[U-Boot] [PATCH v2] usb: fix usb_stor_read/write on DM

Fri Jul 21 10:48:58 UTC 2017

+Tom for question below

Hi,

On 20 July 2017 at 03:40, Marek Vasut <marex at denx.de> wrote:
> On 07/20/2017 11:38 AM, Bin Meng wrote:
>> +Simon,
>>
>> On Thu, Jul 20, 2017 at 5:00 PM, Marek Vasut <marex at denx.de> wrote:
>>> On 07/20/2017 09:49 AM, Masahiro Yamada wrote:
>>>> 2017-07-20 2:33 GMT+09:00 Marek Vasut <marex at denx.de>:
>>>>> On 07/19/2017 05:38 PM, Masahiro Yamada wrote:
>>>>>> 2017-07-15 21:57 GMT+09:00 Marek Vasut <marex at denx.de>:
>>>>>>> On 07/15/2017 01:30 AM, Benoît Thébaudeau wrote:
>>>>>>>> On Sat, Jul 15, 2017 at 12:06 AM, Marek Vasut <marex at denx.de> wrote:
>>>>>>>>> On 07/14/2017 11:46 PM, Benoît Thébaudeau wrote:
>>>>>>>>>> On Fri, Jul 14, 2017 at 1:50 PM, Marek Vasut <marex at denx.de> wrote:
>>>>>>>>>>> On 07/14/2017 01:03 PM, Masahiro Yamada wrote:
>>>>>>>>>>>> 2017-07-14 19:07 GMT+09:00 Marek Vasut <marex at denx.de>:
>>>>>>>>>>>>> On 07/14/2017 04:31 AM, Masahiro Yamada wrote:
>>>>>>>>>>>>>> Prior to DM, we could not enable different types of USB controllers
>>>>>>>>>>>>>> at the same time.  DM was supposed to loosen the limitation.  It is
>>>>>>>>>>>>>> true that we can compile drivers, but they do not work.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> For example, if EHCI is enabled, xHCI fails as follows:
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>   => usb read 82000000 0 2000
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>   USB read: device 0 block # 0, count 8192 ... WARN halted endpoint, queueing URB anyway.
>>>>>>>>>>>>>>   Unexpected XHCI event TRB, skipping... (3fb54010 00000001 13000000 01008401)
>>>>>>>>>>>>>>   BUG: failure at drivers/usb/host/xhci-ring.c:489/abort_td()!
>>>>>>>>>>>>>>   BUG!
>>>>>>>>>>>>>>   ### ERROR ### Please RESET the board ###
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> The cause of the error seems the following code:
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>   #ifdef CONFIG_USB_EHCI_HCD
>>>>>>>>>>>>>>   /*
>>>>>>>>>>>>>>    * The U-Boot EHCI driver can handle any transfer length as long as there is
>>>>>>>>>>>>>>    * enough free heap space left, but the SCSI READ(10) and WRITE(10) commands are
>>>>>>>>>>>>>>    * limited to 65535 blocks.
>>>>>>>>>>>>>>    */
>>>>>>>>>>>>>>   #define USB_MAX_XFER_BLK    65535
>>>>>>>>>>>>>>   #else
>>>>>>>>>>>>>>   #define USB_MAX_XFER_BLK    20
>>>>>>>>>>>>>>   #endif
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> To fix the problem, choose the chunk size at run-time for CONFIG_BLK.
>>>>>>>>>>>>>
>>>>>>>>>>>>> What happens if CONFIG_BLK is not set ?
>>>>>>>>>>>>
>>>>>>>>>>>> USB_MAX_XFER_BLK is chosen.
>>>>>>>>>>>
>>>>>>>>>>> And can we fix that even for non-CONFIG_BLK ?
>>>>>>>>>>>
>>>>>>>>>>>>> Why is it 20 for XHCI anyway ?
>>>>>>>>>>>>
>>>>>>>>>>>> You are the maintainer.
>>>>>>>>>>>> (I hope) you have better knowledge with this.
>>>>>>>>>>>
>>>>>>>>>>> Heh, way to deflect the question. I seem to remember some discussion
>>>>>>>>>>> about the DMA (?) limitation on XHCI, but I'd have to dig through the ML
>>>>>>>>>>> archives myself.
>>>>>>>>>>>
>>>>>>>>>>>> Looks like the following commit was picked up by you.
>>>>>>>>>>>
>>>>>>>>>>> 5 years ago, way before DM was what it is today .
>>>>>>>>>>
>>>>>>>>>> And even way before the introduction of XHCI into U-Boot, which means
>>>>>>>>>> that this 20 was targeting OHCI or proprietary HCDs, not XHCI.
>>>>>>>>>> USB_MAX_READ_BLK was already set to 20 in the initial revision of
>>>>>>>>>> usb_storage.c. As I said in the commit message, this 20 was certainly
>>>>>>>>>> not optimal for these non-EHCI HCDs, but it restored the previous
>>>>>>>>>> (i.e. pre-5dd95cf) behavior for these HCDs instead of using the 5 * 4
>>>>>>>>>> KiB code, which was specific to ehci-hcd.c at that time. Without
>>>>>>>>>> knowing the rationale for the legacy 20 blocks, the safest approach
>>>>>>>>>> for non-EHCI HCDs was to use this value in order to avoid breaking a
>>>>>>>>>> platform or something. Looking at ohci-hcd.c, it limits the transfer
>>>>>>>>>> size to (N_URB_TD - 2) * 4 KiB, with N_URB_TD set to 48, so the
>>>>>>>>>> maximum number of transfers would depend on the MSC block size.
>>>>>>>>>> dwc2.c, isp116x-hcd.c, r8a66597-hcd.c, and sl811-hcd.c do not seem to
>>>>>>>>>> have any limit caused by these drivers. The limit with the current
>>>>>>>>>> XHCI code seems to be 64 * 64 KiB. So, nowadays, USB_MAX_XFER_BLK
>>>>>>>>>> could be set to 65535 for all HCDs but OHCI and XHCI, which require
>>>>>>>>>> specific rules depending on the MSC block size.
>>>>>>>>>
>>>>>>>>> For whatever reason, something tells me that setting the block size to
>>>>>>>>> 64k for XHCI broke things, but I cannot locate the thread. But there's
>>>>>>>>> something in the back of my head ...
>>>>>>>>
>>>>>>>> Indeed: according to what I said above, USB_MAX_XFER_BLK cannot be set
>>>>>>>> to 65535 for XHCI. With an MSC block size of blksz = 512 bytes /
>>>>>>>> block, USB_MAX_XFER_BLK can be set to at most 1 segment *
>>>>>>>> (TRBS_PER_SEGMENT = 64 TRBs / segment) * (TRB_MAX_BUFF_SIZE = 65536
>>>>>>>> bytes / TRB) / blksz = 8192 blocks for XHCI. And for OHCI, the limit
>>>>>>>> is (N_URB_TD - 2 = 46 TDs) * (4096 bytes / TD) / blksz = 368 blocks.
>>>>>>>> The buffer alignment may also have to be taken into account to adjust
>>>>>>>> these values, which would require a USB_MAX_XFER_BLK(host_if, start,
>>>>>>>> blksz) macro or function. USB_MAX_XFER_BLK can however be set to 65535
>>>>>>>> regardless of blksz for all the other HCDs (i.e. EHCI, dwc2.c,
>>>>>>>> isp116x-hcd.c, r8a66597-hcd.c, and sl811-hcd.c).
>>>>>>>
>>>>>>> That's probably what I was looking for, thanks.
>>>>>>>
>>>>>>
>>>>>>
>>>>>> So, how shall we handle this?
>>>>>>
>>>>>> If somebody can fix this in a correct way,
>>>>>> I am happy to hand over this.
>>>>>
>>>>> Any way to fix it for !CONFIG_BLK ?
>>>>
>>>>
>>>> common/usb_storage.c is sprinkled with ugly #ifdef CONFIG_BLK
>>>>
>>>> IIUC, !CONFIG_BLK code will be removed after migration.
>>>>
>>>> Is it worthwhile to save !CONFIG_BLK case?
>>>
>>> Hmmmmmm, sigh. When is the migration happening, how far is it ?
>>
>> One idea is to force all board to switch to driver model at a preset
>> timeline. After the deadline, boards do not switch to DM will get
>> dropped by the mainline. I noticed that not all boards are actively
>> maintained...
>
> Be my guest, there's a few which I'd like to see removed myself :-)

That makes sense although I'm not sure what the deadline should be.
CONFIG_BLK is invasive and it is a pain to carry the #ifdefs.

Maybe end of year, or is that too short?

>
> --
> Best regards,
> Marek Vasut