Andrew Morton a écrit :
> On Wed, 19 Dec 2007 01:14:33 +0100
> Eric Dumazet <dada1@cosmosbay.com> wrote:
> 
>> groups_sort() can be quite long if user loads a large gid table.
>>
>> This is because GROUP_AT(group_info, some_integer) uses an integer divide.
>> So having to do XXX thousand divides during one syscall can lead to very high 
>> latencies. (NGROUPS_MAX=65536)
>>
>> In the past (25 Mar 2006), an analog problem was found in groups_search()
>> (commit d74beb9f33a5f16d2965f11b275e401f225c949d ) and at that time I changed 
>> some variables to unsigned int.
>>
>> I believe that a more generic fix is to make sure NGROUPS_PER_BLOCK is unsigned.
>>
> 
>> diff --git a/include/linux/sched.h b/include/linux/sched.h
>> index ac3d496..725a491 100644
>> --- a/include/linux/sched.h
>> +++ b/include/linux/sched.h
>> @@ -775,7 +775,7 @@ static inline int above_background_load(void)
>>  
>>  struct io_context;			/* See blkdev.h */
>>  #define NGROUPS_SMALL		32
>> -#define NGROUPS_PER_BLOCK	((int)(PAGE_SIZE / sizeof(gid_t)))
> 
> That was a dopey thing to do.  Both PAGE_SIZE and sizeof() are unsigned and
> we went and cast it to a signed thing, even though it is a nonsense to ever
> consider a negative value of this.

Maybe the author intention was to cast it to an "int", instead of "long", I 
dont know...

> 
>> +#define NGROUPS_PER_BLOCK	((unsigned int)(PAGE_SIZE / sizeof(gid_t)))
>>  struct group_info {
>>  	int ngroups;
>>  	atomic_t usage;
>> diff --git a/kernel/sys.c b/kernel/sys.c
>> index d1fe71e..091e58f 100644
>> --- a/kernel/sys.c
>> +++ b/kernel/sys.c
>> @@ -1148,7 +1148,7 @@ static int groups_to_user(gid_t __user *grouplist,
>>  	int count = group_info->ngroups;
>>  
>>  	for (i = 0; i < group_info->nblocks; i++) {
>> -		int cp_count = min(NGROUPS_PER_BLOCK, count);
>> +		int cp_count = min_t(int, NGROUPS_PER_BLOCK, count);
>>  		int off = i * NGROUPS_PER_BLOCK;
>>  		int len = cp_count * sizeof(*grouplist);
>>  
>> @@ -1168,7 +1168,7 @@ static int groups_from_user(struct group_info *group_info,
>>  	int count = group_info->ngroups;
>>  
>>  	for (i = 0; i < group_info->nblocks; i++) {
>> -		int cp_count = min(NGROUPS_PER_BLOCK, count);
>> +		int cp_count = min_t(int, NGROUPS_PER_BLOCK, count);
>>  		int off = i * NGROUPS_PER_BLOCK;
>>  		int len = cp_count * sizeof(*grouplist);
>>  
> 
> We now have a jumble of signed and unsigned, stuck together with min_t
> sticky tape.
> 
> Please, take a closer look in there and see if there's anything which
> *needs* to be signed: is there anything which can actually, sensibly have a
> negative quantity?  I bet there isn't, and I bet the code would be cleaner,
> faster and more obviously correct if it was converted to unsigned throughout.
> 
> I blame C.  Negative quantities are rare, and C's default of treating
> scalars as signed was a mistake.  Oh well.
> 
> 

Yes I agree that 'unsigned' should be the default...

Well, I was trying to keep patch small, but as you have some time, I will make 
it bigger :)

[PATCH] kernel/sys.c : Get rid of expensive divides in groups_sort()

groups_sort() can be quite long if user loads a large gid table.

This is because GROUP_AT(group_info, some_integer) uses an integer divide.
So having to do XXX thousand divides during one syscall can lead to very high 
latencies. (NGROUPS_MAX=65536)

In the past (25 Mar 2006), an analog problem was found in groups_search()
(commit d74beb9f33a5f16d2965f11b275e401f225c949d ) and at that time I changed 
some variables to unsigned int.

I believe that a more generic fix is to make sure NGROUPS_PER_BLOCK is unsigned.

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>