linux/Documentation/io_ordering.txt
<<
>>
Prefs
   1On some platforms, so-called memory-mapped I/O is weakly ordered.  On such
   2platforms, driver writers are responsible for ensuring that I/O writes to
   3memory-mapped addresses on their device arrive in the order intended.  This is
   4typically done by reading a 'safe' device or bridge register, causing the I/O
   5chipset to flush pending writes to the device before any reads are posted.  A
   6driver would usually use this technique immediately prior to the exit of a
   7critical section of code protected by spinlocks.  This would ensure that
   8subsequent writes to I/O space arrived only after all prior writes (much like a
   9memory barrier op, mb(), only with respect to I/O).
  10
  11A more concrete example from a hypothetical device driver:
  12
  13        ...
  14CPU A:  spin_lock_irqsave(&dev_lock, flags)
  15CPU A:  val = readl(my_status);
  16CPU A:  ...
  17CPU A:  writel(newval, ring_ptr);
  18CPU A:  spin_unlock_irqrestore(&dev_lock, flags)
  19        ...
  20CPU B:  spin_lock_irqsave(&dev_lock, flags)
  21CPU B:  val = readl(my_status);
  22CPU B:  ...
  23CPU B:  writel(newval2, ring_ptr);
  24CPU B:  spin_unlock_irqrestore(&dev_lock, flags)
  25        ...
  26
  27In the case above, the device may receive newval2 before it receives newval,
  28which could cause problems.  Fixing it is easy enough though:
  29
  30        ...
  31CPU A:  spin_lock_irqsave(&dev_lock, flags)
  32CPU A:  val = readl(my_status);
  33CPU A:  ...
  34CPU A:  writel(newval, ring_ptr);
  35CPU A:  (void)readl(safe_register); /* maybe a config register? */
  36CPU A:  spin_unlock_irqrestore(&dev_lock, flags)
  37        ...
  38CPU B:  spin_lock_irqsave(&dev_lock, flags)
  39CPU B:  val = readl(my_status);
  40CPU B:  ...
  41CPU B:  writel(newval2, ring_ptr);
  42CPU B:  (void)readl(safe_register); /* maybe a config register? */
  43CPU B:  spin_unlock_irqrestore(&dev_lock, flags)
  44
  45Here, the reads from safe_register will cause the I/O chipset to flush any
  46pending writes before actually posting the read to the chipset, preventing
  47possible data corruption.
  48