Re: [EXT] Re: [PATCH v1 3/4] serial: 8250_pci: Always try MSI/MSI-X

[Ralf Ramsauer]

On 14/07/2021 15:35, Andy Shevchenko wrote:
> On Wed, Jul 14, 2021 at 3:56 PM Ralf Ramsauer
> <ralf.ramsauer@xxxxxxxxxxxxxxxxx> wrote:
>> On 14/07/2021 08:54, Jiri Slaby wrote:
>>> On 13. 07. 21, 12:40, Andy Shevchenko wrote:
> 
>>> Hmm, have you checked the commit which introduced the whitelist?
>>>
>>>     Nevertheless, this needs to handled with care: while many 8250 devices
>>>     actually claim to support MSI(-X) interrupts it should not be
>>> enabled be
>>>     default. I had at least one device in my hands with broken MSI
>>>     implementation.
>>>
>>>     So better introduce a whitelist with devices that are known to support
>>>     MSI(-X) interrupts. I tested all devices mentioned in the patch.
>>>
>>>
>>> You should have at least CCed the author for an input.
>>
>> Yep, back then I was testing three different 8250 pci cards. All of them
>> claimed to support MSI, while one really worked with MSI, the one that I
>> whitelisted. So I thought it would be better to use legacy IRQs as long
>> as no one tested a specific card to work with MSI.
> 
> Can you shed a light eventually what those cards are?

So I found a no-name el-cheapo card that has some issues with MSI:

18:00.0 Serial controller: Device 1c00:3253 (rev 10) (prog-if 05 [16850])

The card comes with two serial lines. It comes perfectly up, if I enable
it to use MSI in the whitelist:

serial 0000:18:00.0: Using MSI(-X) interrupts
serial 0000:18:00.0: Setup PCI port: port 40c0, irq 104, type 0
0000:18:00.0: ttyS6 at I/O 0x40c0 (irq = 104, base_baud = 115200) is a
XR16850
serial 0000:18:00.0: Setup PCI port: port 40c8, irq 104, type 0
0000:18:00.0: ttyS7 at I/O 0x40c8 (irq = 104, base_baud = 115200) is a
XR16850

After loading 8250_pci, lspci -vvs 18:0.0 tells:

	Capabilities: [68] MSI: Enable+ Count=1/32 Maskable+ 64bit+
		Address: 00000000fee000b8  Data: 0000
		Masking: ffffffff  Pending: 00000000

Looks good so far. Now let's echo to the device.

$ echo asdf > /dev/ttyS6

-- stuck. The echoing process stucks at close():

write(1, "asdf\n", 5)                   = 5
close(1

Stuck in the sense of: the echo is still killable, no crashes. The same
happens if I try to access the device with stty. So something is odd
here. However, the Netmos cards that I whitelisted do a great job.

So I can't tell if I was just unlucky to grab a card that has issues
with MSI, and this is an exception rather than the rule…

HTH,
  Ralf


> 
>> Don't do that… And don't convert it to a blacklist. A blacklist will
>> break users until they report that something doesn't work.
> 
> White list is not okay either. MSI in general is a right thing to do.
> preventing users from MSI is asking for the performance degradation
> and IRQ resource conflicts (in case the IRQ line is shared).
> 
> Besides that, shouldn't it be rather the specific field in private (to
> 8250_pci) structure than constantly growing list?
>