cli-sti-removal.txt 3.91 KB
Newer Older
Linus Torvalds's avatar
Linus Torvalds committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133

#### cli()/sti() removal guide, started by Ingo Molnar <mingo@redhat.com>


as of 2.5.28, five popular macros have been removed on SMP, and
are being phased out on UP:

 cli(), sti(), save_flags(flags), save_flags_cli(flags), restore_flags(flags)

until now it was possible to protect driver code against interrupt
handlers via a cli(), but from now on other, more lightweight methods
have to be used for synchronization, such as spinlocks or semaphores.

for example, driver code that used to do something like:

	struct driver_data;

	irq_handler (...)
	{
		....
		driver_data.finish = 1;
		driver_data.new_work = 0;
		....
	}

	...

	ioctl_func (...)
	{
		...
		cli();
		...
		driver_data.finish = 0;
		driver_data.new_work = 2;
		...
		sti();
		...
	}

was SMP-correct because the cli() function ensured that no
interrupt handler (amongst them the above irq_handler()) function
would execute while the cli()-ed section is executing.

but from now on a more direct method of locking has to be used:

	spinlock_t driver_lock = SPIN_LOCK_UNLOCKED;
	struct driver_data;

	irq_handler (...)
	{
		unsigned long flags;
		....
		spin_lock_irqsave(&driver_lock, flags);
		....
		driver_data.finish = 1;
		driver_data.new_work = 0;
		....
		spin_unlock_irqrestore(&driver_lock, flags);
		....
	}

	...

	ioctl_func (...)
	{
		...
		spin_lock_irq(&driver_lock);
		...
		driver_data.finish = 0;
		driver_data.new_work = 2;
		...
		spin_unlock_irq(&driver_lock);
		...
	}

the above code has a number of advantages:

- the locking relation is easier to understand - actual lock usage
  pinpoints the critical sections. cli() usage is too opaque.
  Easier to understand means it's easier to debug.

- it's faster, because spinlocks are faster to acquire than the
  potentially heavily-used IRQ lock. Furthermore, your driver does
  not have to wait eg. for a big heavy SCSI interrupt to finish,
  because the driver_lock spinlock is only used by your driver.
  cli() on the other hand was used by many drivers, and extended
  the critical section to the whole IRQ handler function - creating
  serious lock contention.

 
to make the transition easier, we've still kept the cli(), sti(),
save_flags(), save_flags_cli() and restore_flags() macros defined
on UP systems - but their usage will be phased out until 2.6 is
released.

drivers that want to disable local interrupts (interrupts on the
current CPU), can use the following five macros:

  local_irq_disable(), local_irq_enable(), local_save_flags(flags),
  local_irq_save(flags), local_irq_restore(flags)

but beware, their meaning and semantics are much simpler, far from
that of the old cli(), sti(), save_flags(flags) and restore_flags(flags)
SMP meaning:

    local_irq_disable()       => turn local IRQs off

    local_irq_enable()        => turn local IRQs on

    local_save_flags(flags)   => save the current IRQ state into flags. The
                                 state can be on or off. (on some
                                 architectures there's even more bits in it.)

    local_irq_save(flags)     => save the current IRQ state into flags and
                                 disable interrupts.

    local_irq_restore(flags)  => restore the IRQ state from flags.

(local_irq_save can save both irqs on and irqs off state, and
local_irq_restore can restore into both irqs on and irqs off state.)

another related change is that synchronize_irq() now takes a parameter:
synchronize_irq(irq). This change too has the purpose of making SMP
synchronization more lightweight - this way you can wait for your own
interrupt handler to finish, no need to wait for other IRQ sources.


why were these changes done? The main reason was the architectural burden
of maintaining the cli()/sti() interface - it became a real problem. The
new interrupt system is much more streamlined, easier to understand, debug,
and it's also a bit faster - the same happened to it that will happen to
cli()/sti() using drivers once they convert to spinlocks :-)