Home Home > GIT Browse > stable-xen
summaryrefslogtreecommitdiff
blob: 70374907c020fdf8a33ac66706aa43a1d14d79e4 (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
Lemma 1:
	If ps_tq is scheduled, ps_tq_active is 1.  ps_tq_int() can be called
	only when ps_tq_active is 1.
Proof:	All assignments to ps_tq_active and all scheduling of ps_tq happen
	under ps_spinlock.  There are three places where that can happen:
	one in ps_set_intr() (A) and two in ps_tq_int() (B and C).
	Consider the sequnce of these events.  A can not be preceded by
	anything except B, since it is under if (!ps_tq_active) under
	ps_spinlock.  C is always preceded by B, since we can't reach it
	other than through B and we don't drop ps_spinlock between them.
	IOW, the sequence is A?(BA|BC|B)*.  OTOH, number of B can not exceed
	the sum of numbers of A and C, since each call of ps_tq_int() is
	the result of ps_tq execution.  Therefore, the sequence starts with
	A and each B is preceded by either A or C.  Moments when we enter
	ps_tq_int() are sandwiched between {A,C} and B in that sequence,
	since at any time number of B can not exceed the number of these
	moments which, in turn, can not exceed the number of A and C.
	In other words, the sequence of events is (A or C set ps_tq_active to
	1 and schedule ps_tq, ps_tq is executed, ps_tq_int() is entered,
	B resets ps_tq_active)*.


consider the following area:
	* in do_pd_request1(): to calls of pi_do_claimed() and return in
	  case when pd_req is NULL.
	* in next_request(): to call of do_pd_request1()
	* in do_pd_read(): to call of ps_set_intr()
	* in do_pd_read_start(): to calls of pi_do_claimed(), next_request()
and ps_set_intr()
	* in do_pd_read_drq(): to calls of pi_do_claimed() and next_request()
	* in do_pd_write(): to call of ps_set_intr()
	* in do_pd_write_start(): to calls of pi_do_claimed(), next_request()
and ps_set_intr()
	* in do_pd_write_done(): to calls of pi_do_claimed() and next_request()
	* in ps_set_intr(): to check for ps_tq_active and to scheduling
	  ps_tq if ps_tq_active was 0.
	* in ps_tq_int(): from the moment when we get ps_spinlock() to the
	  return, call of con() or scheduling ps_tq.
	* in pi_schedule_claimed() when called from pi_do_claimed() called from
	  pd.c, everything until returning 1 or setting or setting ->claim_cont
	  on the path that returns 0
	* in pi_do_claimed() when called from pd.c, everything until the call
	  of pi_do_claimed() plus the everything until the call of cont() if
	  pi_do_claimed() has returned 1.
	* in pi_wake_up() called for PIA that belongs to pd.c, everything from
	  the moment when pi_spinlock has been acquired.

Lemma 2:
	1) at any time at most one thread of execution can be in that area or
	be preempted there.
	2) When there is such a thread, pd_busy is set or pd_lock is held by
	that thread.
	3) When there is such a thread, ps_tq_active is 0 or ps_spinlock is
	held by that thread.
	4) When there is such a thread, all PIA belonging to pd.c have NULL
	->claim_cont or pi_spinlock is held by thread in question.

Proof:	consider the first moment when the above is not true.

(1) can become not true if some thread enters that area while another is there.
	a) do_pd_request1() can be called from next_request() or do_pd_request()
	   In the first case the thread was already in the area.  In the second,
	   the thread was holding pd_lock and found pd_busy not set, which would
	   mean that (2) was already not true.
	b) ps_set_intr() and pi_schedule_claimed() can be called only from the
	   area.
	c) pi_do_claimed() is called by pd.c only from the area.
	d) ps_tq_int() can enter the area only when the thread is holding
	   ps_spinlock and ps_tq_active is 1 (due to Lemma 1).  It means that
	   (3) was already not true.
	e) do_pd_{read,write}* could be called only from the area.  The only
	   case that needs consideration is call from pi_wake_up() and there
	   we would have to be called for the PIA that got ->claimed_cont
	   from pd.c.  That could happen only if pi_do_claimed() had been
	   called from pd.c for that PIA, which happens only for PIA belonging
	   to pd.c.
	f) pi_wake_up() can enter the area only when the thread is holding
	   pi_spinlock and ->claimed_cont is non-NULL for PIA belonging to
	   pd.c.  It means that (4) was already not true.

(2) can become not true only when pd_lock is released by the thread in question.
	Indeed, pd_busy is reset only in the area and thread that resets
	it is holding pd_lock.	The only place within the area where we
	release pd_lock is in pd_next_buf() (called from within the area).
	But that code does not reset pd_busy, so pd_busy would have to be
	0 when pd_next_buf() had acquired pd_lock.  If it become 0 while
	we were acquiring the lock, (1) would be already false, since
	the thread that had reset it would be in the area simulateously.
	If it was 0 before we tried to acquire pd_lock, (2) would be
	already false.

For similar reasons, (3) can become not true only when ps_spinlock is released
by the thread in question.  However, all such places within the area are right
after resetting ps_tq_active to 0.

(4) is done the same way - all places where we release pi_spinlock within
the area are either after resetting ->claimed_cont to NULL while holding
pi_spinlock, or after not tocuhing ->claimed_cont since acquiring pi_spinlock
also in the area.  The only place where ->claimed_cont is made non-NULL is
in the area, under pi_spinlock and we do not release it until after leaving
the area.

QED.


Corollary 1: ps_tq_active can be killed.  Indeed, the only place where we
check its value is in ps_set_intr() and if it had been non-zero at that
point, we would have violated either (2.1) (if it was set while ps_set_intr()
was acquiring ps_spinlock) or (2.3) (if it was set when we started to
acquire ps_spinlock).

Corollary 2: ps_spinlock can be killed.  Indeed, Lemma 1 and Lemma 2 show
that the only possible contention is between scheduling ps_tq followed by
immediate release of spinlock and beginning of execution of ps_tq on
another CPU.

Corollary 3: assignment to pd_busy in do_pd_read_start() and do_pd_write_start()
can be killed.  Indeed, we are not holding pd_lock and thus pd_busy is already
1 here.

Corollary 4: in ps_tq_int() uses of con can be replaced with uses of
ps_continuation, since the latter is changed only from the area.
We don't need to reset it to NULL, since we are guaranteed that there
will be a call of ps_set_intr() before we look at ps_continuation again.
We can remove the check for ps_continuation being NULL for the same
reason - the value is guaranteed to be set by the last ps_set_intr() and
we never pass it NULL.  Assignements in the beginning of ps_set_intr()
can be taken to callers as long as they remain within the area.