aboutsummaryrefslogtreecommitdiff
path: root/Documentation/RCU/UP.txt
blob: 90ec5341ee981a0a397710d2ffd1261eb2557896 (plain)
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
134
135
RCU on Uniprocessor Systems


A common misconception is that, on UP systems, the call_rcu() primitive
may immediately invoke its function.  The basis of this misconception
is that since there is only one CPU, it should not be necessary to
wait for anything else to get done, since there are no other CPUs for
anything else to be happening on.  Although this approach will -sort- -of-
work a surprising amount of the time, it is a very bad idea in general.
This document presents three examples that demonstrate exactly how bad
an idea this is.


Example 1: softirq Suicide

Suppose that an RCU-based algorithm scans a linked list containing
elements A, B, and C in process context, and can delete elements from
this same list in softirq context.  Suppose that the process-context scan
is referencing element B when it is interrupted by softirq processing,
which deletes element B, and then invokes call_rcu() to free element B
after a grace period.

Now, if call_rcu() were to directly invoke its arguments, then upon return
from softirq, the list scan would find itself referencing a newly freed
element B.  This situation can greatly decrease the life expectancy of
your kernel.

This same problem can occur if call_rcu() is invoked from a hardware
interrupt handler.


Example 2: Function-Call Fatality

Of course, one could avert the suicide described in the preceding example
by having call_rcu() directly invoke its arguments only if it was called
from process context.  However, this can fail in a similar manner.

Suppose that an RCU-based algorithm again scans a linked list containing
elements A, B, and C in process contexts, but that it invokes a function
on each element as it is scanned.  Suppose further that this function
deletes element B from the list, then passes it to call_rcu() for deferred
freeing.  This may be a bit unconventional, but it is perfectly legal
RCU usage, since call_rcu() must wait for a grace period to elapse.
Therefore, in this case, allowing call_rcu() to immediately invoke
its arguments would cause it to fail to make the fundamental guarantee
underlying RCU, namely that call_rcu() defers invoking its arguments until
all RCU read-side critical sections currently executing have completed.

Quick Quiz #1: why is it -not- legal to invoke synchronize_rcu() in
	this case?


Example 3: Death by Deadlock

Suppose that call_rcu() is invoked while holding a lock, and that the
callback function must acquire this same lock.  In this case, if
call_rcu() were to directly invoke the callback, the result would
be self-deadlock.

In some cases, it would possible to restructure to code so that
the call_rcu() is delayed until after the lock is released.  However,
there are cases where this can be quite ugly:

1.	If a number of items need to be passed to call_rcu() within
	the same critical section, then the code would need to create
	a list of them, then traverse the list once the lock was
	released.

2.	In some cases, the lock will be held across some kernel API,
	so that delaying the call_rcu() until the lock is released
	requires that the data item be passed up via a common API.
	It is far better to guarantee that callbacks are invoked
	with no locks held than to have to modify such APIs to allow
	arbitrary data items to be passed back up through them.

If call_rcu() directly invokes the callback, painful locking restrictions
or API changes would be required.

Quick Quiz #2: What locking restriction must RCU callbacks respect?


Summary

Permitting call_rcu() to immediately invoke its arguments breaks RCU,
even on a UP system.  So do not do it!  Even on a UP system, the RCU
infrastructure -must- respect grace periods, and -must- invoke callbacks
from a known environment in which no locks are held.

It -is- safe for synchronize_sched() and synchronize_rcu_bh() to return
immediately on an UP system.  It is also safe for synchronize_rcu()
to return immediately on UP systems, except when running preemptable
RCU.

Quick Quiz #3: Why can't synchronize_rcu() return immediately on
	UP systems running preemptable RCU?


Answer to Quick Quiz #1:
	Why is it -not- legal to invoke synchronize_rcu() in this case?

	Because the calling function is scanning an RCU-protected linked
	list, and is therefore within an RCU read-side critical section.
	Therefore, the called function has been invoked within an RCU
	read-side critical section, and is not permitted to block.

Answer to Quick Quiz #2:
	What locking restriction must RCU callbacks respect?

	Any lock that is acquired within an RCU callback must be
	acquired elsewhere using an _irq variant of the spinlock
	primitive.  For example, if "mylock" is acquired by an
	RCU callback, then a process-context acquisition of this
	lock must use something like spin_lock_irqsave() to
	acquire the lock.

	If the process-context code were to simply use spin_lock(),
	then, since RCU callbacks can be invoked from softirq context,
	the callback might be called from a softirq that interrupted
	the process-context critical section.  This would result in
	self-deadlock.

	This restriction might seem gratuitous, since very few RCU
	callbacks acquire locks directly.  However, a great many RCU
	callbacks do acquire locks -indirectly-, for example, via
	the kfree() primitive.

Answer to Quick Quiz #3:
	Why can't synchronize_rcu() return immediately on UP systems
	running preemptable RCU?

	Because some other task might have been preempted in the middle
	of an RCU read-side critical section.  If synchronize_rcu()
	simply immediately returned, it would prematurely signal the
	end of the grace period, which would come as a nasty shock to
	that other thread when it started running again.