Re: [RFT] major libata update

[Jeff Garzik <jeff@garzik.org>]

Tejun Heo wrote:
> Jeff Garzik wrote:
>> Tejun Heo wrote:
>>> Jeff Garzik wrote:
>>>> Jeff Garzik wrote:
>>>>> Tejun Heo wrote:
>>>>>> Hmmm.. The drive is issuing SDB FIS which completes already 
>>>>>> completed tags.  This could be dangerous.  Depending on timing, it 
>>>>>> might end up finishing a command which occupied the slot which 
>>>>>> hasn't been processed yet.  If a drive does this, NCQ shouldn't be 
>>>>>> enabled for it.  Can you post full boot dmesg?
>>>>>
>>>>> I'm not sure the data supports that conclusion?  PORT_IRQ_SDB_FIS 
>>>>> is quite normal and expected during NCQ operation, if that 
>>>>> interrupt is enabled.  Just normal SDB:Entry and SDB:SetIntr states.
>>>>
>>>> Strike that last part:  PORT_IRQ_SDB_FIS will appear, as with other 
>>>> status bits, even if the enable bit is not set.
>>>>
>>>> So, you'll see that whenever you get an SDB FIS during normal 
>>>> operation.
>>>
>>> The problem is with the second dword.  Here are some of spurious SDB 
>>> FISes Ric's AHCI was receiving.
>>>
>>> 004040a1:10000000
>>> 004040a1:00000020
>>> 004040a1:00000080
>>>
>>> If the second dword were all zero, it's simply SDB FIS turning on IRQ 
>>> (bit 14 of the first dword) and there's nothing to worry about. 
>>> However, all those spurious SDBs have one bit set in the second dword 
>>> - meaning the SDB completes the corresponding tag, but the tag isn't 
>>> active when those SDBs are received.
>>>
>>> This is okay as long as the controller thinks the tags are unoccupied 
>>> when those SDBs are received, but it's not something which can be 
>>> guaranteed.  NCQ command synchronization depends on devices not 
>>> completing the same commands more than once.
>>>
>>> The duplicate completions might be okay if the drive guarantees it 
>>> doesn't send it if it loses to command issuance.  e.g.
>>>
>>> 1. drive sends completion for tag x
>>> 2. drive shortly schedules another completion for tag x (spurious)
>>> 3. ahci/driver complete tag x
>>> 4. ahci/driver issues tag x
>>> 5. drive receives command for tag x before sending the spurious 
>>> completion and determines not to send the spurious completion. (not 
>>> very likely)
>>>
>>> If above is true, the drive might be okay, but nobody can guarantee 
>>> how  various controllers react.  It depends on how controllers manage 
>>> SActive (when to turn bits on).  At any rate, it's dangerous IMHO.
>>
>> If the silicon is screwing up SActive bits, then we have bigger 
>> problems than spurious interrupts.
>>
>> So, the typical policy of Internet servers applies here:  "be liberal 
>> in what you accept."  For smart controllers like AHCI, we will simply 
>> set the desired IRQ mask, then happily receive and ack events anytime 
>> the controller decides to raise them.  If the controller decides to 
>> send us a no-op, don't worry about it.  This is particularly true when 
>> we turn on Command Coalescing, where we'll have a run of work 
>> initiated [sometimes] by an internal timer, rather than an actual FIS 
>> reception.
> 
> I wish I could explain it better.  This is a clear protocol violation 
> from the drive.  Depending on specific implementation of the drive and 
> the controller, it can result in completion of command which is not 
> processed yet (data corruption!).

I quite understand the implications.  My argument comes from a different 
angle:  I don't feel we should be adding tons of code that essentially 
validates the silicon.  There are plenty of chances for the hardware to 
fuck up in a way that corrupts data, and is also difficult to detect. 
Pre-production BIOS have even done silly things like turn off data 
verification (checksum) by default.  Talk about subtle corruption...

So I feel the best path is to use the hardware programming sequences 
described in the spec, because that's what the chip designers and Q/A 
engineers validate with (read: the Windows driver).

Once we have deployed drivers with the standard programming sequences, 
_then_ we can consider looking into proper spurious interrupt 
accounting.  The current AHCI interrupt accounting stuff is not nearly 
as accurate as it should be, which implies that the code simply should 
not exist at the present time.

	Jeff