Re: [PATCH v2] fpga: altera-cvp: Truncated bitstream error support

[Xu Yilun]

On Tue, May 31, 2022 at 09:53:56AM +0800, tien.sung.ang@xxxxxxxxx wrote:
> hi,
> 
> I will fix the "blank line" issues. 

Please don't top post.

And you don't have to clean too much context, otherwise people have to
search previous mails to remember what is it. People may not read the
mail at once.

> 
> >> The current AIB CvP firmware is not capable of handling a
> >> data stream smaller than 4096bytes. The firmware's limitation
> 
> >So why don't you check the image size on write_init(), and just prevent
> >the DMA writing at the very beginning?
> We try not to add any kind of logic in the CvP driver to read
> or understand the correctness of the bitstream. In future, this
> logic may change. The CvP driver to our best interest should

But you are adding the logic in altera_cvp_write() to keep the
correctness. What's the difference adding it somewhere else? And as I
can see, preventing the writing at the very beginning is a much cleaner
way.

> only be responsible to perform the transfer of the bitstream
> and perform the respective handshake. We have plans to have
> the CvP firmware on the other side be more intelligent to 
> handle errors and inform the CvP driver on why a certain 
> configuration session fails. We are in the process of making
> the next generation CvP to close this gaps.
> 
> >> To resolve this design limitation, both firmware and CvP
> >> driver made several changes. At the CvP driver, we just
> >> have to ensure that anything lesser than 4096bytes are
> >> padded with extra bytes. The CvP will then, initiate the
> >> tear-down by clearing the START_XFER and CVP_CONFIG bits.
> 
> > The driver pads the data block to 4096 bytes, then why the CvP still
> > should fail the reprograming?
> The block size of 4096bytes containts headers and signatures that
> comprise a bitstream to FPGA. Any of the missing pieces will 
> cause the FPGA configuration to fail at the SDM (Secure device
> manager). We are at Intel are just trying to avoid a user from
> using a bad data to reconfigure the FPGA via CvP and then, hitting
> into a fault that is non-recoverable. Only a hard-reboot of the 
> entire system can undo it. 
> 
> > If the image size is larger than 1 Page but is not aligned to 1 Page,
> > will the reprogramming still fail?
> Yes, the reconfiguration will fail. The above tear-down  is to prevent
> that CvP Hardware/firmware in the FPGA from entering into a dead-lock.

So if the image size is not aligned to 1 Page, the reprogram should fail
anyway, is it? Then I really recommend you fail the reprogramming at the
very beginning.

> 
> 
> >> +		memcpy(conf->send_buf, data, len);
> 
> >Any padding value is OK?
> Yes, any padding value is fine. The reason for this is that, the
> CvP Firmware/Hardware on the other side is using DMA to perform
> transfers of 4096bytes. The firmware will wait forever for the
> DMA to complete. The firmware is not able to check for status
> or respond to any errors while in the wait state. 
> 
> >> +		altera_cvp_send_block(conf, (const u32 *)conf->send_buf,
> >> +				      conf->priv->block_size);
> 
> >If the len equals block_size, is the copy still needed?
> Actually, not required. But to maintain a simple design, we use
> a common flow for all so you can skip the check.

I don't think so. We should make the code reasonable. Blindly copy the
whole buffer impacts all normal cases, and makes people confused. The
code seems shorter now, but not simpler, it takes people more time to
figure out why.

> 
> >> +	ret = altera_read_config_dword(conf, VSE_CVP_STATUS, &val);
> >> +	if (ret || (val & VSE_CVP_STATUS_CFG_ERR)) {
> >> +		dev_err(&mgr->dev, "CVP_CONFIG_ERROR!\n");
> 
> >So this new error checking covers the previous "latched error" case?
> Yes, because the new feature requires for the CvP driver to 
> be able to subsequently re-perform a fresh new CvP session after
> an error. With the above, the CvP driver will forever report
> an error and caused the entire CvP to be permanently in error. 
> This is not desirable and the user would need to power-cycle the
> or hard reboot the entire system to recover. Intel sees this as
> a terrible user-experience espcially as CvP is performed 
> normally remotely.
> 
> >> +	/* Allocate the 4096 block size transmit buffer */
> >> +	conf->send_buf = devm_kzalloc(&pdev->dev, conf->priv->block_size, GFP_KERNEL);
> 
> >If block_size == ALTERA_CVP_V1_SIZE, the copy is still needed?
> Yes, the firmware and hardware on the other side uses DMA
> and V1 to my understanding has a smaller buffer size. Intel 
> later improved this and introduced a new back-pressure credit 
> data flow control to help increase the performance.
> 
> >> +	if (!conf->send_buf) {
> >> +		ret = -ENOMEM;
> >> +		goto err_unmap;
> >> +	}
> 
> >Maybe it is better move the buffer allocation to write_init()
> The altera_cvp_conf has not been created yet at this point.
> The init routine in CvP so far just registers the driver.

I meant moving it to altera_cvp_write_init(). But now I think it is not
necessary. Bouncing buffer is not necessary at all.

Thanks,
Yilun