kern/115572: ata disk bug

[Søren Schmidt]

Poul-Henning Kamp wrote:
> In message <200708182120.l7ILKBvF046099 at freefall.freebsd.org>, Graham writes:
>
>   
>> 2. attempt (say)....
>> rabbit# dd if=/dev/zero of=/dev/ad4s1 oseek=2097151 count=1 bs=64k
>> and the result is....
>> dd: /dev/ad4s1: Input/output error
>> 1+0 records in
>> 0+0 records out
>> 0 bytes transferred in 0.000325 secs (0 bytes/sec)
>>
>> (If dd is performed on the raw drive, /dev/ad4 then block boundary is
>> always a power of 2, and blocksize a smaller power of 2.  That's always
>> ok.  But we can't assume we use drives that way.)
>>
>> So a transfer which starts in the 28-bit zone, but extends over into
>> the 48-bit region, fails.  Such transfers happen in the superblock of
>> certain size drives, and that plays havoc.  The sector mapping of gbde
>> can do this, but soft-update gets screwed by this happening.  It's not
>> actually to do with the crypto as I first suspected.
>>     
>
> This is a problem in the ata disk driver.
>   
Yeah, the crossover point from using 28bit to 48bit addressing is 
flawed, below patch should fix that:

--- ata-all.c   23 Feb 2007 16:25:08 -0000      1.279
+++ ata-all.c   19 Aug 2007 09:25:58 -0000
@@ -738,7 +738,7 @@
 
     atadev->flags &= ~ATA_D_48BIT_ACTIVE;
 
-    if ((request->u.ata.lba >= ATA_MAX_28BIT_LBA ||
+    if (((request->u.ata.lba + request->u.ata.count) >= 
ATA_MAX_28BIT_LBA ||
         request->u.ata.count > 256) &&
        atadev->param.support.command2 & ATA_SUPPORT_ADDRESS48) {
 


-Søren
>
>
>