Skip to content

L
linux
Commit f2d652b1 authored by Andrew Morton's avatar Andrew Morton Committed by Linus Torvalds
Browse files
Options

[PATCH] Kernel Locking Documentation update 

From: Rusty Russell <rusty@rustcorp.com.au>

Entirely revised, and largely rewritten.  Has a continuing example now, which
I think makes things clearer.  Also covers Read Copy Update.  This version
further deprecates rwlock_t, shuffles sections for better organization.
parent 8f6f7260
Show whitespace changes
Inline Side-by-side
Showing with 1322 additions and 451 deletions
Documentation/DocBook/kernel-locking.tmpl
View file @ f2d652b1
......@@ -6,8 +6,7 @@
<authorgroup>
<author>
<firstname>Paul</firstname>
<othername>Rusty</othername>
<firstname>Rusty</firstname>
<surname>Russell</surname>
<affiliation>
<address>
......@@ -18,8 +17,8 @@
</authorgroup>
<copyright>
<year>2000</year>
<holder>Paul Russell</holder>
<year>2003</year>
<holder>Rusty Russell</holder>
</copyright>
<legalnotice>
......@@ -58,16 +57,17 @@
<para>
Welcome, to Rusty's Remarkably Unreliable Guide to Kernel
Locking issues. This document describes the locking systems in
the Linux Kernel as we approach 2.4.
the Linux Kernel in 2.6.
</para>
<para>
It looks like <firstterm linkend="gloss-smp"><acronym>SMP</acronym>
</firstterm> is here to stay; so everyone hacking on the kernel
these days needs to know the fundamentals of concurrency and locking
for SMP.
With the wide availability of HyperThreading, and <firstterm
linkend="gloss-preemption">preemption </firstterm> in the Linux
Kernel, everyone hacking on the kernel needs to know the
fundamentals of concurrency and locking for
<firstterm linkend="gloss-smp"><acronym>SMP</acronym></firstterm>.
</para>
<sect1 id="races">
<chapter id="races">
<title>The Problem With Concurrency</title>
<para>
(Skip this if you know what a Race Condition is).
......@@ -169,14 +169,22 @@
</tgroup>
</table>
<sect1 id="race-condition">
<title>Race Conditions and Critical Regions</title>
<para>
This overlap, where what actually happens depends on the
relative timing of multiple tasks, is called a race condition.
This overlap, where the result depends on the
relative timing of multiple tasks, is called a <firstterm>race condition</firstterm>.
The piece of code containing the concurrency issue is called a
critical region. And especially since Linux starting running
<firstterm>critical region</firstterm>. And especially since Linux starting running
on SMP machines, they became one of the major issues in kernel
design and implementation.
</para>
<para>
Preemption can have the same effect, even if there is only one
CPU: by preempting one task during the critical region, we have
exactly the same race condition. In this case the thread which
preempts might run the critical region itself.
</para>
<para>
The solution is to recognize when these simultaneous accesses
occur, and use locks to make sure that only one instance can
......@@ -185,10 +193,28 @@
And then there are the unfriendly primitives, but I'll pretend
they don't exist.
</para>
</sect1>
</chapter>
<chapter id="locks">
<title>Locking in the Linux Kernel</title>
<para>
If I could give you one piece of advice: never sleep with anyone
crazier than yourself. But if I had to give you advice on
locking: <emphasis>keep it simple</emphasis>.
</para>
<para>
Be reluctant to introduce new locks.
</para>
<para>
Strangely enough, this last one is the exact reverse of my advice when
you <emphasis>have</emphasis> slept with someone crazier than yourself.
And you should think about getting a big dog.
</para>
<sect1 id="lock-intro">
<title>Two Main Types of Kernel Locks: Spinlocks and Semaphores</title>
<para>
......@@ -213,23 +239,32 @@
</para>
<para>
Neither type of lock is recursive: see
<xref linkend="techniques-deadlocks">.
<xref linkend="deadlock">.
</para>
</sect1>
<sect1 id="uniprocessor">
<title>Locks and Uniprocessor Kernels</title>
<para>
For kernels compiled without <symbol>CONFIG_SMP</symbol>, spinlocks
do not exist at all. This is an excellent design decision: when
no-one else can run at the same time, there is no reason to
have a lock at all.
For kernels compiled without <symbol>CONFIG_SMP</symbol>, and
without <symbol>CONFIG_PREEMPT</symbol> spinlocks do not exist at
all. This is an excellent design decision: when no-one else can
run at the same time, there is no reason to have a lock.
</para>
<para>
If the kernel is compiled without <symbol>CONFIG_SMP</symbol>,
but <symbol>CONFIG_PREEMPT</symbol> is set, then spinlocks
simply disable preemption, which is sufficient to prevent any
races. For most purposes, we can think of preemption as
equivalent to SMP, and not worry about it separately.
</para>
<para>
You should always test your locking code with <symbol>CONFIG_SMP</symbol>
enabled, even if you don't have an SMP test box, because it
will still catch some (simple) kinds of deadlock.
and <symbol>CONFIG_PREEMPT</symbol> enabled, even if you don't have an SMP test box, because it
will still catch some kinds of locking bugs.
</para>
<para>
......@@ -239,25 +274,6 @@
</para>
</sect1>
<sect1 id="rwlocks">
<title>Read/Write Lock Variants</title>
<para>
Both spinlocks and semaphores have read/write variants:
<type>rwlock_t</type> and <structname>struct rw_semaphore</structname>.
These divide users into two classes: the readers and the writers. If
you are only reading the data, you can get a read lock, but to write to
the data you need the write lock. Many people can hold a read lock,
but a writer must be sole holder.
</para>
<para>
This means much smoother locking if your code divides up
neatly along reader/writer lines. All the discussions below
also apply to read/write variants.
</para>
</sect1>
<sect1 id="usercontextlocking">
<title>Locking Only In User Context</title>
......@@ -289,17 +305,26 @@
</sect1>
<sect1 id="lock-user-bh">
<title>Locking Between User Context and BHs</title>
<title>Locking Between User Context and Softirqs</title>
<para>
If a <firstterm linkend="gloss-bh">bottom half</firstterm> shares
If a <firstterm linkend="gloss-softirq">softirq</firstterm> shares
data with user context, you have two problems. Firstly, the current
user context can be interrupted by a bottom half, and secondly, the
user context can be interrupted by a softirq, and secondly, the
critical region could be entered from another CPU. This is where
<function>spin_lock_bh()</function>
(<filename class=headerfile>include/linux/spinlock.h</filename>) is
used. It disables bottom halves on that CPU, then grabs the lock.
<function>spin_unlock_bh()</function> does the reverse.
used. It disables softirqs on that CPU, then grabs the lock.
<function>spin_unlock_bh()</function> does the reverse. (The
'_bh' suffix is a historical reference to "Bottom Halves", the
old name for software interrupts. It should really be
called spin_lock_softirq()' in a perfect world).
</para>
<para>
Note that you can also use <function>spin_lock_irq()</function>
or <function>spin_lock_irqsave()</function> here, which stop
hardware interrupts as well: see <xref linkend="hardirq-context">.
</para>
<para>
......@@ -307,70 +332,41 @@
</acronym></firstterm> as well: the spin lock vanishes, and this macro
simply becomes <function>local_bh_disable()</function>
(<filename class=headerfile>include/linux/interrupt.h</filename>), which
protects you from the bottom half being run.
protects you from the softirq being run.
</para>
</sect1>
<sect1 id="lock-user-tasklet">
<title>Locking Between User Context and Tasklets/Soft IRQs</title>
<title>Locking Between User Context and Tasklets</title>
<para>
This is exactly the same as above, because
<function>local_bh_disable()</function> actually disables all
softirqs and <firstterm linkend="gloss-tasklet">tasklets</firstterm>
on that CPU as well. It should really be
called `local_softirq_disable()', but the name has been preserved
for historical reasons. Similarly,
<function>spin_lock_bh()</function> would now be called
spin_lock_softirq() in a perfect world.
This is exactly the same as above, because <firstterm
linkend="gloss-tasklet">tasklets</firstterm> are actually run
from a softirq.
</para>
</sect1>
<sect1 id="lock-bh">
<title>Locking Between Bottom Halves</title>
<sect1 id="lock-user-timers">
<title>Locking Between User Context and Timers</title>
<para>
Sometimes a bottom half might want to share data with
another bottom half (especially remember that timers are run
off a bottom half).
This, too, is exactly the same as above, because <firstterm
linkend="gloss-timers">timers</firstterm> are actually run from
a softirq. From a locking point of view, tasklets and timers
are identical.
</para>
<sect2 id="lock-bh-same">
<title>The Same BH</title>
<para>
Since a bottom half is never run on two CPUs at once, you
don't need to worry about your bottom half being run twice
at once, even on SMP.
</para>
</sect2>
<sect2 id="lock-bh-different">
<title>Different BHs</title>
<para>
Since only one bottom half ever runs at a time once, you
don't need to worry about race conditions with other bottom
halves. Beware that things might change under you, however,
if someone changes your bottom half to a tasklet. If you
want to make your code future-proof, pretend you're already
running from a tasklet (see below), and doing the extra
locking. Of course, if it's five years before that happens,
you're gonna look like a damn fool.
</para>
</sect2>
</sect1>
<sect1 id="lock-tasklets">
<title>Locking Between Tasklets</title>
<title>Locking Between Tasklets/Timers</title>
<para>
Sometimes a tasklet might want to share data with another
tasklet, or a bottom half.
Sometimes a tasklet or timer might want to share data with
another tasklet or timer.
</para>
<sect2 id="lock-tasklets-same">
<title>The Same Tasklet</title>
<title>The Same Tasklet/Timer</title>
<para>
Since a tasklet is never run on two CPUs at once, you don't
need to worry about your tasklet being reentrant (running
......@@ -379,10 +375,10 @@
</sect2>
<sect2 id="lock-tasklets-different">
<title>Different Tasklets</title>
<title>Different Tasklets/Timers</title>
<para>
If another tasklet (or bottom half, such as a timer) wants
to share data with your tasklet, you will both need to use
If another tasklet/timer wants
to share data with your tasklet or timer , you will both need to use
<function>spin_lock()</function> and
<function>spin_unlock()</function> calls.
<function>spin_lock_bh()</function> is
......@@ -396,8 +392,8 @@
<title>Locking Between Softirqs</title>
<para>
Often a <firstterm linkend="gloss-softirq">softirq</firstterm> might
want to share data with itself, a tasklet, or a bottom half.
Often a softirq might
want to share data with itself or a tasklet/timer.
</para>
<sect2 id="lock-softirqs-same">
......@@ -423,8 +419,8 @@
<para>
You'll need to use <function>spin_lock()</function> and
<function>spin_unlock()</function> for shared data, whether it
be a timer (which can be running on a different CPU), bottom half,
tasklet or the same or another softirq.
be a timer, tasklet, different softirq or the same or another
softirq: any of them could be running on a different CPU.
</para>
</sect2>
</sect1>
......@@ -434,13 +430,13 @@
<title>Hard IRQ Context</title>
<para>
Hardware interrupts usually communicate with a bottom half,
Hardware interrupts usually communicate with a
tasklet or softirq. Frequently this involves putting work in a
queue, which the BH/softirq will take out.
queue, which the softirq will take out.
</para>
<sect1 id="hardirq-softirq">
<title>Locking Between Hard IRQ and Softirqs/Tasklets/BHs</title>
<title>Locking Between Hard IRQ and Softirqs/Tasklets</title>
<para>
If a hardware irq handler shares data with a softirq, you have
......@@ -452,6 +448,16 @@
the lock. <function>spin_unlock_irq()</function> does the reverse.
</para>
<para>
The irq handler does not to use
<function>spin_lock_irq()</function>, because the softirq cannot
run while the irq handler is running: it can use
<function>spin_lock()</function>, which is slightly faster. The
only exception would be if a different hardware irq handler uses
the same lock: <function>spin_lock_irq()</function> will stop
that from interrupting us.
</para>
<para>
This works perfectly for UP as well: the spin lock vanishes,
and this macro simply becomes <function>local_irq_disable()</function>
......@@ -468,60 +474,766 @@
interrupts are already off) and in softirqs (where the irq
disabling is required).
</para>
</sect1>
</chapter>
<chapter id="common-techniques">
<title>Common Techniques</title>
<para>
Note that softirqs (and hence tasklets and timers) are run on
return from hardware interrupts, so
<function>spin_lock_irq()</function> also stops these. In that
sense, <function>spin_lock_irqsave()</function> is the most
general and powerful locking function.
</para>
</sect1>
<sect1 id="hardirq-hardirq">
<title>Locking Between Two Hard IRQ Handlers</title>
<para>
This section lists some common dilemmas and the standard
solutions used in the Linux kernel code. If you use these,
people will find your code simpler to understand.
It is rare to have to share data between two IRQ handlers, but
if you do, <function>spin_lock_irqsave()</function> should be
used: it is architecture-specific whether all interrupts are
disabled inside irq handlers themselves.
</para>
</sect1>
</chapter>
<chapter id="cheatsheet">
<title>Cheat Sheet For Locking</title>
<para>
If I could give you one piece of advice: never sleep with anyone
crazier than yourself. But if I had to give you advice on
locking: <emphasis>keep it simple</emphasis>.
Pete Zaitcev gives the following summary:
</para>
<itemizedlist>
<listitem>
<para>
If you are in a process context (any syscall) and want to
lock other process out, use a semaphore. You can take a semaphore
and sleep (<function>copy_from_user*(</function> or
<function>kmalloc(x,GFP_KERNEL)</function>).
</para>
</listitem>
<listitem>
<para>
Otherwise (== data can be touched in an interrupt), use
<function>spin_lock_irqsave()</function> and
<function>spin_unlock_irqrestore()</function>.
</para>
</listitem>
<listitem>
<para>
Avoid holding spinlock for more than 5 lines of code and
across any function call (except accessors like
<function>readb</function>).
</para>
</listitem>
</itemizedlist>
<sect1 id="minimum-lock-reqirements">
<title>Table of Minimum Requirements</title>
<para> The following table lists the <emphasis>minimum</emphasis>
locking requirements between various contexts. In some cases,
the same context can only be running on one CPU at a time, so
no locking is required for that context (eg. a particular
thread can only run on one CPU at a time, but if it needs
shares data with another thread, locking is required).
</para>
<para>
Lock data, not code.
Remember the advice above: you can always use
<function>spin_lock_irqsave()</function>, which is a superset
of all other spinlock primitives.
</para>
<table>
<title>Table of Locking Requirements</title>
<TGROUP COLS=11>
<TBODY>
<ROW>
<ENTRY></ENTRY>
<ENTRY>IRQ Handler A</ENTRY>
<ENTRY>IRQ Handler B</ENTRY>
<ENTRY>Softirq A</ENTRY>
<ENTRY>Softirq B</ENTRY>
<ENTRY>Tasklet A</ENTRY>
<ENTRY>Tasklet B</ENTRY>
<ENTRY>Timer A</ENTRY>
<ENTRY>Timer B</ENTRY>
<ENTRY>User Context A</ENTRY>
<ENTRY>User Context B</ENTRY>
</ROW>
<ROW>
<ENTRY>IRQ Handler A</ENTRY>
<ENTRY>None</ENTRY>
</ROW>
<ROW>
<ENTRY>IRQ Handler B</ENTRY>
<ENTRY>spin_lock_irqsave</ENTRY>
<ENTRY>None</ENTRY>
</ROW>
<ROW>
<ENTRY>Softirq A</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock</ENTRY>
</ROW>
<ROW>
<ENTRY>Softirq B</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
</ROW>
<ROW>
<ENTRY>Tasklet A</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>None</ENTRY>
</ROW>
<ROW>
<ENTRY>Tasklet B</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>None</ENTRY>
</ROW>
<ROW>
<ENTRY>Timer A</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>None</ENTRY>
</ROW>
<ROW>
<ENTRY>Timer B</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>spin_lock</ENTRY>
<ENTRY>None</ENTRY>
</ROW>
<ROW>
<ENTRY>User Context A</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>None</ENTRY>
</ROW>
<ROW>
<ENTRY>User Context B</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_irq</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>spin_lock_bh</ENTRY>
<ENTRY>down_interruptible</ENTRY>
<ENTRY>None</ENTRY>
</ROW>
</TBODY>
</TGROUP>
</TABLE>
</chapter>
<chapter id="Examples">
<title>Common Examples</title>
<para>
Let's step through a simple example: a cache of number to name
mappings. The cache keeps a count of how often each of the objects is
used, and when it gets full, throws out the least used one.
</para>
<sect1 id="examples-usercontext">
<title>All In User Context</title>
<para>
For our first example, we assume that all operations are in user
context (ie. from system calls), so we can sleep. This means we can
use a semaphore to protect the cache and all the objects within
it. Here's the code:
</para>
<programlisting>
#include &lt;linux/list.h&gt;
#include &lt;linux/slab.h&gt;
#include &lt;linux/string.h&gt;
#include &lt;asm/semaphore.h&gt;
#include &lt;asm/errno.h&gt;
struct object
{
struct list_head list;
int id;
char name[32];
int popularity;
};
/* Protects the cache, cache_num, and the objects within it */
static DECLARE_MUTEX(cache_lock);
static LIST_HEAD(cache);
static unsigned int cache_num = 0;
#define MAX_CACHE_SIZE 10
/* Must be holding cache_lock */
static struct object *__cache_find(int id)
{
struct object *i;
list_for_each_entry(i, &amp;cache, list)
if (i-&gt;id == id) {
i-&gt;popularity++;
return i;
}
return NULL;
}
/* Must be holding cache_lock */
static void __cache_delete(struct object *obj)
{
BUG_ON(!obj);
list_del(&amp;obj-&gt;list);
kfree(obj);
cache_num--;
}
/* Must be holding cache_lock */
static void __cache_add(struct object *obj)
{
list_add(&amp;obj-&gt;list, &amp;cache);
if (++cache_num > MAX_CACHE_SIZE) {
struct object *i, *outcast = NULL;
list_for_each_entry(i, &amp;cache, list) {
if (!outcast || i-&gt;popularity &lt; outcast-&gt;popularity)
outcast = i;
}
__cache_delete(outcast);
}
}
int cache_add(int id, const char *name)
{
struct object *obj;
if ((obj = kmalloc(sizeof(*obj), GFP_KERNEL)) == NULL)
return -ENOMEM;
strlcpy(obj-&gt;name, name, sizeof(obj-&gt;name));
obj-&gt;id = id;
obj-&gt;popularity = 0;
down(&amp;cache_lock);
__cache_add(obj);
up(&amp;cache_lock);
return 0;
}
void cache_delete(int id)
{
down(&amp;cache_lock);
__cache_delete(__cache_find(id));
up(&amp;cache_lock);
}
int cache_find(int id, char *name)
{
struct object *obj;
int ret = -ENOENT;
down(&amp;cache_lock);
obj = __cache_find(id);
if (obj) {
ret = 0;
strcpy(name, obj-&gt;name);
}
up(&amp;cache_lock);
return ret;
}
</programlisting>
<para>
Be reluctant to introduce new locks.
Note that we always make sure we have the cache_lock when we add,
delete, or look up the cache: both the cache infrastructure itself and
the contents of the objects are protected by the lock. In this case
it's easy, since we copy the data for the user, and never let them
access the objects directly.
</para>
<para>
There is a slight (and common) optimization here: in
<function>cache_add</function> we set up the fields of the object
before grabbing the lock. This is safe, as no-one else can access it
until we put it in cache.
</para>
</sect1>
<sect1 id="examples-interrupt">
<title>Accessing From Interrupt Context</title>
<para>
Now consider the case where <function>cache_find</function> can be
called from interrupt context: either a hardware interrupt or a
softirq. An example would be a timer which deletes object from the
cache.
</para>
<para>
The change is shown below, in standard patch format: the
<symbol>-</symbol> are lines which are taken away, and the
<symbol>+</symbol> are lines which are added.
</para>
<programlisting>
--- cache.c.usercontext 2003-12-09 13:58:54.000000000 +1100
+++ cache.c.interrupt 2003-12-09 14:07:49.000000000 +1100
@@ -12,7 +12,7 @@
int popularity;
};
-static DECLARE_MUTEX(cache_lock);
+static spinlock_t cache_lock = SPIN_LOCK_UNLOCKED;
static LIST_HEAD(cache);
static unsigned int cache_num = 0;
#define MAX_CACHE_SIZE 10
@@ -55,6 +55,7 @@
int cache_add(int id, const char *name)
{
struct object *obj;
+ unsigned long flags;
if ((obj = kmalloc(sizeof(*obj), GFP_KERNEL)) == NULL)
return -ENOMEM;
@@ -63,30 +64,33 @@
obj-&gt;id = id;
obj-&gt;popularity = 0;
- down(&amp;cache_lock);
+ spin_lock_irqsave(&amp;cache_lock, flags);
__cache_add(obj);
- up(&amp;cache_lock);
+ spin_unlock_irqrestore(&amp;cache_lock, flags);
return 0;
}
void cache_delete(int id)
{
- down(&amp;cache_lock);
+ unsigned long flags;
+
+ spin_lock_irqsave(&amp;cache_lock, flags);
__cache_delete(__cache_find(id));
- up(&amp;cache_lock);
+ spin_unlock_irqrestore(&amp;cache_lock, flags);
}
int cache_find(int id, char *name)
{
struct object *obj;
int ret = -ENOENT;
+ unsigned long flags;
- down(&amp;cache_lock);
+ spin_lock_irqsave(&amp;cache_lock, flags);
obj = __cache_find(id);
if (obj) {
ret = 0;
strcpy(name, obj-&gt;name);
}
- up(&amp;cache_lock);
+ spin_unlock_irqrestore(&amp;cache_lock, flags);
return ret;
}
</programlisting>
<para>
Strangely enough, this is the exact reverse of my advice when
you <emphasis>have</emphasis> slept with someone crazier than yourself.
Note that the <function>spin_lock_irqsave</function> will turn off
interrupts if they are on, otherwise does nothing (if we are already
in an interrupt handler), hence these functions are safe to call from
any context.
</para>
<para>
Unfortunately, <function>cache_add</function> calls
<function>kmalloc</function> with the <symbol>GFP_KERNEL</symbol>
flag, which is only legal in user context. I have assumed that
<function>cache_add</function> is still only called in user context,
otherwise this should become a parameter to
<function>cache_add</function>.
</para>
</sect1>
<sect1 id="examples-refcnt">
<title>Exposing Objects Outside This File</title>
<para>
If our objects contained more information, it might not be sufficient
to copy the information in and out: other parts of the code might want
to keep pointers to these objects, for example, rather than looking up
the id every time. This produces two problems.
</para>
<para>
The first problem is that we use the <symbol>cache_lock</symbol> to
protect objects: we'd need to make this non-static so the rest of the
code can use it. This makes locking trickier, as it is no longer all
in one place.
</para>
<para>
The second problem is the lifetime problem: if another structure keeps
a pointer to an object, it presumably expects that pointer to remain
valid. Unfortunately, this is only guaranteed while you hold the
lock, otherwise someone might call <function>cache_delete</function>
and even worse, add another object, re-using the same address.
</para>
<para>
As there is only one lock, you can't hold it forever: no-one else would
get any work done.
</para>
<para>
The solution to this problem is to use a reference count: everyone who
has a pointer to the object increases it when they first get the
object, and drops the reference count when they're finished with it.
Whoever drops it to zero knows it is unused, and can actually delete it.
</para>
<para>
Here is the code:
</para>
<programlisting>
--- cache.c.interrupt 2003-12-09 14:25:43.000000000 +1100
+++ cache.c.refcnt 2003-12-09 14:33:05.000000000 +1100
@@ -7,6 +7,7 @@
struct object
{
struct list_head list;
+ unsigned int refcnt;
int id;
char name[32];
int popularity;
@@ -17,6 +18,35 @@
static unsigned int cache_num = 0;
#define MAX_CACHE_SIZE 10
+static void __object_put(struct object *obj)
+{
+ if (--obj-&gt;refcnt == 0)
+ kfree(obj);
+}
+
+static void __object_get(struct object *obj)
+{
+ obj-&gt;refcnt++;
+}
+
+void object_put(struct object *obj)
+{
+ unsigned long flags;
+
+ spin_lock_irqsave(&amp;cache_lock, flags);
+ __object_put(obj);
+ spin_unlock_irqrestore(&amp;cache_lock, flags);
+}
+
+void object_get(struct object *obj)
+{
+ unsigned long flags;
+
+ spin_lock_irqsave(&amp;cache_lock, flags);
+ __object_get(obj);
+ spin_unlock_irqrestore(&amp;cache_lock, flags);
+}
+
/* Must be holding cache_lock */
static struct object *__cache_find(int id)
{
@@ -35,6 +65,7 @@
{
BUG_ON(!obj);
list_del(&amp;obj-&gt;list);
+ __object_put(obj);
cache_num--;
}
@@ -63,6 +94,7 @@
strlcpy(obj-&gt;name, name, sizeof(obj-&gt;name));
obj-&gt;id = id;
obj-&gt;popularity = 0;
+ obj-&gt;refcnt = 1; /* The cache holds a reference */
spin_lock_irqsave(&amp;cache_lock, flags);
__cache_add(obj);
@@ -79,18 +111,15 @@
spin_unlock_irqrestore(&amp;cache_lock, flags);
}
-int cache_find(int id, char *name)
+struct object *cache_find(int id)
{
struct object *obj;
- int ret = -ENOENT;
unsigned long flags;
spin_lock_irqsave(&amp;cache_lock, flags);
obj = __cache_find(id);
- if (obj) {
- ret = 0;
- strcpy(name, obj-&gt;name);
- }
+ if (obj)
+ __object_get(obj);
spin_unlock_irqrestore(&amp;cache_lock, flags);
- return ret;
+ return obj;
}
</programlisting>
<para>
We encapsulate the reference counting in the standard 'get' and 'put'
functions. Now we can return the object itself from
<function>cache_find</function> which has the advantage that the user
can now sleep holding the object (eg. to
<function>copy_to_user</function> to name to userspace).
</para>
<para>
The other point to note is that I said a reference should be held for
every pointer to the object: thus the reference count is 1 when first
inserted into the cache. In some versions the framework does not hold
a reference count, but they are more complicated.
</para>
<sect2 id="examples-refcnt-atomic">
<title>Using Atomic Operations For The Reference Count</title>
<para>
In practice, <type>atomic_t</type> would usually be used for
<structfield>refcnt</structfield>. There are a number of atomic
operations defined in
<filename class=headerfile>include/asm/atomic.h</filename>: these are
guaranteed to be seen atomically from all CPUs in the system, so no
lock is required. In this case, it is simpler than using spinlocks,
although for anything non-trivial using spinlocks is clearer. The
<function>atomic_inc</function> and
<function>atomic_dec_and_test</function> are used instead of the
standard increment and decrement operators, and the lock is no longer
used to protect the reference count itself.
</para>
<programlisting>
--- cache.c.refcnt 2003-12-09 15:00:35.000000000 +1100
+++ cache.c.refcnt-atomic 2003-12-11 15:49:42.000000000 +1100
@@ -7,7 +7,7 @@
struct object
{
struct list_head list;
- unsigned int refcnt;
+ atomic_t refcnt;
int id;
char name[32];
int popularity;
@@ -18,33 +18,15 @@
static unsigned int cache_num = 0;
#define MAX_CACHE_SIZE 10
-static void __object_put(struct object *obj)
-{
- if (--obj-&gt;refcnt == 0)
- kfree(obj);
-}
-
-static void __object_get(struct object *obj)
-{
- obj-&gt;refcnt++;
-}
-
void object_put(struct object *obj)
{
- unsigned long flags;
-
- spin_lock_irqsave(&amp;cache_lock, flags);
- __object_put(obj);
- spin_unlock_irqrestore(&amp;cache_lock, flags);
+ if (atomic_dec_and_test(&amp;obj-&gt;refcnt))
+ kfree(obj);
}
<sect1 id="techniques-no-writers">
<title>No Writers in Interrupt Context</title>
void object_get(struct object *obj)
{
- unsigned long flags;
-
- spin_lock_irqsave(&amp;cache_lock, flags);
- __object_get(obj);
- spin_unlock_irqrestore(&amp;cache_lock, flags);
+ atomic_inc(&amp;obj-&gt;refcnt);
}
/* Must be holding cache_lock */
@@ -65,7 +47,7 @@
{
BUG_ON(!obj);
list_del(&amp;obj-&gt;list);
- __object_put(obj);
+ object_put(obj);
cache_num--;
}
@@ -94,7 +76,7 @@
strlcpy(obj-&gt;name, name, sizeof(obj-&gt;name));
obj-&gt;id = id;
obj-&gt;popularity = 0;
- obj-&gt;refcnt = 1; /* The cache holds a reference */
+ atomic_set(&amp;obj-&gt;refcnt, 1); /* The cache holds a reference */
spin_lock_irqsave(&amp;cache_lock, flags);
__cache_add(obj);
@@ -119,7 +101,7 @@
spin_lock_irqsave(&amp;cache_lock, flags);
obj = __cache_find(id);
if (obj)
- __object_get(obj);
+ object_get(obj);
spin_unlock_irqrestore(&amp;cache_lock, flags);
return obj;
}
</programlisting>
</sect1>
<sect1 id="examples-lock-per-obj">
<title>Protecting The Objects Themselves</title>
<para>
In these examples, we assumed that the objects (except the reference
counts) never changed once they are created. If we wanted to allow
the name to change, there are three possibilities:
</para>
<itemizedlist>
<listitem>
<para>
There is a fairly common case where an interrupt handler needs
access to a critical region, but does not need write access.
In this case, you do not need to use
<function>read_lock_irq()</function>, but only
<function>read_lock()</function> everywhere (since if an interrupt
occurs, the irq handler will only try to grab a read lock, which
won't deadlock). You will still need to use
<function>write_lock_irq()</function>.
You can make <symbol>cache_lock</symbol> non-static, and tell people
to grab that lock before changing the name in any object.
</para>
</listitem>
<listitem>
<para>
You can provide a <function>cache_obj_rename</function> which grabs
this lock and changes the name for the caller, and tell everyone to
use that function.
</para>
</listitem>
<listitem>
<para>
You can make the <symbol>cache_lock</symbol> protect only the cache
itself, and use another lock to protect the name.
</para>
</listitem>
</itemizedlist>
<para>
Similar logic applies to locking between softirqs/tasklets/BHs
which never need a write lock, and user context:
<function>read_lock()</function> and
<function>write_lock_bh()</function>.
Theoretically, you can make the locks as fine-grained as one lock for
every field, for every object. In practice, the most common variants
are:
</para>
<itemizedlist>
<listitem>
<para>
One lock which protects the infrastructure (the <symbol>cache</symbol>
list in this example) and all the objects. This is what we have done
so far.
</para>
</sect1>
</listitem>
<listitem>
<para>
One lock which protects the infrastructure (including the list
pointers inside the objects), and one lock inside the object which
protects the rest of that object.
</para>
</listitem>
<listitem>
<para>
Multiple locks to protect the infrastructure (eg. one lock per hash
chain), possibly with a separate per-object lock.
</para>
</listitem>
</itemizedlist>
<sect1 id="techniques-deadlocks">
<para>
Here is the "lock-per-object" implementation:
</para>
<programlisting>
--- cache.c.refcnt-atomic 2003-12-11 15:50:54.000000000 +1100
+++ cache.c.perobjectlock 2003-12-11 17:15:03.000000000 +1100
@@ -6,11 +6,17 @@
struct object
{
+ /* These two protected by cache_lock. */
struct list_head list;
+ int popularity;
+
atomic_t refcnt;
+
+ /* Doesn't change once created. */
int id;
+
+ spinlock_t lock; /* Protects the name */
char name[32];
- int popularity;
};
static spinlock_t cache_lock = SPIN_LOCK_UNLOCKED;
@@ -77,6 +84,7 @@
obj-&gt;id = id;
obj-&gt;popularity = 0;
atomic_set(&amp;obj-&gt;refcnt, 1); /* The cache holds a reference */
+ spin_lock_init(&amp;obj-&gt;lock);
spin_lock_irqsave(&amp;cache_lock, flags);
__cache_add(obj);
</programlisting>
<para>
Note that I decide that the <structfield>popularity</structfield>
count should be protected by the <symbol>cache_lock</symbol> rather
than the per-object lock: this is because it (like the
<structname>struct list_head</structname> inside the object) is
logically part of the infrastructure. This way, I don't need to grab
the lock of every object in <function>__cache_add</function> when
seeking the least popular.
</para>
<para>
I also decided that the <structfield>id</structfield> member is
unchangeable, so I don't need to grab each object lock in
<function>__cache_find()</function> to examine the
<structfield>id</structfield>: the object lock is only used by a
caller who wants to read or write the <structfield>name</structfield>
field.
</para>
<para>
Note also that I added a comment describing what data was protected by
which locks. This is extremely important, as it describes the runtime
behavior of the code, and can be hard to gain from just reading. And
as Alan Cox says, <quote>Lock data, not code</quote>.
</para>
</chapter>
<chapter id="common-problems">
<title>Common Problems</title>
<sect1 id="deadlock">
<title>Deadlock: Simple and Advanced</title>
<para>
......@@ -535,10 +1247,10 @@
<para>
For a slightly more complex case, imagine you have a region
shared by a bottom half and user context. If you use a
shared by a softirq and user context. If you use a
<function>spin_lock()</function> call to protect it, it is
possible that the user context will be interrupted by the bottom
half while it holds the lock, and the bottom half will then spin
possible that the user context will be interrupted by the softirq
while it holds the lock, and the softirq will then spin
forever trying to get the same lock.
</para>
......@@ -558,7 +1270,7 @@
</para>
<para>
A more complex problem is the so-called `deadly embrace',
A more complex problem is the so-called 'deadly embrace',
involving two or more locks. Say you have a hash table: each
entry in the table is a spinlock, and a chain of hashed
objects. Inside a softirq handler, you sometimes want to
......@@ -606,7 +1318,7 @@
their lock. It will look, smell, and feel like a crash.
</para>
<sect2 id="techs-deadlock-prevent">
<sect1 id="techs-deadlock-prevent">
<title>Preventing Deadlock</title>
<para>
......@@ -634,7 +1346,6 @@
will do?). Remember, the other programmers are out to get
you, so don't do this.
</para>
</sect2>
<sect2 id="techs-deadlock-overprevent">
<title>Overzealous Prevention Of Deadlocks</title>
......@@ -652,264 +1363,479 @@
</sect2>
</sect1>
<sect1 id="per-cpu">
<title>Per-CPU Data</title>
<sect1 id="racing-timers">
<title>Racing Timers: A Kernel Pastime</title>
<para>
A great technique for avoiding locking which is used fairly
widely is to duplicate information for each CPU. For example,
if you wanted to keep a count of a common condition, you could
use a spin lock and a single counter. Nice and simple.
Timers can produce their own special problems with races.
Consider a collection of objects (list, hash, etc) where each
object has a timer which is due to destroy it.
</para>
<para>
If that was too slow [it's probably not], you could instead
use a counter for each CPU [don't], then none of them need an
exclusive lock [you're wasting your time here]. To make sure
the CPUs don't have to synchronize caches all the time, align
the counters to cache boundaries by appending
`__cacheline_aligned' to the declaration
(<filename class=headerfile>include/linux/cache.h</filename>).
[Can't you think of anything better to do?]
If you want to destroy the entire collection (say on module
removal), you might do the following:
</para>
<para>
They will need a read lock to access their own counters,
however. That way you can use a write lock to grant exclusive
access to all of them at once, to tally them up.
</para>
</sect1>
<programlisting>
/* THIS CODE BAD BAD BAD BAD: IF IT WAS ANY WORSE IT WOULD USE
HUNGARIAN NOTATION */
spin_lock_bh(&amp;list_lock);
<sect1 id="brlock">
<title>Big Reader Locks</title>
while (list) {
struct foo *next = list-&gt;next;
del_timer(&amp;list-&gt;timer);
kfree(list);
list = next;
}
<para>
A classic example of per-CPU information is Ingo's `big
reader' locks
(<filename class=headerfile>linux/include/brlock.h</filename>). These
use the Per-CPU Data techniques described above to create a lock which
is very fast to get a read lock, but agonizingly slow for a write
lock.
</para>
spin_unlock_bh(&amp;list_lock);
</programlisting>
<para>
Fortunately, there are a limited number of these locks
available: you have to go through a strict interview process
to get one.
Sooner or later, this will crash on SMP, because a timer can
have just gone off before the <function>spin_lock_bh()</function>,
and it will only get the lock after we
<function>spin_unlock_bh()</function>, and then try to free
the element (which has already been freed!).
</para>
</sect1>
<sect1 id="lock-avoidance-rw">
<title>Avoiding Locks: Read And Write Ordering</title>
<para>
Sometimes it is possible to avoid locking. Consider the
following case from the 2.2 firewall code, which inserted an
element into a single linked list in user context:
This can be avoided by checking the result of
<function>del_timer()</function>: if it returns
<returnvalue>1</returnvalue>, the timer has been deleted.
If <returnvalue>0</returnvalue>, it means (in this
case) that it is currently running, so we can do:
</para>
<programlisting>
new-&gt;next = i-&gt;next;
i-&gt;next = new;
</programlisting>
<para>
Here the author (Alan Cox, who knows what he's doing) assumes
that the pointer assignments are atomic. This is important,
because networking packets would traverse this list on bottom
halves without a lock. Depending on their exact timing, they
would either see the new element in the list with a valid
<structfield>next</structfield> pointer, or it would not be in the
list yet. A lock is still required against other CPUs inserting
or deleting from the list, of course.
</para>
retry:
spin_lock_bh(&amp;list_lock);
<para>
Of course, the writes <emphasis>must</emphasis> be in this
order, otherwise the new element appears in the list with an
invalid <structfield>next</structfield> pointer, and any other
CPU iterating at the wrong time will jump through it into garbage.
Because modern CPUs reorder, Alan's code actually read as follows:
</para>
while (list) {
struct foo *next = list-&gt;next;
if (!del_timer(&amp;list-&gt;timer)) {
/* Give timer a chance to delete this */
spin_unlock_bh(&amp;list_lock);
goto retry;
}
kfree(list);
list = next;
}
<programlisting>
new-&gt;next = i-&gt;next;
wmb();
i-&gt;next = new;
spin_unlock_bh(&amp;list_lock);
</programlisting>
<para>
The <function>wmb()</function> is a write memory barrier
(<filename class=headerfile>include/asm/system.h</filename>): neither
the compiler nor the CPU will allow any writes to memory after the
<function>wmb()</function> to be visible to other hardware
before any of the writes before the <function>wmb()</function>.
Another common problem is deleting timers which restart
themselves (by calling <function>add_timer()</function> at the end
of their timer function). Because this is a fairly common case
which is prone to races, you should use <function>del_timer_sync()</function>
(<filename class=headerfile>include/linux/timer.h</filename>)
to handle this case. It returns the number of times the timer
had to be deleted before we finally stopped it from adding itself back
in.
</para>
</sect1>
<para>
As i386 does not do write reordering, this bug would never
show up on that platform. On other SMP platforms, however, it
will.
</para>
<sect1 id="sparc">
<title>The Fucked Up Sparc</title>
<para>
There is also <function>rmb()</function> for read ordering: to ensure
any previous variable reads occur before following reads. The simple
<function>mb()</function> macro combines both
<function>rmb()</function> and <function>wmb()</function>.
Alan Cox says <quote>the irq disable/enable is in the register
window on a sparc</quote>. Andi Kleen says <quote>when you do
restore_flags in a different function you mess up all the
register windows</quote>.
</para>
<para>
Some atomic operations are defined to act as a memory barrier
(ie. as per the <function>mb()</function> macro, but if in
doubt, be explicit.
<!-- Rusty Russell 2 May 2001, 2.4.4 -->
Also,
spinlock operations act as partial barriers: operations after
gaining a spinlock will never be moved to precede the
<function>spin_lock()</function> call, and operations before
releasing a spinlock will never be moved after the
<function>spin_unlock()</function> call.
<!-- Manfred Spraul <manfreds@colorfullife.com>
24 May 2000 2.3.99-pre9 -->
So never pass the flags word set by
<function>spin_lock_irqsave()</function> and brethren to another
function (unless it's declared <type>inline</type>). Usually no-one
does this, but now you've been warned. Dave Miller can never do
anything in a straightforward manner (I can say that, because I have
pictures of him and a certain PowerPC maintainer in a compromising
position).
</para>
</sect1>
<sect1 id="lock-avoidance-atomic-ops">
<title>Avoiding Locks: Atomic Operations</title>
</chapter>
<chapter id="Efficiency">
<title>Locking Speed</title>
<para>
There are three main things to worry about when considering speed of
some code which does locking. First is concurrency: how many things
are going to be waiting while someone else is holding a lock. Second
is the time taken to actually acquire and release an uncontended lock.
Third is using fewer, or smarter locks. I'm assuming that the lock is
used fairly often: otherwise, you wouldn't be concerned about
efficiency.
</para>
<para>
Concurrency depends on how long the lock is usually held: you should
hold the lock for as long as needed, but no longer. In the cache
example, we always create the object without the lock held, and then
grab the lock only when we are ready to insert it in the list.
</para>
<para>
Acquisition times depend on how much damage the lock operations do to
the pipeline (pipeline stalls) and how likely it is that this CPU was
the last one to grab the lock (ie. is the lock cache-hot for this
CPU): on a machine with more CPUs, this likelihood drops fast.
Consider a 700MHz Intel Pentium III: an instruction takes about 0.7ns,
an atomic increment takes about 58ns, a lock which is cache-hot on
this CPU takes 160ns, and a cacheline transfer from another CPU takes
an additional 170 to 360ns. (These figures from Paul McKenney's
<ulink url="http://www.linuxjournal.com/article.php?sid=6993"> Linux
Journal RCU article</ulink>).
</para>
<para>
These two aims conflict: holding a lock for a short time might be done
by splitting locks into parts (such as in our final per-object-lock
example), but this increases the number of lock acquisitions, and the
results are often slower than having a single lock. This is another
reason to advocate locking simplicity.
</para>
<para>
The third concern is addressed below: there are some methods to reduce
the amount of locking which needs to be done.
</para>
<sect1 id="efficiency-rwlocks">
<title>Read/Write Lock Variants</title>
<para>
There are a number of atomic operations defined in
<filename class=headerfile>include/asm/atomic.h</filename>: these
are guaranteed to be seen atomically from all CPUs in the system, thus
avoiding races. If your shared data consists of a single counter, say,
these operations might be simpler than using spinlocks (although for
anything non-trivial using spinlocks is clearer).
Both spinlocks and semaphores have read/write variants:
<type>rwlock_t</type> and <structname>struct rw_semaphore</structname>.
These divide users into two classes: the readers and the writers. If
you are only reading the data, you can get a read lock, but to write to
the data you need the write lock. Many people can hold a read lock,
but a writer must be sole holder.
</para>
<para>
Note that the atomic operations do in general not act as memory
barriers. Instead you can insert a memory barrier before or
after <function>atomic_inc()</function> or
<function>atomic_dec()</function> by inserting
<function>smp_mb__before_atomic_inc()</function>,
<function>smp_mb__after_atomic_inc()</function>,
<function>smp_mb__before_atomic_dec()</function> or
<function>smp_mb__after_atomic_dec()</function>
respectively. The advantage of using those macros instead of
<function>smp_mb()</function> is, that they are cheaper on some
platforms.
<!-- Sebastian Wilhelmi <seppi@seppi.de> 2002-03-04 -->
If your code divides neatly along reader/writer lines (as our
cache code does), and the lock is held by readers for
significant lengths of time, using these locks can help. They
are slightly slower than the normal locks though, so in practice
<type>rwlock_t</type> is not usually worthwhile.
</para>
</sect1>
<sect1 id="ref-counts">
<title>Protecting A Collection of Objects: Reference Counts</title>
<sect1 id="efficiency-read-copy-update">
<title>Avoiding Locks: Read Copy Update</title>
<para>
Locking a collection of objects is fairly easy: you get a
single spinlock, and you make sure you grab it before
searching, adding or deleting an object.
There is a special method of read/write locking called Read Copy
Update. Using RCU, the readers can avoid taking a lock
altogether: as we expect our cache to be read more often than
updated (otherwise the cache is a waste of time), it is a
candidate for this optimization.
</para>
<para>
The purpose of this lock is not to protect the individual
objects: you might have a separate lock inside each one for
that. It is to protect the <emphasis>data structure
containing the objects</emphasis> from race conditions. Often
the same lock is used to protect the contents of all the
objects as well, for simplicity, but they are inherently
orthogonal (and many other big words designed to confuse).
How do we get rid of read locks? Getting rid of read locks
means that writers may be changing the list underneath the
readers. That is actually quite simple: we can read a linked
list while an element is being added if the writer adds the
element very carefully. For example, adding
<symbol>new</symbol> to a single linked list called
<symbol>list</symbol>:
</para>
<para>
Changing this to a read-write lock will often help markedly if
reads are far more common that writes. If not, there is
another approach you can use to reduce the time the lock is
held: reference counts.
</para>
<programlisting>
new-&gt;next = list-&gt;next;
wmb();
list-&gt;next = new;
</programlisting>
<para>
In this approach, an object has an owner, who sets the
reference count to one. Whenever you get a pointer to the
object, you increment the reference count (a `get' operation).
Whenever you relinquish a pointer, you decrement the reference
count (a `put' operation). When the owner wants to destroy
it, they mark it dead, and do a put.
The <function>wmb()</function> is a write memory barrier. It
ensures that the first operation (setting the new element's
<symbol>next</symbol> pointer) is complete and will be seen by
all CPUs, before the second operation is (putting the new
element into the list). This is important, since modern
compilers and modern CPUs can both reorder instructions unless
told otherwise: we want a reader to either not see the new
element at all, or see the new element with the
<symbol>next</symbol> pointer correctly pointing at the rest of
the list.
</para>
<para>
Whoever drops the reference count to zero (usually implemented
with <function>atomic_dec_and_test()</function>) actually cleans
up and frees the object.
Fortunately, there is a function to do this for standard
<structname>struct list_head</structname> lists:
<function>list_add_rcu()</function>
(<filename>include/linux/list.h</filename>).
</para>
<para>
This means that you are guaranteed that the object won't
vanish underneath you, even though you no longer have a lock
for the collection.
Removing an element from the list is even simpler: we replace
the pointer to the old element with a pointer to its successor,
and readers will either see it, or skip over it.
</para>
<programlisting>
list-&gt;next = old-&gt;next;
</programlisting>
<para>
Here's some skeleton code:
</para>
There is <function>list_del_rcu()</function>
(<filename>include/linux/list.h</filename>) which does this (the
normal version poisons the old object, which we don't want).
</para>
<para>
The reader must also be careful: some CPUs can look through the
<symbol>next</symbol> pointer to start reading the contents of
the next element early, but don't realize that the pre-fetched
contents is wrong when the <symbol>next</symbol> pointer changes
underneath them. Once again, there is a
<function>list_for_each_entry_rcu()</function>
(<filename>include/linux/list.h</filename>) to help you. Of
course, writers can just use
<function>list_for_each_entry()</function>, since there cannot
be two simultaneous writers.
</para>
<para>
Our final dilemma is this: when can we actually destroy the
removed element? Remember, a reader might be stepping through
this element in the list right now: it we free this element and
the <symbol>next</symbol> pointer changes, the reader will jump
off into garbage and crash. We need to wait until we know that
all the readers who were traversing the list when we deleted the
element are finished. We use <function>call_rcu()</function> to
register a callback which will actually destroy the object once
the readers are finished.
</para>
<para>
But how does Read Copy Update know when the readers are
finished? The method is this: firstly, the readers always
traverse the list inside
<function>rcu_read_lock()</function>/<function>rcu_read_unlock()</function>
pairs: these simply disable preemption so the reader won't go to
sleep while reading the list.
</para>
<para>
RCU then waits until every other CPU has slept at least once:
since readers cannot sleep, we know that any readers which were
traversing the list during the deletion are finished, and the
callback is triggered. The real Read Copy Update code is a
little more optimized than this, but this is the fundamental
idea.
</para>
<programlisting>
--- cache.c.perobjectlock 2003-12-11 17:15:03.000000000 +1100
+++ cache.c.rcupdate 2003-12-11 17:55:14.000000000 +1100
@@ -1,15 +1,18 @@
#include &lt;linux/list.h&gt;
#include &lt;linux/slab.h&gt;
#include &lt;linux/string.h&gt;
+#include &lt;linux/rcupdate.h&gt;
#include &lt;asm/semaphore.h&gt;
#include &lt;asm/errno.h&gt;
struct object
{
- /* These two protected by cache_lock. */
+ /* This is protected by RCU */
struct list_head list;
int popularity;
<programlisting>
void create_foo(struct foo *x)
+ struct rcu_head rcu;
+
atomic_t refcnt;
/* Doesn't change once created. */
@@ -40,7 +43,7 @@
{
atomic_set(&amp;x-&gt;use, 1);
spin_lock_bh(&amp;list_lock);
... insert in list ...
spin_unlock_bh(&amp;list_lock);
struct object *i;
- list_for_each_entry(i, &amp;cache, list) {
+ list_for_each_entry_rcu(i, &amp;cache, list) {
if (i-&gt;id == id) {
i-&gt;popularity++;
return i;
@@ -49,19 +52,25 @@
return NULL;
}
struct foo *get_foo(int desc)
+/* Final discard done once we know no readers are looking. */
+static void cache_delete_rcu(void *arg)
+{
+ object_put(arg);
+}
+
/* Must be holding cache_lock */
static void __cache_delete(struct object *obj)
{
struct foo *ret;
spin_lock_bh(&amp;list_lock);
... find in list ...
if (ret) atomic_inc(&amp;ret-&gt;use);
spin_unlock_bh(&amp;list_lock);
return ret;
BUG_ON(!obj);
- list_del(&amp;obj-&gt;list);
- object_put(obj);
+ list_del_rcu(&amp;obj-&gt;list);
cache_num--;
+ call_rcu(&amp;obj-&gt;rcu, cache_delete_rcu, obj);
}
void put_foo(struct foo *x)
/* Must be holding cache_lock */
static void __cache_add(struct object *obj)
{
- list_add(&amp;obj-&gt;list, &amp;cache);
+ list_add_rcu(&amp;obj-&gt;list, &amp;cache);
if (++cache_num > MAX_CACHE_SIZE) {
struct object *i, *outcast = NULL;
list_for_each_entry(i, &amp;cache, list) {
@@ -85,6 +94,7 @@
obj-&gt;popularity = 0;
atomic_set(&amp;obj-&gt;refcnt, 1); /* The cache holds a reference */
spin_lock_init(&amp;obj-&gt;lock);
+ INIT_RCU_HEAD(&amp;obj-&gt;rcu);
spin_lock_irqsave(&amp;cache_lock, flags);
__cache_add(obj);
@@ -104,12 +114,11 @@
struct object *cache_find(int id)
{
if (atomic_dec_and_test(&amp;x-&gt;use))
kfree(foo);
struct object *obj;
- unsigned long flags;
- spin_lock_irqsave(&amp;cache_lock, flags);
+ rcu_read_lock();
obj = __cache_find(id);
if (obj)
object_get(obj);
- spin_unlock_irqrestore(&amp;cache_lock, flags);
+ rcu_read_unlock();
return obj;
}
</programlisting>
<para>
Note that the reader will alter the
<structfield>popularity</structfield> member in
<function>__cache_find()</function>, and now it doesn't hold a lock.
One solution would be to make it an <type>atomic_t</type>, but for
this usage, we don't really care about races: an approximate result is
good enough, so I didn't change it.
</para>
<para>
The result is that <function>cache_find()</function> requires no
synchronization with any other functions, so is almost as fast on SMP
as it would be on UP.
</para>
<para>
There is a furthur optimization possible here: remember our original
cache code, where there were no reference counts and the caller simply
held the lock whenever using the object? This is still possible: if
you hold the lock, noone can delete the object, so you don't need to
get and put the reference count.
</para>
<para>
Now, because the 'read lock' in RCU is simply disabling preemption, a
caller which always has preemption disabled between calling
<function>cache_find()</function> and
<function>object_put()</function> does not need to actually get and
put the reference count: we could expose
<function>__cache_find()</function> by making it non-static, and
such callers could simply call that.
</para>
<para>
The benefit here is that the reference count is not written to: the
object is not altered in any way, which is much faster on SMP
machines due to caching.
</para>
</sect1>
void destroy_foo(struct foo *x)
{
spin_lock_bh(&amp;list_lock);
... remove from list ...
spin_unlock_bh(&amp;list_lock);
<sect1 id="per-cpu">
<title>Per-CPU Data</title>
put_foo(x);
}
</programlisting>
<para>
Another technique for avoiding locking which is used fairly
widely is to duplicate information for each CPU. For example,
if you wanted to keep a count of a common condition, you could
use a spin lock and a single counter. Nice and simple.
</para>
<sect2 id="helpful-macros">
<title>Macros To Help You</title>
<para>
There are a set of debugging macros tucked inside
<filename class=headerfile>include/linux/netfilter_ipv4/lockhelp.h</filename>
and <filename class=headerfile>listhelp.h</filename>: these are very
useful for ensuring that locks are held in the right places to protect
infrastructure.
If that was too slow (it's usually not, but if you've got a
really big machine to test on and can show that it is), you
could instead use a counter for each CPU, then none of them need
an exclusive lock. See <function>DEFINE_PER_CPU()</function>,
<function>get_cpu_var()</function> and
<function>put_cpu_var()</function>
(<filename class=headerfile>include/linux/percpu.h</filename>).
</para>
<para>
Of particular use for simple per-cpu counters is the
<type>local_t</type> type, and the
<function>cpu_local_inc()</function> and related functions,
which are more efficient than simple code on some architectures
(<filename class=headerfile>include/asm/local.h</filename>).
</para>
<para>
Note that there is no simple, reliable way of getting an exact
value of such a counter, without introducing more locks. This
is not a problem for some uses.
</para>
</sect2>
</sect1>
<sect1 id="sleeping-things">
<title>Things Which Sleep</title>
<sect1 id="mostly-hardirq">
<title>Data Which Mostly Used By An IRQ Handler</title>
<para>
If data is always accessed from within the same IRQ handler, you
don't need a lock at all: the kernel already guarantees that the
irq handler will not run simultaneously on multiple CPUs.
</para>
<para>
Manfred Spraul points out that you can still do this, even if
the data is very occasionally accessed in user context or
softirqs/tasklets. The irq handler doesn't use a lock, and
all other accesses are done as so:
</para>
<programlisting>
spin_lock(&amp;lock);
disable_irq(irq);
...
enable_irq(irq);
spin_unlock(&amp;lock);
</programlisting>
<para>
You can never call the following routines while holding a
spinlock, as they may sleep. This also means you need to be in
user context.
The <function>disable_irq()</function> prevents the irq handler
from running (and waits for it to finish if it's currently
running on other CPUs). The spinlock prevents any other
accesses happening at the same time. Naturally, this is slower
than just a <function>spin_lock_irq()</function> call, so it
only makes sense if this type of access happens extremely
rarely.
</para>
</sect1>
</chapter>
<chapter id="sleeping-things">
<title>What Functions Are Safe To Call From Interrupts?</title>
<para>
Many functions in the kernel sleep (ie. call schedule())
directly or indirectly: you can never call them while holding a
spinlock, or with preemption disabled. This also means you need
to be in user context: calling them from an interrupt is illegal.
</para>
<sect1 id="sleeping">
<title>Some Functions Which Sleep</title>
<para>
The most common ones are listed below, but you usually have to
read the code to find out if other calls are safe. If everyone
else who calls it can sleep, you probably need to be able to
sleep, too. In particular, registration and deregistration
functions usually expect to be called from user context, and can
sleep.
</para>
<itemizedlist>
......@@ -961,106 +1887,31 @@
</listitem>
</itemizedlist>
<para>
<function>printk()</function> can be called in
<emphasis>any</emphasis> context, interestingly enough.
</para>
</sect1>
<sect1 id="sparc">
<title>The Fucked Up Sparc</title>
<para>
Alan Cox says <quote>the irq disable/enable is in the register
window on a sparc</quote>. Andi Kleen says <quote>when you do
restore_flags in a different function you mess up all the
register windows</quote>.
</para>
<para>
So never pass the flags word set by
<function>spin_lock_irqsave()</function> and brethren to another
function (unless it's declared <type>inline</type>. Usually no-one
does this, but now you've been warned. Dave Miller can never do
anything in a straightforward manner (I can say that, because I have
pictures of him and a certain PowerPC maintainer in a compromising
position).
</para>
</sect1>
<sect1 id="racing-timers">
<title>Racing Timers: A Kernel Pastime</title>
<sect1 id="dont-sleep">
<title>Some Functions Which Don't Sleep</title>
<para>
Timers can produce their own special problems with races.
Consider a collection of objects (list, hash, etc) where each
object has a timer which is due to destroy it.
Some functions are safe to call from any context, or holding
almost any lock.
</para>
<itemizedlist>
<listitem>
<para>
If you want to destroy the entire collection (say on module
removal), you might do the following:
</para>
<programlisting>
/* THIS CODE BAD BAD BAD BAD: IF IT WAS ANY WORSE IT WOULD USE
HUNGARIAN NOTATION */
spin_lock_bh(&amp;list_lock);
while (list) {
struct foo *next = list-&gt;next;
del_timer(&amp;list-&gt;timer);
kfree(list);
list = next;
}
spin_unlock_bh(&amp;list_lock);
</programlisting>
<para>
Sooner or later, this will crash on SMP, because a timer can
have just gone off before the <function>spin_lock_bh()</function>,
and it will only get the lock after we
<function>spin_unlock_bh()</function>, and then try to free
the element (which has already been freed!).
<function>printk()</function>
</para>
</listitem>
<listitem>
<para>
This can be avoided by checking the result of
<function>del_timer()</function>: if it returns
<returnvalue>1</returnvalue>, the timer has been deleted.
If <returnvalue>0</returnvalue>, it means (in this
case) that it is currently running, so we can do:
<function>kfree()</function>
</para>
<programlisting>
retry:
spin_lock_bh(&amp;list_lock);
while (list) {
struct foo *next = list-&gt;next;
if (!del_timer(&amp;list-&gt;timer)) {
/* Give timer a chance to delete this */
spin_unlock_bh(&amp;list_lock);
goto retry;
}
kfree(list);
list = next;
}
spin_unlock_bh(&amp;list_lock);
</programlisting>
</listitem>
<listitem>
<para>
Another common problem is deleting timers which restart
themselves (by calling <function>add_timer()</function> at the end
of their timer function). Because this is a fairly common case
which is prone to races, you should use <function>del_timer_sync()</function>
(<filename class=headerfile>include/linux/timer.h</filename>)
to handle this case. It returns the number of times the timer
had to be deleted before we finally stopped it from adding itself back
in.
<function>add_timer()</function> and <function>del_timer()</function>
</para>
</listitem>
</itemizedlist>
</sect1>
</chapter>
......@@ -1101,8 +1952,9 @@
<para>
Thanks to Martin Pool, Philipp Rumpf, Stephen Rothwell, Paul
Mackerras, Ruedi Aschwanden, Alan Cox, Manfred Spraul and Tim
Waugh for proofreading, correcting, flaming, commenting.
Mackerras, Ruedi Aschwanden, Alan Cox, Manfred Spraul, Tim
Waugh, Pete Zaitcev, James Morris, Robert Love, Paul McKenney,
John Ashby for proofreading, correcting, flaming, commenting.
</para>
<para>
......@@ -1113,12 +1965,27 @@
<glossary id="glossary">
<title>Glossary</title>
<glossentry id="gloss-preemption">
<glossterm>preemption</glossterm>
<glossdef>
<para>
Prior to 2.5, or when <symbol>CONFIG_PREEMPT</symbol> is
unset, processes in user context inside the kernel would not
preempt each other (ie. you had that CPU until you have it up,
except for interrupts). With the addition of
<symbol>CONFIG_PREEMPT</symbol> in 2.5.4, this changed: when
in user context, higher priority tasks can "cut in": spinlocks
were changed to disable preemption, even on UP.
</para>
</glossdef>
</glossentry>
<glossentry id="gloss-bh">
<glossterm>bh</glossterm>
<glossdef>
<para>
Bottom Half: for historical reasons, functions with
`_bh' in them often now refer to any software interrupt, e.g.
'_bh' in them often now refer to any software interrupt, e.g.
<function>spin_lock_bh()</function> blocks any software interrupt
on the current CPU. Bottom halves are deprecated, and will
eventually be replaced by tasklets. Only one bottom half will be
......@@ -1132,8 +1999,7 @@
<glossdef>
<para>
Hardware interrupt request. <function>in_irq()</function> returns
<returnvalue>true</returnvalue> in a hardware interrupt handler (it
also returns true when interrupts are blocked).
<returnvalue>true</returnvalue> in a hardware interrupt handler.
</para>
</glossdef>
</glossentry>
......@@ -1144,8 +2010,7 @@
<para>
Not user context: processing a hardware irq or software irq.
Indicated by the <function>in_interrupt()</function> macro
returning <returnvalue>true</returnvalue> (although it also
returns true when interrupts or BHs are blocked).
returning <returnvalue>true</returnvalue>.
</para>
</glossdef>
</glossentry>
......@@ -1161,35 +2026,40 @@
</glossentry>
<glossentry id="gloss-softirq">
<glossterm>softirq</glossterm>
<glossterm>Software Interrupt / softirq</glossterm>
<glossdef>
<para>
Strictly speaking, one of up to 32 enumerated software
Software interrupt handler. <function>in_irq()</function> returns
<returnvalue>false</returnvalue>; <function>in_softirq()</function>
returns <returnvalue>true</returnvalue>. Tasklets and softirqs
both fall into the category of 'software interrupts'.
</para>
<para>
Strictly speaking a softirq is one of up to 32 enumerated software
interrupts which can run on multiple CPUs at once.
Sometimes used to refer to tasklets and bottom halves as
Sometimes used to refer to tasklets as
well (ie. all software interrupts).
</para>
</glossdef>
</glossentry>
<glossentry id="gloss-swinterrupt">
<glossterm>Software Interrupt / Software IRQ</glossterm>
<glossentry id="gloss-tasklet">
<glossterm>tasklet</glossterm>
<glossdef>
<para>
Software interrupt handler. <function>in_irq()</function> returns
<returnvalue>false</returnvalue>; <function>in_softirq()</function>
returns <returnvalue>true</returnvalue>. Tasklets, softirqs and
bottom halves all fall into the category of `software interrupts'.
A dynamically-registrable software interrupt,
which is guaranteed to only run on one CPU at a time.
</para>
</glossdef>
</glossentry>
<glossentry id="gloss-tasklet">
<glossterm>tasklet</glossterm>
<glossentry id="gloss-timers">
<glossterm>timer</glossterm>
<glossdef>
<para>
A dynamically-registrable software interrupt,
which is guaranteed to only run on one CPU at a time.
A dynamically-registrable software interrupt, which is run at
(or close to) a given time. When running, it is just like a
tasklet (in fact, they are called from the TIMER_SOFTIRQ).
</para>
</glossdef>
</glossentry>
......@@ -1207,10 +2077,11 @@
<glossterm>User Context</glossterm>
<glossdef>
<para>
The kernel executing on behalf of a particular
process or kernel thread (given by the <function>current()</function>
macro.) Not to be confused with userspace. Can be interrupted by
software or hardware interrupts.
The kernel executing on behalf of a particular process (ie. a
system call or trap) or kernel thread. You can tell which
process with the <symbol>current</symbol> macro.) Not to
be confused with userspace. Can be interrupted by software or
hardware interrupts.
</para>
</glossdef>
</glossentry>
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7