[PATCH 0/3 V2] time: use __builtin_constant_p() in msecs_to_jiffies

* [PATCH 0/3 V2] time: use __builtin_constant_p() in msecs_to_jiffies
@ 2015-04-12 12:13 Nicholas Mc Guire
  2015-04-12 12:13 ` [PATCH 1/3 V2] time: move timeconst.h into include/generated Nicholas Mc Guire
                   ` (2 more replies)
  0 siblings, 3 replies; 10+ messages in thread
From: Nicholas Mc Guire @ 2015-04-12 12:13 UTC (permalink / raw)
  To: Michal Marek
  Cc: Masahiro Yamada, Sam Ravnborg, Thomas Gleixner, H. Peter Alvin,
	Joe Perches, John Stultz, Andrew Hunter, Paul Turner,
	Aaron Sierra, Brian Norris, linux-kernel, Nicholas Mc Guire

In the overall kernel source there currently are
  2544 msecs_to_jiffies
  126  usecs_to_jiffies
and a few places that are using  var * HZ / 1000  constructs
which are not always safe (no check of corner cases) and should
be switched to msecs_to_jiffies (roughly 25 left).
Allowing gcc to fold constants for these calls that in most
cases are passing in constants (roughly 95%) has some potential
to improve performance (and should save a few bytes).

size impact is marginal and for Powerpc it is
actually a slight increase.

As the changes to the top level Kbuild will impact every architecture
this is probably not enough - but I think suitable for a first review

Once this is clean a patch for usecs_to_jiffies will be provided as well

The patch set:
 0001  moves timeconst.h from kernel/time/ to include/generated/ and makes
       it available early enough so that the build can use the constants
       for msecs_to_jiffies
 0002  rename msecs_to_jiffies() to __msecs_to_jiffies()
       move the common HZ range specific #ifdef'ed code into a inline helper
       and #ifdef the variant to use. This allows to re-use the helpers in
       both the const and non-const variant of msecs_to_jiffies()
       modified msecs_to_jiffies() to checks for constant via 
       call to __builtin_constant_p() and call __msecs_to_jiffies() if it 
       can't determine that the argument is constant.
 0003  documentation update and reformatting to kernel-doc format  
       for msecs_to_jiffies() and __msecs_to_jiffies() - for the helpers
       its left as comments.

Verification:

kernel configs tested are defconfigs with e.g.
make x86_64_defconfig + CONFIG_TEST_LKM=m, GCONFIG_HZ_300=y. CONFIG_HZ=300

check kernel/softirq.c
#define MAX_SOFTIRQ_TIME  msecs_to_jiffies(2)
used in __do_softirq() is folded into a single addq    $1, %rax  
conversely kernel/sched/core.c:sched_rr_handler() is not constant and
is a full call    __msecs_to_jiffies

added the test-case proposed by Joe Perches <joe@perches.com>
into lib/test_module.c:test_module_init()
<snip>
        unsigned int m;                                                         

        for (m = 10; m < 200; m += 10)                                          
                pr_info("msecs_to_jiffies(%u) is %lu\n",                        
                        m, msecs_to_jiffies(m));                                

        pr_info("msecs_to_jiffies(%u) is %lu\n",                                
                10, msecs_to_jiffies(10));                                      
        pr_info("msecs_to_jiffies(%u) is %lu\n",                                
                100, msecs_to_jiffies(100));                                    
        pr_info("msecs_to_jiffies(%u) is %lu\n",                                
                1000, msecs_to_jiffies(1000));                                  
<snip>

without the patch applied:
test_module_init:
	pushq	%rbp	#
	movq	%rsp, %rbp	#,
	pushq	%rbx	#
	movl	$10, %ebx	#, m
	pushq	%rcx	#
.L2:
	movl	%ebx, %edi	# m,
	call	msecs_to_jiffies	#
	movl	%ebx, %esi	# m,
	movq	%rax, %rdx	# D.14503,
	movq	$.LC0, %rdi	#,
	xorl	%eax, %eax	#
	addl	$10, %ebx	#, m
	call	printk	#
	cmpl	$200, %ebx	#, m
	jne	.L2	#,
	movl	$10, %edi	#,   <--- msecs_to_jiffies(10) 
	call	msecs_to_jiffies	#  <--- runtime conversion
	movl	$10, %esi	#,
	movq	%rax, %rdx	# D.14504,
	movq	$.LC0, %rdi	#,
	xorl	%eax, %eax	#
	call	printk	#
	movl	$100, %edi	#,
	call	msecs_to_jiffies	#
	movl	$100, %esi	#,
	movq	%rax, %rdx	# D.14505,
	movq	$.LC0, %rdi	#,
	xorl	%eax, %eax	#
	call	printk	#
	movl	$1000, %edi	#,
	call	msecs_to_jiffies	#
	movl	$1000, %esi	#,
	movq	%rax, %rdx	# D.14506,
	movq	$.LC0, %rdi	#,
	xorl	%eax, %eax	#
	call	printk	#
	movq	$.LC1, %rdi	#,
	xorl	%eax, %eax	#
	call	printk	#
	popq	%rdx	#
	popq	%rbx	#
	xorl	%eax, %eax	#
	popq	%rbp	#
	ret

with the patch applied:
test_module_init:
	pushq	%rbp	#
	movq	%rsp, %rbp	#,
	pushq	%rbx	#
	movl	$10, %ebx	#, m
	pushq	%rcx	#
.L2:
	movl	%ebx, %edi	# m,
	call	__msecs_to_jiffies	#
	movl	%ebx, %esi	# m,
	movq	%rax, %rdx	# D.14545,
	movq	$.LC0, %rdi	#,
	xorl	%eax, %eax	#
	addl	$10, %ebx	#, m
	call	printk	#
	cmpl	$200, %ebx	#, m
	jne	.L2	#,
	movl	$3, %edx	#,   <--- msecs_to_jiffies(10) == 3 jiffies
	movl	$10, %esi	#,   <--- const 10 passed to printk
	movq	$.LC0, %rdi	#,
	xorl	%eax, %eax	#
	call	printk	#
	movl	$30, %edx	#,
	movl	$100, %esi	#,
	movq	$.LC0, %rdi	#,
	xorl	%eax, %eax	#
	call	printk	#
	movl	$300, %edx	#,
	movl	$1000, %esi	#,
	movq	$.LC0, %rdi	#,
	xorl	%eax, %eax	#
	call	printk	#
	movq	$.LC1, %rdi	#,
	xorl	%eax, %eax	#
	call	printk	#
	popq	%rdx	#
	popq	%rbx	#
	xorl	%eax, %eax	#
	popq	%rbp	#
	ret

^ permalink raw reply	[flat|nested] 10+ messages in thread