Re: [Qemu-block] Virtio-BLK/SCSI write requests and data payload checksums

[Peter Lieven]

> Am 07.06.2017 um 01:17 schrieb Paolo Bonzini <address@hidden>:
> 
> 
> 
> ----- Original Message -----
>> From: "Peter Lieven" <address@hidden>
>> To: "qemu block" <address@hidden>
>> Cc: "Paolo Bonzini" <address@hidden>, address@hidden, address@hidden, "Max 
>> Reitz" <address@hidden>,
>> "Michael S. Tsirkin" <address@hidden>
>> Sent: Tuesday, June 6, 2017 10:13:51 PM
>> Subject: Virtio-BLK/SCSI write requests and data payload checksums
>> 
>> Hi,
>> 
>> 
>> ich have spend several hours debugging a strange checksum error issue and
>> finally found the cause, but I am totally unsure if whats happening
>> is correct or not.
>> 
>> Imagine a Protocol like iSCSI which has a Data Digest and which receives its
>> data via zero copy straight from the guest kernel through Qemu and
>> 
>> then out of the socket to the network. It calculates the CRC32C data digest
>> from the buffer it gets from Qemu. It totally relies thereby that the buffer
>> is not mangled after the write request has been submitted. But it seems this
>> assumption is not true.
>> 
>> [...] the contents of the buffer submitted to Qemu change while Qemu 
>> processes
>> the write.
>> In the end several write requests follow and the /etc/network/interfaces file
>> has the right content, but
>> is this at all legal?
> 
> Yes, it's ugly but it's legal.  It probably doesn't happen on real hardware
> that computes the checksum after or during DMA and has some kind of buffer
> inside the board.  But on virt there is only one copy until we reach the 
> actual
> physical hardware.

okay, so it's no bug. and there is no way to detect that the page is dirty 
(again). so i need a bounce buffer if i have data checksums?

Peter

> 
> Paolo
> 
>> If yes, checksum calculation of payload and zero copy
>> would not work together.
>