Nasty IDE crasher in 2.6.9rc1
You never never issue unknown commands to drives. Thats how Mandrake destroyed CD-ROM drives. I knew this was in -mm and supposed to be getting sorted I was somewhat horrified to find it in 2.6.9rc1.
This patch crashes two of my CF cards (one so badly you have to reformat it to get it back) and anything attached to an IT8212 controller. The correct fix is to do what the standard actually says and always check for cache flush. Contrary to the comment in the patch drives do report this correctly its just that some of them nop unknown commands.
Please fix this patch segment for rc2, its not just wrong, its dangerous.
Another problem with barrier is that it can take several minutes worst case for the command to complete on a large modern drive (timings c/o friendly ide drive engineer). That causes two problems I've pointed out to Jens that we need to fix before barriers are IMHO production grade
1. Anything based on fairness and latency is screwed. Throughput
apparently is up so it makes sense for some users, and probably for others we should write cache off as Jens suggested.
2. The timeouts on the command issue appear to be too small, and
we will time out and reset the drive in loaded situations.
Thankfully next generation ATA has both cache bypass writes and tagging.
drivers/ide/ide-disk.c +21 -0
@@ -1543,6 +1608,27 @@ drive->wcache =1 ;write_cache (drive,1 ); + +/* +* decide if we can sanely support flushes and barriers on +* this drive. unfortunately not all drives advertise FLUSH_CACHE +* support even if they support it. So assume FLUSH_CACHE is there +* always. LBA48 drives are newer, so expect it to flag support +* properly. We can safely support FLUSH_CACHE on lba48, if capacity +* doesn't exceed lba28 +*/ + barrier =1 ; +if (drive->addressing ==1 ) { +if (capacity > (1ULL <<28 ) && !ide_id_has_flush_cache_ext (id)) + barrier =0 ; + } + +printk ("%s: cache flushes %ssupported\n" , + drive->name, barrier ?"" :"not " ); +if (barrier) { +blk_queue_ordered (drive->queue,1 ); +blk_queue_issue_flush_fn (drive->queue, idedisk_issue_flush); + } }static void ide_cacheflush_p (ide_drive_t *drive)
(suse.dk is not related to suse.de and it helpfully eats all messages sent to unknown users. not so great :(
Quoted 11 lines · Alan Cox
You never never issue unknown commands to drives. Thats how Mandrake destroyed CD-ROM drives. I knew this was in -mm and supposed to be getting sorted I was somewhat horrified to find it in 2.6.9rc1.
This patch crashes two of my CF cards (one so badly you have to reformat it to get it back) and anything attached to an IT8212 controller. The correct fix is to do what the standard actually says and always check for cache flush. Contrary to the comment in the patch drives do report this correctly its just that some of them nop unknown commands.
Please fix this patch segment for rc2, its not just wrong, its dangerous.
Ugh, that's bad. I agree with the change, thanks. Linus passed it on.
Quoted 4 lines
Another problem with barrier is that it can take several minutes worst case for the command to complete on a large modern drive (timings c/o friendly ide drive engineer). That causes two problems I've pointed out to Jens that we need to fix before barriers are IMHO production grade
Can you pass me his results?
Quoted 3 lines
1. Anything based on fairness and latency is screwed. Throughput
apparently is up so it makes sense for some users, and probably for others we should write cache off as Jens suggested.
Yes, it's a tradeoff. The user can decide himself what is most important. It all depends on the work load, of course.
Quoted 2 lines
2. The timeouts on the command issue appear to be too small, and
we will time out and reset the drive in loaded situations.
You don't seem to address that in your patch?
Quoted 1 line
Thankfully next generation ATA has both cache bypass writes and tagging.
But the tagging still isn't useful for this. Have they added WIN_WRITE_DMA_EXT_QUEUED_FUA?
Signature · 1 line
Jens Axboe
Quoted 2 lines · Jens Axboe
(suse.dk is not related to suse.de and it helpfully eats all messages sent to unknown users. not so great :(
Ah sorry.
Quoted 6 lines
Quoted 4 lines
Another problem with barrier is that it can take several minutes worst case for the command to complete on a large modern drive (timings c/o friendly ide drive engineer). That causes two problems I've pointed out to Jens that we need to fix before barriers are IMHO production grade
Can you pass me his results?
I can ask. Its NDA data (not Maxtor). Or Eric might have public info ? The later mail I reported my tests trying to make it as slow as possible and I couldn't get worse than 7 seconds for the command.
Quoted 4 lines
Quoted 2 lines
2. The timeouts on the command issue appear to be too small, and
we will time out and reset the drive in loaded situations.You don't seem to address that in your patch?
I'm not sure what the right answer is.
Quoted 11 lines · Alan Cox
Quoted 2 lines · Jens Axboe
(suse.dk is not related to suse.de and it helpfully eats all messages sent to unknown users. not so great :(
Ah sorry.
Quoted 5 lines
Quoted 4 lines
Another problem with barrier is that it can take several minutes worst case for the command to complete on a large modern drive (timings c/o friendly ide drive engineer). That causes two problems I've pointed out to Jens that we need to fix before barriers are IMHO production grade
Quoted 5 lines
Quoted 1 line
Can you pass me his results?
I can ask. Its NDA data (not Maxtor). Or Eric might have public info ? The later mail I reported my tests trying to make it as slow as possible and I couldn't get worse than 7 seconds for the command.
IIRC, 7 seconds is the magic number that Microsoft uses for when a command times out in the kernel... That might make the results a little suspicious :)
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
2. The timeouts on the command issue appear to be too small, and
we will time out and reset the drive in loaded situations.You don't seem to address that in your patch?
I'm not sure what the right answer is.
I guess as a first measure just increasing the timeout two-fold will cover most of the problem.
Signature · 1 line
Jens Axboe