Re: Freescale P2020 CPU Freeze over PCIe abort signal

[Eran Liberty <liberty@extricom.com>]

Eran Liberty wrote:
> This should probably go to the Freescale support, as it feels like a 
> hardware issue yet the end result is a very frozen Linux kernel so I 
> post here first...
>
> I have a programmable FPGA PCIe device connected to a Freescale's 
> P2020 PCIe port. As part of the bring-up tests, we are testing two 
> faulty scenarios:
> 1. The FPGA totally ignores the PCIe transaction.
> 2. The FPGA return a transaction abort.
>
> Both are plausible PCIe behavior and their should be outcome is 
> documented in the PCIe spec. The first should be terminated by the 
> transaction requestor timeout mechanism and raise an error, the second 
> should abort the transaction and raise and error.
>
> In P2020 if I do any of those the CPU is left hung over the transaction.
>
> something like:
> in_le32(addr)
>
> is turned into:
> 7c 00 04 ac     sync   7c 00 4c 2c     lwbrx   r0,0,r9
> 0c 00 00 00     twi     0,r0,0
> 4c 00 01 2c     isync
>
> assembly code, where in r9 (in this example) hold an address which is 
> physically mapped into the PCIe resource space.
>
> The CPU will hang over the load instruction.
>
> Just for the fun of it, I have wrote my own assembly function omitting 
> everything but the load instruction; still freeze.
> Replace "lwbrx" with a simple "lwz"; still freeze.
>
> It looks like the CPU snoozes till the PCIe transaction is done with 
> no timeouts, ignoring any abort signal.
>
> I am going to:
> A. Try to reach the Freescale support.
> B. Asked the FPGA designed to give me a new behavior that will stall 
> the PCIe transaction replay for 10 sec, but after those return ok.
> C. report back here with either A or B.
>
> If you have any ideas I would love to hear them.
>
> -- Liberty
>
Some more info:

As said the the FPGA designer provided me a PCIe device that will stall 
its response to a variable amount of time. The CPU became un-frozen 
after this amount of time. More over, we have found that in that period 
till it un-froze the PCIe core did a retry to that transaction over and 
over every 40 ms. This gave me the bright idea to look for the word 
"retry" in the Freescale documentation which rewarded me with these 
registers:

------------------------------------------------------- snip 
-------------------------------------------------------
16.3.2.3        PCI Express Outbound Completion Timeout Register
                (PEX_OTB_CPL_TOR)
The PCI Express outbound completion timeout register, shown in Figure 
16-4, contains the maximum wait
time for a response to come back as a result of an outbound non-posted 
request before a timeout condition
occurs.
 Offset 
0x00C                                                                                                
Access: Read/Write
         0   1              5     7   
8                                                                                      
31
     R
        TD            
—                                                            TC
     W
 Reset 0     0  0  0   0   0   0  0   0   0   0   1    0  0   0  0    
1   1  1    1   1  1   1   1   1  1   1   1  1  1   1  1
            Figure 16-4. PCI Express Outbound Completion Timeout 
Register (PEX_OTB_CPL_TOR)
Table 16-6 describes the PCI Express outbound completion timeout 
register fields.
                                 Table 16-6. PEX_OTB_CPL_TOR Field 
Descriptions
  Bits     Name                                                     
Description
   0        TD     Timeout disable. This bit controls the 
enabling/disabling of the timeout function.
                   0 Enable completion timeout
                   1 Disable completion timeout
  1–7        —     Reserved
 8–31       TC     Timeout counter. This is the value that is used to 
load the response counter of the completion timeout.
                   One TC unit is 8× the PCI Express controller clock 
period; that is, one TC unit is 20 ns at 400 MHz, and 30
                   ns at 266.66 MHz.
                   The following are examples of timeout periods based 
on different TC settings:
                   0x00_0000 Reserved
                   0x10_FFFF 22.28 ms at 400 MHz controller clock; 33.34 
ms at 266.66 MHz controller clock
                   0xFF_FFFF 335.54 ms at 400 MHz controller clock; 
503.31 ms at 266.66 MHz controller clock


16.3.2.4       PCI Express Configuration Retry Timeout Register
               (PEX_CONF_RTY_TOR)
The PCI Express configuration retry timeout register, shown in Figure 
16-5, contains the maximum time
period during which retries of configuration transactions which resulted 
in a CRS response occur.
 Offset 
0x010                                                                               
Access: Read/Write
         0  1     3   
4                                                                                     
31
     R
        RD     —                                                 TC
     W
 Reset 0    0  0  0  0   1  0  0  0  0   0  0   0  0  0  0   1  1  1  1  
1  1   1   1 1 1   1  1   1 1   1  1
           Figure 16-5. PCI Express Configuration Retry Timeout Register 
(PEX_CONF_RTY_TOR)
                            QorIQ P2020 Integrated Processor Reference 
Manual, Rev. 0
16-12                                                                                   
Freescale Semiconductor
                                                                                                 
PCI Express Interface Controller
Table 16-7 describes the PCI Express configuration retry timeout 
register fields.
                            Table 16-7. PEX_CONF_RTY_TOR Field Descriptions
  Bits  Name                                                     Description
   0     RD    Retry disable. This bit disables the retry of a 
configuration transaction that receives a CRS status response
               packet.
               0 Enable retry of a configuration transaction in response 
to receiving a CRS status response until the timeout
                  counter (defined by the PEX_CONF_RTY_TOR[TC] field) 
has expired.
               1 Disable retry of a configuration transaction regardless 
of receiving a CRS status response.
  1–3     —    Reserved
 4–31    TC    Timeout counter. This is the value that is used to load 
the CRS response counter.
               One TC unit is 8× the PCI Express controller clock 
period; that is, one TC unit is 20 ns at 400 MHz and 30 ns
               at 266.66 MHz.
               Timeout period based on different TC settings:
               0x000_0000        Reserved
               0x400_FFFF        1.34 s at 400 MHz controller clock, 
2.02 s at 266.66 MHz controller clock
               0xFFF_FFFF        5.37 s at 400 MHz controller clock, 
8.05 s at 266.66 MHz controller clock
------------------------------------------------------- snap 
-------------------------------------------------------

Now this is all nice on the paper, but what the P2020 seems to be doing 
in reality is
1. never expire
2. do re-tries even in the non configuration access

I am going to try to disable completion timeout and see if I get better 
behavior.

-- Liberty