigned int timeout);
+
+Lock a previously-assigned hwspinlock with a timeout limit (specified in
+msecs). If the hwspinlock is already taken, the function will busy loop
+waiting for it to be released, but give up when the timeout elapses.
+Upon a successful return from this function, preemption and the local
+interrupts are disabled, so the caller must not sleep, and is advised to
+release the hwspinlock as soon as possible.
+
+Returns 0 when successful and an appropriate error code otherwise (most
+notably -ETIMEDOUT if the hwspinlock is still busy after timeout msecs).
+The function will never sleep.
+
+::
+
+ int hwspin_lock_timeout_irqsave(struct hwspinlock *hwlock, unsigned int to,
+ unsigned long *flags);
+
+Lock a previously-assigned hwspinlock with a timeout limit (specified in
+msecs). If the hwspinlock is already taken, the function will busy loop
+waiting for it to be released, but give up when the timeout elapses.
+Upon a successful return from this function, preemption is disabled,
+local interrupts are disabled and their previous state is saved at the
+given flags placeholder. The caller must not sleep, and is advised to
+release the hwspinlock as soon as possible.
+
+Returns 0 when successful and an appropriate error code otherwise (most
+notably -ETIMEDOUT if the hwspinlock is still busy after timeout msecs).
+
+The function will never sleep.
+
+::
+
+ int hwspin_lock_timeout_raw(struct hwspinlock *hwlock, unsigned int timeout);
+
+Lock a previously-assigned hwspinlock with a timeout limit (specified in
+msecs). If the hwspinlock is already taken, the function will busy loop
+waiting for it to be released, but give up when the timeout elapses.
+
+Caution: User must protect the routine of getting hardware lock with mutex
+or spinlock to avoid dead-lock, that will let user can do some time-consuming
+or sleepable operations under the hardware lock.
+
+Returns 0 when successful and an appropriate error code otherwise (most
+notably -ETIMEDOUT if the hwspinlock is still busy after timeout msecs).
+
+The function will never sleep.
+
+::
+
+ int hwspin_lock_timeout_in_atomic(struct hwspinlock *hwlock, unsigned int to);
+
+Lock a previously-assigned hwspinlock with a timeout limit (specified in
+msecs). If the hwspinlock is already taken, the function will busy loop
+waiting for it to be released, but give up when the timeout elapses.
+
+This function shall be called only from an atomic context and the timeout
+value shall not exceed a few msecs.
+
+Returns 0 when successful and an appropriate error code otherwise (most
+notably -ETIMEDOUT if the hwspinlock is still busy after timeout msecs).
+
+The function will never sleep.
+
+::
+
+ int hwspin_trylock(struct hwspinlock *hwlock);
+
+
+Attempt to lock a previously-assigned hwspinlock, but immediately fail if
+it is already taken.
+
+Upon a successful return from this function, preemption is disabled so
+caller must not sleep, and is advised to release the hwspinlock as soon as
+possible, in order to minimize remote cores polling on the hardware
+interconnect.
+
+Returns 0 on success and an appropriate error code otherwise (most
+notably -EBUSY if the hwspinlock was already taken).
+The function will never sleep.
+
+::
+
+ int hwspin_trylock_irq(struct hwspinlock *hwlock);
+
+
+Attempt to lock a previously-assigned hwspinlock, but immediately fail if
+it is already taken.
+
+Upon a successful return from this function, preemption and the local
+interrupts are disabled so caller must not sleep, and is advised to
+release the hwspinlock as soon as possible.
+
+Returns 0 on success and an appropriate error code otherwise (most
+notably -EBUSY if the hwspinlock was already taken).
+
+The function will never sleep.
+
+::
+
+ int hwspin_trylock_irqsave(struct hwspinlock *hwlock, unsigned long *flags);
+
+Attempt to lock a previously-assigned hwspinlock, but immediately fail if
+it is already taken.
+
+Upon a successful return from this function, preemption is disabled,
+the local interrupts are disabled and their previous state is saved
+at the given flags placeholder. The caller must not sleep, and is advised
+to release the hwspinlock as soon as possible.
+
+Returns 0 on success and an appropriate error code otherwise (most
+notably -EBUSY if the hwspinlock was already taken).
+The function will never sleep.
+
+::
+
+ int hwspin_trylock_raw(struct hwspinlock *hwlock);
+
+Attempt to lock a previously-assigned hwspinlock, but immediately fail if
+it is already taken.
+
+Caution: User must protect the routine of getting hardware lock with mutex
+or spinlock to avoid dead-lock, that will let user can do some time-consuming
+or sleepable operations under the hardware lock.
+
+Returns 0 on success and an appropriate error code otherwise (most
+notably -EBUSY if the hwspinlock was already taken).
+The function will never sleep.
+
+::
+
+ int hwspin_trylock_in_atomic(struct hwspinlock *hwlock);
+
+Attempt to lock a previously-assigned hwspinlock, but immediately fail if
+it is already taken.
+
+This function shall be called only from an atomic context.
+
+Returns 0 on success and an appropriate error code otherwise (most
+notably -EBUSY if the hwspinlock was already taken).
+The function will never sleep.
+
+::
+
+ void hwspin_unlock(struct hwspinlock *hwlock);
+
+Unlock a previously-locked hwspinlock. Always succeed, and can be called
+from any context (the function never sleeps).
+
+.. note::
+
+ code should **never** unlock an hwspinlock which is already unlocked
+ (there is no protection against this).
+
+::
+
+ void hwspin_unlock_irq(struct hwspinlock *hwlock);
+
+Unlock a previously-locked hwspinlock and enable local interrupts.
+The caller should **never** unlock an hwspinlock which is already unlocked.
+
+Doing so is considered a bug (there is no protection against this).
+Upon a successful return from this function, preemption and local
+interrupts are enabled. This function will never sleep.
+
+::
+
+ void
+ hwspin_unlock_irqrestore(struct hwspinlock *hwlock, unsigned long *flags);
+
+Unlock a previously-locked hwspinlock.
+
+The caller should **never** unlock an hwspinlock which is already unlocked.
+Doing so is considered a bug (there is no protection against this).
+Upon a successful return from this function, preemption is reenabled,
+and the state of the local interrupts is restored to the state saved at
+the given flags. This function will never sleep.
+
+::
+
+ void hwspin_unlock_raw(struct hwspinlock *hwlock);
+
+Unlock a previously-locked hwspinlock.
+
+The caller should **never** unlock an hwspinlock which is already unlocked.
+Doing so is considered a bug (there is no protection against this).
+This function will never sleep.
+
+::
+
+ void hwspin_unlock_in_atomic(struct hwspinlock *hwlock);
+
+Unlock a previously-locked hwspinlock.
+
+The caller should **never** unlock an hwspinlock which is already unlocked.
+Doing so is considered a bug (there is no protection against this).
+This function will never sleep.
+
+::
+
+ int hwspin_lock_get_id(struct hwspinlock *hwlock);
+
+Retrieve id number of a given hwspinlock. This is needed when an
+hwspinlock is dynamically assigned: before it can be used to achieve
+mutual exclusion with a remote cpu, the id number should be communicated
+to the remote task with which we want to synchronize.
+
+Returns the hwspinlock id number, or -EINVAL if hwlock is null.
+
+Typical usage
+=============
+
+::
+
+ #include <linux/hwspinlock.h>
+ #include <linux/err.h>
+
+ int hwspinlock_example1(void)
+ {
+ struct hwspinlock *hwlock;
+ int ret;
+
+ /* dynamically assign a hwspinlock */
+ hwlock = hwspin_lock_request();
+ if (!hwlock)
+ ...
+
+ id = hwspin_lock_get_id(hwlock);
+ /* probably need to communicate id to a remote processor now */
+
+ /* take the lock, spin for 1 sec if it's already taken */
+ ret = hwspin_lock_timeout(hwlock, 1000);
+ if (ret)
+ ...
+
+ /*
+ * we took the lock, do our thing now, but do NOT sleep
+ */
+
+ /* release the lock */
+ hwspin_unlock(hwlock);
+
+ /* free the lock */
+ ret = hwspin_lock_free(hwlock);
+ if (ret)
+ ...
+
+ return ret;
+ }
+
+ int hwspinlock_example2(void)
+ {
+ struct hwspinlock *hwlock;
+ int ret;
+
+ /*
+ * assign a specific hwspinlock id - this should be called early
+ * by board init code.
+ */
+ hwlock = hwspin_lock_request_specific(PREDEFINED_LOCK_ID);
+ if (!hwlock)
+ ...
+
+ /* try to take it, but don't spin on it */
+ ret = hwspin_trylock(hwlock);
+ if (!ret) {
+ pr_info("lock is already taken\n");
+ return -EBUSY;
+ }
+
+ /*
+ * we took the lock, do our thing now, but do NOT sleep
+ */
+
+ /* release the lock */
+ hwspin_unlock(hwlock);
+
+ /* free the lock */
+ ret = hwspin_lock_free(hwlock);
+ if (ret)
+ ...
+
+ return ret;
+ }
+
+
+API for implementors
+====================
+
+::
+
+ int hwspin_lock_register(struct hwspinlock_device *bank, struct device *dev,
+ const struct hwspinlock_ops *ops, int base_id, int num_locks);
+
+To be called from the underlying platform-specific implementation, in
+order to register a new hwspinlock device (which is usually a bank of
+numerous locks). Should be called from a process context (this function
+might sleep).
+
+Returns 0 on success, or appropriate error code on failure.
+
+::
+
+ int hwspin_lock_unregister(struct hwspinlock_device *bank);
+
+To be called from the underlying vendor-specific implementation, in order
+to unregister an hwspinlock device (which is usually a bank of numerous
+locks).
+
+Should be called from a process context (this function might sleep).
+
+Returns the address of hwspinlock on success, or NULL on error (e.g.
+if the hwspinlock is still in use).
+
+Important structs
+=================
+
+struct hwspinlock_device is a device which usually contains a bank
+of hardware locks. It is registered by the underlying hwspinlock
+implementation using the hwspin_lock_register() API.
+
+::
+
+ /**
+ * struct hwspinlock_device - a device which usually spans numerous hwspinlocks
+ * @dev: underlying device, will be used to invoke runtime PM api
+ * @ops: platform-specific hwspinlock handlers
+ * @base_id: id index of the first lock in this device
+ * @num_locks: number of locks in this device
+ * @lock: dynamically allocated array of 'struct hwspinlock'
+ */
+ struct hwspinlock_device {
+ struct device *dev;
+ const struct hwspinlock_ops *ops;
+ int base_id;
+ int num_locks;
+ struct hwspinlock lock[0];
+ };
+
+struct hwspinlock_device contains an array of hwspinlock structs, each
+of which represents a single hardware lock::
+
+ /**
+ * struct hwspinlock - this struct represents a single hwspinlock instance
+ * @bank: the hwspinlock_device structure which owns this lock
+ * @lock: initialized and used by hwspinlock core
+ * @priv: private data, owned by the underlying platform-specific hwspinlock drv
+ */
+ struct hwspinlock {
+ struct hwspinlock_device *bank;
+ spinlock_t lock;
+ void *priv;
+ };
+
+When registering a bank of locks, the hwspinlock driver only needs to
+set the priv members of the locks. The rest of the members are set and
+initialized by the hwspinlock core itself.
+
+Implementation callbacks
+========================
+
+There are three possible callbacks defined in 'struct hwspinlock_ops'::
+
+ struct hwspinlock_ops {
+ int (*trylock)(struct hwspinlock *lock);
+ void (*unlock)(struct hwspinlock *lock);
+ void (*relax)(struct hwspinlock *lock);
+ };
+
+The first two callbacks are mandatory:
+
+The ->trylock() callback should make a single attempt to take the lock, and
+return 0 on failure and 1 on success. This callback may **not** sleep.
+
+The ->unlock() callback releases the lock. It always succeed, and it, too,
+may **not** sleep.
+
+The ->relax() callback is optional. It is called by hwspinlock core while
+spinning on a lock, and can be used by the underlying implementation to force
+a delay between two successive invocations of ->trylock(). It may **not** sleep.
diff --git a/Documentation/locking/index.rst b/Documentation/locking/index.rst
index 5d6800a723dc..d785878cad65 100644
--- a/Documentation/locking/index.rst
+++ b/Documentation/locking/index.rst
@@ -16,6 +16,13 @@ locking
rt-mutex
spinlocks
ww-mutex-design
+ preempt-locking
+ pi-futex
+ futex-requeue-pi
+ hwspinlock
+ percpu-rw-semaphore
+ robust-futexes
+ robust-futex-ABI
.. only:: subproject and html
diff --git a/Documentation/locking/locktorture.rst b/Documentation/locking/locktorture.rst
index 5bcb99ba7bd9..8012a74555e7 100644
--- a/Documentation/locking/locktorture.rst
+++ b/Documentation/locking/locktorture.rst
@@ -110,7 +110,7 @@ stutter
same period of time. Defaults to "stutter=5", so as
to run and pause for (roughly) five-second intervals.
Specifying "stutter=0" causes the test to run continuously
- without pausing, which is the old default behavior.
+ without pausing.
shuffle_interval
The number of seconds to keep the test threads affinitied
diff --git a/Documentation/locking/percpu-rw-semaphore.rst b/Documentation/locking/percpu-rw-semaphore.rst
new file mode 100644
index 000000000000..247de6410855
--- /dev/null
+++ b/Documentation/locking/percpu-rw-semaphore.rst
@@ -0,0 +1,28 @@
+====================
+Percpu rw semaphores
+====================
+
+Percpu rw semaphores is a new read-write semaphore design that is
+optimized for locking for reading.
+
+The problem with traditional read-write semaphores is that when multiple
+cores take the lock for reading, the cache line containing the semaphore
+is bouncing between L1 caches of the cores, causing performance
+degradation.
+
+Locking for reading is very fast, it uses RCU and it avoids any atomic
+instruction in the lock and unlock path. On the other hand, locking for
+writing is very expensive, it calls synchronize_rcu() that can take
+hundreds of milliseconds.
+
+The lock is declared with "struct percpu_rw_semaphore" type.
+The lock is initialized percpu_init_rwsem, it returns 0 on success and
+-ENOMEM on allocation failure.
+The lock must be freed with percpu_free_rwsem to avoid memory leak.
+
+The lock is locked for read with percpu_down_read, percpu_up_read and
+for write with percpu_down_write, percpu_up_write.
+
+The idea of using RCU for optimized rw-lock was introduced by
+Eric Dumazet <eric.dumazet@gmail.com>.
+The code was written by Mikulas Patocka <mpatocka@redhat.com>
diff --git a/Documentation/locking/pi-futex.rst b/Documentation/locking/pi-futex.rst
new file mode 100644
index 000000000000..c33ba2befbf8
--- /dev/null
+++ b/Documentation/locking/pi-futex.rst
@@ -0,0 +1,122 @@
+======================
+Lightweight PI-futexes
+======================
+
+We are calling them lightweight for 3 reasons:
+
+ - in the user-space fastpath a PI-enabled futex involves no kernel work
+ (or any other PI complexity) at all. No registration, no extra kernel
+ calls - just pure fast atomic ops in userspace.
+
+ - even in the slowpath, the system call and scheduling pattern is very
+ similar to normal futexes.
+
+ - the in-kernel PI implementation is streamlined around the mutex
+ abstraction, with strict rules that keep the implementation
+ relatively simple: only a single owner may own a lock (i.e. no
+ read-write lock support), only the owner may unlock a lock, no
+ recursive locking, etc.
+
+Priority Inheritance - why?
+---------------------------
+
+The short reply: user-space PI helps achieving/improving determinism for
+user-space applications. In the best-case, it can help achieve
+determinism and well-bound latencies. Even in the worst-case, PI will
+improve the statistical distribution of locking related application
+delays.
+
+The longer reply
+----------------
+
+Firstly, sharing locks between multiple tasks is a common programming
+technique that often cannot be replaced with lockless algorithms. As we
+can see it in the kernel [which is a quite complex program in itself],
+lockless structures are rather the exception than the norm - the current
+ratio of lockless vs. locky code for shared data structures is somewhere
+between 1:10 and 1:100. Lockless is hard, and the complexity of lockless
+algorithms often endangers to ability to do robust reviews of said code.
+I.e. critical RT apps often choose lock structures to protect critical
+data structures, instead of lockless algorithms. Furthermore, there are
+cases (like shared hardware, or other resource limits) where lockless
+access is mathematically impossible.
+
+Media players (such as Jack) are an example of reasonable application
+design with multiple tasks (with multiple priority levels) sharing
+short-held locks: for example, a highprio audio playback thread is
+combined with medium-prio construct-audio-data threads and low-prio
+display-colory-stuff threads. Add video and decoding to the mix and
+we've got even more priority levels.
+
+So once we accept that synchronization objects (locks) are an
+unavoidable fact of life, and once we accept that multi-task userspace
+apps have a very fair expectation of being able to use locks, we've got
+to think about how to offer the option of a deterministic locking
+implementation to user-space.
+
+Most of the technical counter-arguments against doing priority
+inheritance only apply to kernel-space locks. But user-space locks are
+different, there we cannot disable interrupts or make the task
+non-preemptible in a critical section, so the 'use spinlocks' argument
+does not apply (user-space spinlocks have the same priority inversion
+problems as other user-space locking constructs). Fact is, pretty much
+the only technique that currently enables good determinism for userspace
+locks (such as futex-based pthread mutexes) is priority inheritance:
+
+Currently (without PI), if a high-prio and a low-prio task shares a lock
+[this is a quite common scenario for most non-trivial RT applications],
+even if all critical sections are coded carefully to be deterministic
+(i.e. all critical sections are short in duration and only execute a
+limited number of instructions), the kernel cannot guarantee any
+deterministic execution of the high-prio task: any medium-priority task
+could preempt the low-prio task while it holds the shared lock and
+executes the critical section, and could delay it indefinitely.
+
+Implementation
+--------------
+
+As mentioned before, the userspace fastpath of PI-enabled pthread
+mutexes involves no kernel work at all - they behave quite similarly to
+normal futex-based locks: a 0 value means unlocked, and a value==TID
+means locked. (This is the same method as used by list-based robust
+futexes.) Userspace uses atomic ops to lock/unlock these mutexes without
+entering the kernel.
+
+To handle the slowpath, we have added two new futex ops:
+
+ - FUTEX_LOCK_PI
+ - FUTEX_UNLOCK_PI
+
+If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to
+TID fails], then FUTEX_LOCK_PI is called. The kernel does all the
+remaining work: if there is no futex-queue attached to the futex address
+yet then the code looks up the task that owns the futex [it has put its
+own TID into the futex value], and attaches a 'PI state' structure to
+the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware,
+kernel-based synchronization object. The 'other' task is made the owner
+of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the
+futex value. Then this task tries to lock the rt-mutex, on which it
+blocks. Once it returns, it has the mutex acquired, and it sets the
+futex value to its own TID and returns. Userspace has no other work to
+perform - it now owns the lock, and futex value contains
+FUTEX_WAITERS|TID.
+
+If the unlock side fastpath succeeds, [i.e. userspace manages to do a
+TID -> 0 atomic transition of the futex value], then no kernel work is
+triggered.
+
+If the unlock fastpath fails (because the FUTEX_WAITERS bit is set),
+then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the
+behalf of userspace - and it also unlocks the attached
+pi_state->rt_mutex and thus wakes up any potential waiters.
+
+Note that under this approach, contrary to previous PI-futex approaches,
+there is no prior 'registration' of a PI-futex. [which is not quite
+possible anyway, due to existing ABI properties of pthread mutexes.]
+
+Also, under this scheme, 'robustness' and 'PI' are two orthogonal
+properties of futexes, and all four combinations are possible: futex,
+robust-futex, PI-futex, robust+PI-futex.
+
+More details about priority inheritance can be found in
+Documentation/locking/rt-mutex.rst.
diff --git a/Documentation/locking/preempt-locking.rst b/Documentation/locking/preempt-locking.rst
new file mode 100644
index 000000000000..dce336134e54
--- /dev/null
+++ b/Documentation/locking/preempt-locking.rst
@@ -0,0 +1,144 @@
+===========================================================================
+Proper Locking Under a Preemptible Kernel: Keeping Kernel Code Preempt-Safe
+===========================================================================
+
+:Author: Robert Love <rml@tech9.net>
+
+
+Introduction
+============
+
+
+A preemptible kernel creates new locking issues. The issues are the same as
+those under SMP: concurrency and reentrancy. Thankfully, the Linux preemptible
+kernel model leverages existing SMP locking mechanisms. Thus, the kernel
+requires explicit additional locking for very few additional situations.
+
+This document is for all kernel hackers. Developing code in the kernel
+requires protecting these situations.
+
+
+RULE #1: Per-CPU data structures need explicit protection
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+
+Two similar problems arise. An example code snippet::
+
+ struct this_needs_locking tux[NR_CPUS];
+ tux[smp_processor_id()] = some_value;
+ /* task is preempted here... */
+ something = tux[smp_processor_id()];
+
+First, since the data is per-CPU, it may not have explicit SMP locking, but
+require it otherwise. Second, when a preempted task is finally rescheduled,
+the previous value of smp_processor_id may not equal the current. You must
+protect these situations by disabling preemption around them.
+
+You can also use put_cpu() and get_cpu(), which will disable preemption.
+
+
+RULE #2: CPU state must be protected.
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+
+Under preemption, the state of the CPU must be protected. This is arch-
+dependent, but includes CPU structures and state not preserved over a context
+switch. For example, on x86, entering and exiting FPU mode is now a critical
+section that must occur while preemption is disabled. Think what would happen
+if the kernel is executing a floating-point instruction and is then preempted.
+Remember, the kernel does not save FPU state except for user tasks. Therefore,
+upon preemption, the FPU registers will be sold to the lowest bidder. Thus,
+preemption must be disabled around such regions.
+
+Note, some FPU functions are already explicitly preempt safe. For example,
+kernel_fpu_begin and kernel_fpu_end will disable and enable preemption.
+
+
+RULE #3: Lock acquire and release must be performed by same task
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+
+A lock acquired in one task must be released by the same task. This
+means you can't do oddball things like acquire a lock and go off to
+play while another task releases it. If you want to do something
+like this, acquire and release the task in the same code path and
+have the caller wait on an event by the other task.
+
+
+Solution
+========
+
+
+Data protection under preemption is achieved by disabling preemption for the
+duration of the critical region.
+
+::
+
+ preempt_enable() decrement the preempt counter
+ preempt_disable() increment the preempt counter
+ preempt_enable_no_resched() decrement, but do not immediately preempt
+ preempt_check_resched() if needed, reschedule
+ preempt_count() return the preempt counter
+
+The functions are nestable. In other words, you can call preempt_disable
+n-times in a code path, and preemption will not be reenabled until the n-th
+call to preempt_enable. The preempt statements define to nothing if
+preemption is not enabled.
+
+Note that you do not need to explicitly prevent preemption if you are holding
+any locks or interrupts are disabled, since preemption is implicitly disabled
+in those cases.
+
+But keep in mind that 'irqs disabled' is a fundamentally unsafe way of
+disabling preemption - any cond_resched() or cond_resched_lock() might trigger
+a reschedule if the preempt count is 0. A simple printk() might trigger a
+reschedule. So use this implicit preemption-disabling property only if you
+know that the affected codepath does not do any of this. Best policy is to use
+this only for small, atomic code that you wrote and which calls no complex
+functions.
+
+Example::
+
+ cpucache_t *cc; /* this is per-CPU */
+ preempt_disable();
+ cc = cc_data(searchp);
+ if (cc && cc->avail) {
+ __free_block(searchp, cc_entry(cc), cc->avail);
+ cc->avail = 0;
+ }
+ preempt_enable();
+ return 0;
+
+Notice how the preemption statements must encompass every reference of the
+critical variables. Another example::
+
+ int buf[NR_CPUS];
+ set_cpu_val(buf);
+ if (buf[smp_processor_id()] == -1) printf(KERN_INFO "wee!\n");
+ spin_lock(&buf_lock);
+ /* ... */
+
+This code is not preempt-safe, but see how easily we can fix it by simply
+moving the spin_lock up two lines.
+
+
+Preventing preemption using interrupt disabling
+===============================================
+
+
+It is possible to prevent a preemption event using local_irq_disable and
+local_irq_save. Note, when doing so, you must be very careful to not cause
+an event that would set need_resched and result in a preemption check. When
+in doubt, rely on locking or explicit preemption disabling.
+
+Note in 2.5 interrupt disabling is now only per-CPU (e.g. local).
+
+An additional concern is proper usage of local_irq_disable and local_irq_save.
+These may be used to protect from preemption, however, on exit, if preemption
+may be enabled, a test to see if preemption is required should be done. If
+these are called from the spin_lock and read/write lock macros, the right thing
+is done. They may also be called within a spin-lock protected region, however,
+if they are ever called outside of this context, a test for preemption should
+be made. Do note that calls from interrupt context or bottom half/ tasklets
+are also protected by preemption locks and so may use the versions which do
+not check preemption.
diff --git a/Documentation/locking/robust-futex-ABI.rst b/Documentation/locking/robust-futex-ABI.rst
new file mode 100644
index 000000000000..f24904f1c16f
--- /dev/null
+++ b/Documentation/locking/robust-futex-ABI.rst
@@ -0,0 +1,184 @@
+====================
+The robust futex ABI
+====================
+
+:Author: Started by Paul Jackson <pj@sgi.com>
+
+
+Robust_futexes provide a mechanism that is used in addition to normal
+futexes, for kernel assist of cleanup of held locks on task exit.
+
+The interesting data as to what futexes a thread is holding is kept on a
+linked list in user space, where it can be updated efficiently as locks
+are taken and dropped, without kernel intervention. The only additional
+kernel intervention required for robust_futexes above and beyond what is
+required for futexes is:
+
+ 1) a one time call, per thread, to tell the kernel where its list of
+ held robust_futexes begins, and
+ 2) internal kernel code at exit, to handle any listed locks held
+ by the exiting thread.
+
+The existing normal futexes already provide a "Fast Userspace Locking"
+mechanism, which handles uncontested locking without needing a system
+call, and handles contested locking by maintaining a list of waiting
+threads in the kernel. Options on the sys_futex(2) system call support
+waiting on a particular futex, and waking up the next waiter on a
+particular futex.
+
+For robust_futexes to work, the user code (typically in a library such
+as glibc linked with the application) has to manage and place the
+necessary list elements exactly as the kernel expects them. If it fails
+to do so, then improperly listed locks will not be cleaned up on exit,
+probably causing deadlock or other such failure of the other threads
+waiting on the same locks.
+
+A thread that anticipates possibly using robust_futexes should first
+issue the system call::
+
+ asmlinkage long
+ sys_set_robust_list(struct robust_list_head __user *head, size_t len);
+
+The pointer 'head' points to a structure in the threads address space
+consisting of three words. Each word is 32 bits on 32 bit arch's, or 64
+bits on 64 bit arch's, and local byte order. Each thread should have
+its own thread private 'head'.
+
+If a thread is running in 32 bit compatibility mode on a 64 native arch
+kernel, then it can actually have two such structures - one using 32 bit
+words for 32 bit compatibility mode, and one using 64 bit words for 64
+bit native mode. The kernel, if it is a 64 bit kernel supporting 32 bit
+compatibility mode, will attempt to process both lists on each task
+exit, if the corresponding sys_set_robust_list() call has been made to
+setup that list.
+
+ The first word in the memory structure at 'head' contains a
+ pointer to a single linked list of 'lock entries', one per lock,
+ as described below. If the list is empty, the pointer will point
+ to itself, 'head'. The last 'lock entry' points back to the 'head'.
+
+ The second word, called 'offset', specifies the offset from the
+ address of the associated 'lock entry', plus or minus, of what will
+ be called the 'lock word', from that 'lock entry'. The 'lock word'
+ is always a 32 bit word, unlike the other words above. The 'lock
+ word' holds 2 flag bits in the upper 2 bits, and the thread id (TID)
+ of the thread holding the lock in the bottom 30 bits. See further
+ below for a description of the flag bits.
+
+ The third word, called 'list_op_pending', contains transient copy of
+ the address of the 'lock entry', during list insertion and removal,
+ and is needed to correctly resolve races should a thread exit while
+ in the middle of a locking or unlocking operation.
+
+Each 'lock entry' on the single linked list starting at 'head' consists
+of just a single word, pointing to the next 'lock entry', or back to
+'head' if there are no more entries. In addition, nearby to each 'lock
+entry', at an offset from the 'lock entry' specified by the 'offset'
+word, is one 'lock word'.
+
+The 'lock word' is always 32 bits, and is intended to be the same 32 bit
+lock variable used by the futex mechanism, in conjunction with
+robust_futexes. The kernel will only be able to wakeup the next thread
+waiting for a lock on a threads exit if that next thread used the futex
+mechanism to register the address of that 'lock word' with the kernel.
+
+For each futex lock currently held by a thread, if it wants this
+robust_futex support for exit cleanup of that lock, it should have one
+'lock entry' on this list, with its associated 'lock word' at the
+specified 'offset'. Should a thread die while holding any such locks,
+the kernel will walk this list, mark any such locks with a bit
+indicating their holder died, and wakeup the next thread waiting for
+that lock using the futex mechanism.
+
+When a thread has invoked the above system call to indicate it
+anticipates using robust_futexes, the kernel stores the passed in 'head'
+pointer for that task. The task may retrieve that value later on by
+using the system call::
+
+ asmlinkage long
+ sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
+ size_t __user *len_ptr);
+
+It is anticipated that threads will use robust_futexes embedded in
+larger, user level locking structures, one per lock. The kernel
+robust_futex mechanism doesn't care what else is in that structure, so
+long as the 'offset' to the 'lock word' is the same for all
+robust_futexes used by that thread. The thread should link those locks
+it currently holds using the 'lock entry' pointers. It may also have
+other links between the locks, such as the reverse side of a double
+linked list, but that doesn't matter to the kernel.
+
+By keeping its locks linked this way, on a list starting with a 'head'
+pointer known to the kernel, the kernel can provide to a thread the
+essential service available for robust_futexes, which is to help clean
+up locks held at the time of (a perhaps unexpectedly) exit.
+
+Actual locking and unlocking, during normal operations, is handled
+entirely by user level code in the contending threads, and by the
+existing futex mechanism to wait for, and wakeup, locks. The kernels
+only essential involvement in robust_futexes is to remember where the
+list 'head' is, and to walk the list on thread exit, handling locks
+still held by the departing thread, as described below.
+
+There may exist thousands of futex lock structures in a threads shared
+memory, on various data structures, at a given point in time. Only those
+lock structures for locks currently held by that thread should be on
+that thread's robust_futex linked lock list a given time.
+
+A given futex lock structure in a user shared memory region may be held
+at different times by any of the threads with access to that region. The
+thread currently holding such a lock, if any, is marked with the threads
+TID in the lower 30 bits of the 'lock word'.
+
+When adding or removing a lock from its list of held locks, in order for
+the kernel to correctly handle lock cleanup regardless of when the task
+exits (perhaps it gets an unexpected signal 9 in the middle of
+manipulating this list), the user code must observe the following
+protocol on 'lock entry' insertion and removal:
+
+On insertion:
+
+ 1) set the 'list_op_pending' word to the address of the 'lock entry'
+ to be inserted,
+ 2) acquire the futex lock,
+ 3) add the lock entry, with its thread id (TID) in the bottom 30 bits
+ of the 'lock word', to the linked list starting at 'head', and
+ 4) clear the 'list_op_pending' word.
+
+On removal:
+
+ 1) set the 'list_op_pending' word to the address of the 'lock entry'
+ to be removed,
+ 2) remove the lock entry for this lock from the 'head' list,
+ 3) release the futex lock, and
+ 4) clear the 'lock_op_pending' word.
+
+On exit, the kernel will consider the address stored in
+'list_op_pending' and the address of each 'lock word' found by walking
+the list starting at 'head'. For each such address, if the bottom 30
+bits of the 'lock word' at offset 'offset' from that address equals the
+exiting threads TID, then the kernel will do two things:
+
+ 1) if bit 31 (0x80000000) is set in that word, then attempt a futex
+ wakeup on that address, which will waken the next thread that has
+ used to the futex mechanism to wait on that address, and
+ 2) atomically set bit 30 (0x40000000) in the 'lock word'.
+
+In the above, bit 31 was set by futex waiters on that lock to indicate
+they were waiting, and bit 30 is set by the kernel to indicate that the
+lock owner died holding the lock.
+
+The kernel exit code will silently stop scanning the list further if at
+any point:
+
+ 1) the 'head' pointer or an subsequent linked list pointer
+ is not a valid address of a user space word
+ 2) the calculated location of the 'lock word' (address plus
+ 'offset') is not the valid address of a 32 bit user space
+ word
+ 3) if the list contains more than 1 million (subject to
+ future kernel configuration changes) elements.
+
+When the kernel sees a list entry whose 'lock word' doesn't have the
+current threads TID in the lower 30 bits, it does nothing with that
+entry, and goes on to the next entry.
diff --git a/Documentation/locking/robust-futexes.rst b/Documentation/locking/robust-futexes.rst
new file mode 100644
index 000000000000..6361fb01c9c1
--- /dev/null
+++ b/Documentation/locking/robust-futexes.rst
@@ -0,0 +1,221 @@
+========================================
+A description of what robust futexes are
+========================================
+
+:Started by: Ingo Molnar <mingo@redhat.com>
+
+Background
+----------
+
+what are robust futexes? To answer that, we first need to understand
+what futexes are: normal futexes are special types of locks that in the
+noncontended case can be acquired/released from userspace without having
+to enter the kernel.
+
+A futex is in essence a user-space address, e.g. a 32-bit lock variable
+field. If userspace notices contention (the lock is already owned and
+someone else wants to grab it too) then the lock is marked with a value
+that says "there's a waiter pending", and the sys_futex(FUTEX_WAIT)
+syscall is used to wait for the other guy to release it. The kernel
+creates a 'futex queue' internally, so that it can later on match up the
+waiter with the waker - without them having to know about each other.
+When the owner thread releases the futex, it notices (via the variable
+value) that there were waiter(s) pending, and does the
+sys_futex(FUTEX_WAKE) syscall to wake them up. Once all waiters have
+taken and released the lock, the futex is again back to 'uncontended'
+state, and there's no in-kernel state associated with it. The kernel
+completely forgets that there ever was a futex at that address. This
+method makes futexes very lightweight and scalable.
+
+"Robustness" is about dealing with crashes while holding a lock: if a
+process exits prematurely while holding a pthread_mutex_t lock that is
+also shared with some other process (e.g. yum segfaults while holding a
+pthread_mutex_t, or yum is kill -9-ed), then waiters for that lock need
+to be notified that the last owner of the lock exited in some irregular
+way.
+
+To solve such types of problems, "robust mutex" userspace APIs were
+created: pthread_mutex_lock() returns an error value if the owner exits
+prematurely - and the new owner can decide whether the data protected by
+the lock can be recovered safely.
+
+There is a big conceptual problem with futex based mutexes though: it is
+the kernel that destroys the owner task (e.g. due to a SEGFAULT), but
+the kernel cannot help with the cleanup: if there is no 'futex queue'
+(and in most cases there is none, futexes being fast lightweight locks)
+then the kernel has no information to clean up after the held lock!
+Userspace has no chance to clean up after the lock either - userspace is
+the one that crashes, so it has no opportunity to clean up. Catch-22.
+
+In practice, when e.g. yum is kill -9-ed (or segfaults), a system reboot
+is needed to release that futex based lock. This is one of the leading
+bugreports against yum.
+
+To solve this problem, the traditional approach was to extend the vma
+(virtual memory area descriptor) concept to have a notion of 'pending
+robust futexes attached to this area'. This approach requires 3 new
+syscall variants to sys_futex(): FUTEX_REGISTER, FUTEX_DEREGISTER and
+FUTEX_RECOVER. At do_exit() time, all vmas are searched to see whether
+they have a robust_head set. This approach has two fundamental problems
+left:
+
+ - it has quite complex locking and race scenarios. The vma-based
+ approach had been pending for years, but they are still not completely
+ reliable.
+
+ - they have to scan _every_ vma at sys_exit() time, per thread!
+
+The second disadvantage is a real killer: pthread_exit() takes around 1
+microsecond on Linux, but with thousands (or tens of thousands) of vmas
+every pthread_exit() takes a millisecond or more, also totally
+destroying the CPU's L1 and L2 caches!
+
+This is very much noticeable even for normal process sys_exit_group()
+calls: the kernel has to do the vma scanning unconditionally! (this is
+because the kernel has no knowledge about how many robust futexes there
+are to be cleaned up, because a robust futex might have been registered
+in another task, and the futex variable might have been simply mmap()-ed
+into this process's address space).
+
+This huge overhead forced the creation of CONFIG_FUTEX_ROBUST so that
+normal kernels can turn it off, but worse than that: the overhead makes
+robust futexes impractical for any type of generic Linux distribution.
+
+So something had to be done.
+
+New approach to robust futexes
+------------------------------
+
+At the heart of this new approach there is a per-thread private list of
+robust locks that userspace is holding (maintained by glibc) - which
+userspace list is registered with the kernel via a new syscall [this
+registration happens at most once per thread lifetime]. At do_exit()
+time, the kernel checks this user-space list: are there any robust futex
+locks to be cleaned up?
+
+In the common case, at do_exit() time, there is no list registered, so
+the cost of robust futexes is just a simple current->robust_list != NULL
+comparison. If the thread has registered a list, then normally the list
+is empty. If the thread/process crashed or terminated in some incorrect
+way then the list might be non-empty: in this case the kernel carefully
+walks the list [not trusting it], and marks all locks that are owned by
+this thread with the FUTEX_OWNER_DIED bit, and wakes up one waiter (if
+any).
+
+The list is guaranteed to be private and per-thread at do_exit() time,
+so it can be accessed by the kernel in a lockless way.
+
+There is one race possible though: since adding to and removing from the
+list is done after the futex is acquired by glibc, there is a few
+instructions window for the thread (or process) to die there, leaving
+the futex hung. To protect against this possibility, userspace (glibc)
+also maintains a simple per-thread 'list_op_pending' field, to allow the
+kernel to clean up if the thread dies after acquiring the lock, but just
+before it could have added itself to the list. Glibc sets this
+list_op_pending field before it tries to acquire the futex, and clears
+it after the list-add (or list-remove) has finished.
+
+That's all that is needed - all the rest of robust-futex cleanup is done
+in userspace [just like with the previous patches].
+
+Ulrich Drepper has implemented the necessary glibc support for this new
+mechanism, which fully enables robust mutexes.
+
+Key differences of this userspace-list based approach, compared to the
+vma based method:
+
+ - it's much, much faster: at thread exit time, there's no need to loop
+ over every vma (!), which the VM-based method has to do. Only a very
+ simple 'is the list empty' op is done.
+
+ - no VM changes are needed - 'struct address_space' is left alone.
+
+ - no registration of individual locks is needed: robust mutexes don't
+ need any extra per-lock syscalls. Robust mutexes thus become a very
+ lightweight primitive - so they don't force the application designer
+ to do a hard choice between performance and robustness - robust
+ mutexes are just as fast.
+
+ - no per-lock kernel allocation happens.
+
+ - no resource limits are needed.
+
+ - no kernel-space recovery call (FUTEX_RECOVER) is needed.
+
+ - the implementation and the locking is "obvious", and there are no
+ interactions with the VM.
+
+Performance
+-----------
+
+I have benchmarked the time needed for the kernel to process a list of 1
+million (!) held locks, using the new method [on a 2GHz CPU]:
+
+ - with FUTEX_WAIT set [contended mutex]: 130 msecs
+ - without FUTEX_WAIT set [uncontended mutex]: 30 msecs
+
+I have also measured an approach where glibc does the lock notification
+[which it currently does for !pshared robust mutexes], and that took 256
+msecs - clearly slower, due to the 1 million FUTEX_WAKE syscalls
+userspace had to do.
+
+(1 million held locks are unheard of - we expect at most a handful of
+locks to be held at a time. Nevertheless it's nice to know that this
+approach scales nicely.)
+
+Implementation details
+----------------------
+
+The patch adds two new syscalls: one to register the userspace list, and
+one to query the registered list pointer::
+
+ asmlinkage long
+ sys_set_robust_list(struct robust_list_head __user *head,
+ size_t len);
+
+ asmlinkage long
+ sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
+ size_t __user *len_ptr);
+
+List registration is very fast: the pointer is simply stored in
+current->robust_list. [Note that in the future, if robust futexes become
+widespread, we could extend sys_clone() to register a robust-list head
+for new threads, without the need of another syscall.]
+
+So there is virtually zero overhead for tasks not using robust futexes,
+and even for robust futex users, there is only one extra syscall per
+thread lifetime, and the cleanup operation, if it happens, is fast and
+straightforward. The kernel doesn't have any internal distinction between
+robust and normal futexes.
+
+If a futex is found to be held at exit time, the kernel sets the
+following bit of the futex word::
+
+ #define FUTEX_OWNER_DIED 0x40000000
+
+and wakes up the next futex waiter (if any). User-space does the rest of
+the cleanup.
+
+Otherwise, robust futexes are acquired by glibc by putting the TID into
+the futex field atomically. Waiters set the FUTEX_WAITERS bit::
+
+ #define FUTEX_WAITERS 0x80000000
+
+and the remaining bits are for the TID.
+
+Testing, architecture support
+-----------------------------
+
+I've tested the new syscalls on x86 and x86_64, and have made sure the
+parsing of the userspace list is robust [ ;-) ] even if the list is
+deliberately corrupted.
+
+i386 and x86_64 syscalls are wired up at the moment, and Ulrich has
+tested the new glibc code (on x86_64 and i386), and it works for his
+robust-mutex testcases.
+
+All other architectures should build just fine too - but they won't have
+the new syscalls yet.
+
+Architectures need to implement the new futex_atomic_cmpxchg_inatomic()
+inline function before writing up the syscalls.
diff --git a/Documentation/locking/rt-mutex.rst b/Documentation/locking/rt-mutex.rst
index c365dc302081..3b5097a380e6 100644
--- a/Documentation/locking/rt-mutex.rst
+++ b/Documentation/locking/rt-mutex.rst
@@ -4,7 +4,7 @@ RT-mutex subsystem with PI support
RT-mutexes with priority inheritance are used to support PI-futexes,
which enable pthread_mutex_t priority inheritance attributes
-(PTHREAD_PRIO_INHERIT). [See Documentation/pi-futex.txt for more details
+(PTHREAD_PRIO_INHERIT). [See Documentation/locking/pi-futex.rst for more details
about PI-futexes.]
This technology was developed in the -rt tree and streamlined for