You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

24 KiB

Synchronization primitives in the Linux kernel. Part 1.

Introduction

This part opens new chapter in the linux-insides book. Timers and time management related stuff was described in the previous chapter. Now time to go next. As you may understand from the part's title, this chapter will describe synchronization primitives in the Linux kernel.

As always, before we will consider something synchronization related, we will try to know what is synchronization primitive in general. Actually, synchronization primitive is a software mechanism which provides ability to two or more parallel processes or threads to not execute simultaneously on the same segment of a code. For example let's look on the following piece of code:

mutex_lock(&clocksource_mutex);
...
...
...
clocksource_enqueue(cs);
clocksource_enqueue_watchdog(cs);
clocksource_select();
...
...
...
mutex_unlock(&clocksource_mutex);

from the kernel/time/clocksource.c source code file. This code is from the __clocksource_register_scale function which adds the given clocksource to the clock sources list. This function produces different operations on a list with registered clock sources. For example the clocksource_enqueue function adds the given clock source to the list with registered clocksources - clocksource_list. Note that these lines of code wrapped to two functions: mutex_lock and mutex_unlock which are takes one parameter - the clocksource_mutex in our case.

These functions represents locking and unlocking based on mutex synchronization primitive. As mutex_lock will be executed, it allows us to prevent situation when two or more threads will execute this code while the mutex_unlock will not be executed by process-owner of the mutex. In other words, we prevent parallel operations on a clocksource_list. Why do we need mutex here? What if two parallel processes will try to register a clock source. As we already know, the clocksource_enqueue function adds the given clock source to the clocksource_list list right after a clock source in the list which has the biggest rating (a registered clock source which has the highest frequency in the system):

static void clocksource_enqueue(struct clocksource *cs)
{
	struct list_head *entry = &clocksource_list;
	struct clocksource *tmp;

	list_for_each_entry(tmp, &clocksource_list, list)
		if (tmp->rating >= cs->rating)
			entry = &tmp->list;
	list_add(&cs->list, entry);
}

If two parallel processes will try to do it simultaneously, both process may found the same entry may occur race condition or in other words, the second process which will execute list_add, will overwrite a clock source from first thread.

Besides this simple example, synchronization primitives are ubiquitous in the Linux kernel. If we will go through the previous chapter or other chapters again or if we will look at the Linux kernel source code in general, we will meet many places like this. We will not consider how mutex is implemented in the Linux kernel. Actually, the Linux kernel provides a set of different synchronization primitives like:

  • mutex;
  • semaphores;
  • seqlocks;
  • atomic operations;
  • etc.

We will start this chapter from the spinlock.

Spinlocks in the Linux kernel.

The spinlock is a low-level synchronization mechanism which in simple words, represents a variable which can be in two states:

  • acquired;
  • released.

Each process which wants to acquire a spinlock, must write a value which represents spinlock acquired state to this variable and write spinlock released state to the variable. If a process tries to execute code which is protected by a spinlock, it will be locked while a process which holds this lock will release it. In this case all related operations must be atomic to prevent race conditions state. The spinlock is represented by the spinlock_t type in the Linux kernel. If we will look at the Linux kernel code, we will see that this type is widely used. The spinlock_t is defined as:

typedef struct spinlock {
        union {
              struct raw_spinlock rlock;
 
#ifdef CONFIG_DEBUG_LOCK_ALLOC
# define LOCK_PADSIZE (offsetof(struct raw_spinlock, dep_map))
                struct {
                        u8 __padding[LOCK_PADSIZE];
                        struct lockdep_map dep_map;
                };
#endif
        };
} spinlock_t;

and located in the include/linux/spinlock_types.h header file. We may see that its implementation depends on the state of the CONFIG_DEBUG_LOCK_ALLOC kernel configuration option. We will skip this now, because all debugging related stuff will be in the end of this part. So, if the CONFIG_DEBUG_LOCK_ALLOC kernel configuration option is disabled, the spinlock_t contains union with one field which is - raw_spinlock:

typedef struct spinlock {
        union {
              struct raw_spinlock rlock;
        };
} spinlock_t;

The raw_spinlock structure defined in the same header file and represents implementation of normal spinlock. Let's look how the raw_spinlock structure is defined:

typedef struct raw_spinlock {
        arch_spinlock_t raw_lock;
#ifdef CONFIG_GENERIC_LOCKBREAK
        unsigned int break_lock;
#endif
} raw_spinlock_t;

where the arch_spinlock_t represents architecture-specific spinlock implementation and the break_lock field which holds value - 1 in a case when one processor starts to wait while the lock is held on another processor on SMP systems. This allows prevent long time locking. As consider the x86_64 architecture in this books, so the arch_spinlock_t is defined in the arch/x86/include/asm/spinlock_types.h header file and looks:

#ifdef CONFIG_QUEUED_SPINLOCKS
#include <asm-generic/qspinlock_types.h>
#else
typedef struct arch_spinlock {
        union {
                __ticketpair_t head_tail;
                struct __raw_tickets {
                        __ticket_t head, tail;
                } tickets;
        };
} arch_spinlock_t;

As we may see, the definition of the arch_spinlock structure depends on the value of the CONFIG_QUEUED_SPINLOCKS kernel configuration option. This configuration option the Linux kernel supports spinlocks with queue. This special type of spinlocks which instead of acquired and released atomic values used atomic operation on a queue. If the CONFIG_QUEUED_SPINLOCKS kernel configuration option is enabled, the arch_spinlock_t will be represented by the following structure:

typedef struct qspinlock {
	atomic_t	val;
} arch_spinlock_t;

from the include/asm-generic/qspinlock_types.h header file.

We will not stop on this structures for now and before we will consider both arch_spinlock and the qspinlock, let's look at the operations on a spinlock. The Linux kernel provides following main operations on a spinlock:

  • spin_lock_init - produces initialization of the given spinlock;
  • spin_lock - acquires given spinlock;
  • spin_lock_bh - disables software interrupts and acquire given spinlock.
  • spin_lock_irqsave and spin_lock_irq - disable interrupts on local processor and preserve/not preserve previous interrupt state in the flags;
  • spin_unlock - releases given spinlock;
  • spin_unlock_bh - releases given spinlock and enables software interrupts;
  • spin_is_locked - returns the state of the given spinlock;
  • and etc.

Let's look on the implementation of the spin_lock_init macro. As I already wrote, this and other macro are defined in the include/linux/spinlock.h header file and the spin_lock_init macro looks:

#define spin_lock_init(_lock)		\
do {							                \
	spinlock_check(_lock);				        \
	raw_spin_lock_init(&(_lock)->rlock);		\
} while (0)

As we may see, the spin_lock_init macro takes a spinlock and executes two operations: check the given spinlock and execute the raw_spin_lock_init. The implementation of the spinlock_check is pretty easy, this function just returns the raw_spinlock_t of the given spinlock to be sure that we got exactly normal raw spinlock:

static __always_inline raw_spinlock_t *spinlock_check(spinlock_t *lock)
{
	return &lock->rlock;
}

The raw_spin_lock_init macro:

# define raw_spin_lock_init(lock)		\
do {                                                  \
    *(lock) = __RAW_SPIN_LOCK_UNLOCKED(lock);         \
} while (0)                                           \

assigns the value of the __RAW_SPIN_LOCK_UNLOCKED with the given spinlock to the given raw_spinlock_t. As we may understand from the name of the __RAW_SPIN_LOCK_UNLOCKED macro, this macro does initialization of the given spinlock and set it to released state. This macro defined in the include/linux/spinlock_types.h header file and expands to the following macros:

#define __RAW_SPIN_LOCK_UNLOCKED(lockname)      \
         (raw_spinlock_t) __RAW_SPIN_LOCK_INITIALIZER(lockname)

#define __RAW_SPIN_LOCK_INITIALIZER(lockname)   \
         {                                                      \
             .raw_lock = __ARCH_SPIN_LOCK_UNLOCKED,             \
             SPIN_DEBUG_INIT(lockname)                          \
             SPIN_DEP_MAP_INIT(lockname)                        \
         }

As I already wrote above, we will not consider stuff which is related to debugging of synchronization primitives. In this case we will not consider the SPIN_DEBUG_INIT and the SPIN_DEP_MAP_INIT macros. So the __RAW_SPINLOCK_UNLOCKED macro will be expanded to the:

*(&(_lock)->rlock) = __ARCH_SPIN_LOCK_UNLOCKED;

where the __ARCH_SPIN_LOCK_UNLOCKED is:

#define __ARCH_SPIN_LOCK_UNLOCKED       { { 0 } }

and:

#define __ARCH_SPIN_LOCK_UNLOCKED       { ATOMIC_INIT(0) }

for the x86_64 architecture. if the CONFIG_QUEUED_SPINLOCKS kernel configuration option is enabled. So, after the expansion of the spin_lock_init macro, a given spinlock will be initialized and its state will be - unlocked.

From this moment we know how to initialize a spinlock, now let's consider API which Linux kernel provides for manipulations of spinlocks. The first is:

static __always_inline void spin_lock(spinlock_t *lock)
{
	raw_spin_lock(&lock->rlock);
}

function which allows us to acquire a spinlock. The raw_spin_lock macro is defined in the same header file and expands to the call of the _raw_spin_lock function:

#define raw_spin_lock(lock)	_raw_spin_lock(lock)

As we may see in the include/linux/spinlock.h header file, definition of the _raw_spin_lock macro depends on the CONFIG_SMP kernel configuration parameter:

#if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
# include <linux/spinlock_api_smp.h>
#else
# include <linux/spinlock_api_up.h>
#endif

So, if the SMP is enabled in the Linux kernel, the _raw_spin_lock macro is defined in the arch/x86/include/asm/spinlock.h header file and looks like:

#define _raw_spin_lock(lock) __raw_spin_lock(lock)

The __raw_spin_lock function looks:

static inline void __raw_spin_lock(raw_spinlock_t *lock)
{
        preempt_disable();
        spin_acquire(&lock->dep_map, 0, 0, _RET_IP_);
        LOCK_CONTENDED(lock, do_raw_spin_trylock, do_raw_spin_lock);
}

As you may see, first of all we disable preemption by the call of the preempt_disable macro from the include/linux/preempt.h (more about this you may read in the ninth part of the Linux kernel initialization process chapter). When we will unlock the given spinlock, preemption will be enabled again:

static inline void __raw_spin_unlock(raw_spinlock_t *lock)
{
        ...
        ...
        ...
        preempt_enable();
}

We need to do this while a process is spinning on a lock, other processes must be prevented to preempt the process which acquired a lock. The spin_acquire macro which through a chain of other macros expands to the call of the:

#define spin_acquire(l, s, t, i)                lock_acquire_exclusive(l, s, t, NULL, i)
#define lock_acquire_exclusive(l, s, t, n, i)           lock_acquire(l, s, t, 0, 1, n, i)

lock_acquire function:

void lock_acquire(struct lockdep_map *lock, unsigned int subclass,
                  int trylock, int read, int check,
                  struct lockdep_map *nest_lock, unsigned long ip)
{
         unsigned long flags;

         if (unlikely(current->lockdep_recursion))
                return;
 
         raw_local_irq_save(flags);
         check_flags(flags);
 
         current->lockdep_recursion = 1;
         trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
         __lock_acquire(lock, subclass, trylock, read, check,
                        irqs_disabled_flags(flags), nest_lock, ip, 0, 0);
         current->lockdep_recursion = 0;
         raw_local_irq_restore(flags);
}

As I wrote above, we will not consider stuff here which is related to debugging or tracing. The main point of the lock_acquire function is to disable hardware interrupts by the call of the raw_local_irq_save macro, because the given spinlock might be acquired with enabled hardware interrupts. In this way the process will not be preempted. Note that in the end of the lock_acquire function we will enable hardware interrupts again with the help of the raw_local_irq_restore macro. As you already may guess, the main work will be in the __lock_acquire function which is defined in the kernel/locking/lockdep.c source code file.

The __lock_acquire function looks big. We will try to understand what does this function do, but not in this part. Actually this function mostly related to the Linux kernel lock validator and it is not topic of this part. If we will return to the definition of the __raw_spin_lock function, we will see that it contains the following definition in the end:

LOCK_CONTENDED(lock, do_raw_spin_trylock, do_raw_spin_lock);

The LOCK_CONTENDED macro is defined in the include/linux/lockdep.h header file and just calls the given function with the given spinlock:

#define LOCK_CONTENDED(_lock, try, lock) \
         lock(_lock)

In our case, the lock is do_raw_spin_lock function from the include/linux/spinlock.h header file and the _lock is the given raw_spinlock_t:

static inline void do_raw_spin_lock(raw_spinlock_t *lock) __acquires(lock)
{
        __acquire(lock);
         arch_spin_lock(&lock->raw_lock);
}

The __acquire here is just sparse related macro and we are not interesting in it in this moment. Location of the definition of the arch_spin_lock function depends on two things: the first is architecture of system and the second do we use queued spinlocks or not. In our case we consider only x86_64 architecture, so the definition of the arch_spin_lock is represented as the macro from the include/asm-generic/qspinlock.h header file:

#define arch_spin_lock(l)               queued_spin_lock(l)

if we are using queued spinlocks. Or in other case, the arch_spin_lock function is defined in the arch/x86/include/asm/spinlock.h header file. Now we will consider only normal spinlock and information related to queued spinlocks we will see later. Let's look again on the definition of the arch_spinlock structure, to understand implementation of the arch_spin_lock function:

typedef struct arch_spinlock {
         union {
                __ticketpair_t head_tail;
                struct __raw_tickets {
                        __ticket_t head, tail;
                } tickets;
        };
} arch_spinlock_t;

This variant of spinlock is called - ticket spinlock. As we may see, it consists from two parts. When lock is acquired, it increments a tail by one every time when a process wants to hold a spinlock. If the tail is not equal to head, the process will be locked, until values of these variables will not be equal. Let's look on the implementation of the arch_spin_lock function:

static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
{
        register struct __raw_tickets inc = { .tail = TICKET_LOCK_INC };

        inc = xadd(&lock->tickets, inc);

        if (likely(inc.head == inc.tail))
                goto out;

        for (;;) {
                 unsigned count = SPIN_THRESHOLD;

                 do {
                       inc.head = READ_ONCE(lock->tickets.head);
                       if (__tickets_equal(inc.head, inc.tail))
                                goto clear_slowpath;
                        cpu_relax();
                 } while (--count);
                 __ticket_lock_spinning(lock, inc.tail);
         }
clear_slowpath:
        __ticket_check_and_clear_slowpath(lock, inc.head);
out:
        barrier();
}

At the beginning of the arch_spin_lock function we can initialization of the __raw_tickets structure with tail - 1:

#define __TICKET_LOCK_INC       1

In the next line we execute xadd operation on the inc and lock->tickets. After this operation the inc will store value of the tickets of the given lock and the tickets.tail will be increased on inc or 1. The tail value was increased on 1 which means that one process started to try to hold a lock. In the next step we do the check that checks that head and tail have the same value. If these values are equal, this means that nobody holds lock and we go to the out label. In the end of the arch_spin_lock function we may see the barrier macro which represents barrier instruction which guarantees that compiler will not change order of operations that access memory (more about memory barriers you can read in the kernel documentation).

If one process held a lock and a second process started to execute the arch_spin_lock function, the head will not be equal to tail, because the tail will be greater than head on 1. In this way, process will occur in the loop. There will be comparison between head and the tail values at each loop iteration. If these values are not equal, the cpu_relax will be called which is just NOP instruction:

#define cpu_relax()     asm volatile("rep; nop")

and the next iteration of the loop will be started. If these values will be equal, this means that the process which held this lock, released this lock and the next process may acquire the lock.

The spin_unlock operation goes through the all macros/function as spin_lock, of course with unlock prefix. In the end the arch_spin_unlock function will be called. If we will look at the implementation of the arch_spin_unlock function, we will see that it increases head of the lock tickets list:

__add(&lock->tickets.head, TICKET_LOCK_INC, UNLOCK_LOCK_PREFIX);

In a combination of the spin_lock and spin_unlock, we get kind of queue where head contains an index number which maps currently executed process which holds a lock and the tail which contains an index number which maps last process which tried to hold the lock:

     +-------+       +-------+
     |       |       |       |
head |   7   | - - - |   7   | tail
     |       |       |       |
     +-------+       +-------+
                         |
                     +-------+
                     |       |
                     |   8   |
                     |       |
                     +-------+
                         |
                     +-------+
                     |       |
                     |   9   |
                     |       |
                     +-------+

That's all for now. We didn't cover spinlock API in full in this part, but I think that the main idea behind this concept must be clear now.

Conclusion

This concludes the first part covering synchronization primitives in the Linux kernel. In this part, we met first synchronization primitive spinlock provided by the Linux kernel. In the next part we will continue to dive into this interesting theme and will see other synchronization related stuff.

If you have questions or suggestions, feel free to ping me in twitter 0xAX, drop me email or just create issue.

Please note that English is not my first language and I am really sorry for any inconvenience. If you found any mistakes please send me PR to linux-insides.