Comment 476 for bug 59695

Alexey Borzenkov (snaury) wrote :

Some more info on the issue. I recently bought a new Samsung HM160HI and started having the issue with Load_Cycle_Count. I tried the solution with setting hdparm -B 255, but what I found is that my drive is already in that mode after reboot. So after hdparm -B 254 Load_Cycle_Count stopped increasing. But what I've found is that after this Load_Retry_Count starts increasing, with an extremely loud click once per several minues, which I found much worse than increasing Load_Cycle_Count.

The real issue, it seems, is that something is constantly waking the disk up, even when disk shouldn't be touched. I've been playing with laptop-mode and a modified (which shows actual bytes for each process in batch mode, skipping the ones that don't do any actual io), and found the following. When I set hdparm -B 255, and hdparm -S 4, the disk spins down after 20 seconds, however after a very short time (less than a minute) firefox writes something to disk and pdflush kicks in dropping in, even though /proc/meminfo shows only several kilobytes under "Dirty", while my dirty ratio is 60 percent. Very strange.

Now I suspected that there might be something going on with pagefile, so I swapoff my swap partition and try all this again. Now pdflush no longer kicks in immediately after firefox or other programs. But things get even more fishy. Even without any I/O activity (both my modified and gkrellm show absolutely nothing) disk spins down, but shortly after that spins up again.

Why would my disk spin up again without any I/O?

Judging from increasing Load_Retry_Count, it seems to me that my drive parks heads with a purpose (and when you don't let it, something bad happens, i.e. that horrible click as if heads drop from somewhere). But what's wrong here is that something pulls them back. Something is constantly touching the drive when it shouldn't.

p.s. I don't have trackerd installed, and during my tests the only processes that showed up were either none, or firefox and [pdflush].