Re: possible bug in ll_rw_blk.c in 2.0.x and 2.1.x

mlord (mlord@pobox.com)
Mon, 24 Nov 1997 20:27:05 -0500

Kelly French wrote:
>
> Here's my 2nd try on this one. Instead of a linked list, it's now an
> ordered array. The search is still linear, but since the list is usually
> short, I didn't see the need for a binary search.
>
> To recap, this replaces the static 4-entry array for the block device
> stats in kstat with a dynamically allocated array so that we can hold more
> devices. This patch also removes the overlapping stats when you have both
> IDE and SCSI HD's.

How about something simpler, using fixed-size allocations per major,
without a built-in linear search prior to each and every disk access?

Something like below?

-ml

/**********************************************************/

++disk_stats[MAJOR(req->rq_dev)][unit].ds_disk_rio;

where "unit" is obtained from a switch() stmt similar to the existing
code.

At this time, I would suggest that a coded limit would be a simpler
alternative to dynamic expansion of the number of possible units per
major,
yielding code that might do something like this:

#define MAX_DISKSTATS_UNITS_PER_MAJOR 8 /* or 16, or
whatever.. */
struct kernel_stat_disk *disk_stats[MAX_BLKDEV];
...

static inline void disk_stat_acct(int cmd, unsigned long
nr_sectors,short disk_index)
{
struct kernel_stat_disk *stats;
stats = disk_stats[MAJOR(req->rq_dev)];

if (!stats) {
stats = disk_stats(MAJOR(req->rq_dev)] =
kmalloc(sizeof(struct kernel_stat_disk)
* MAX_DISKSTATS_UNITS_PER_MAJOR, GFP_KERNEL);
}
if (stats && disk_index < MAX_DISKSTATS_UNITS_PER_MAJOR) {
if (cmd == READ) {
++stats[disk_index].ds_disk_rio;
++stats[disk_index].ds_disk_rblks;
} else if (cmd == WRITE) {
++stats[disk_index].ds_disk_wio;
++stats[disk_index].ds_disk_wblks;
}
}
}

-- 
mlord@pobox.com
The Linux IDE guy