Thanks for your suggestions. The unmapping of the packet buffers is done only after the buffer has been completely transmitted or received. This is known by examining the completion information from the buffer descriptor. So, the possibility of the DMA happening while/after unmapping of buffers is remote - though there might be a bug there to look at. However, I wanted to clarify that I am speaking of unmapping of PCIe device regions, not of system packet buffers. The device registers start showing the 0xffffffff values. Regards, - priya -----Original Message----- From: Mulyadi Santosa [mailto:mulyadi.santosa@xxxxxxxxx] Sent: Thursday, February 14, 2008 7:13 AM To: Priya Suryanarayanan Cc: kernelnewbies@xxxxxxxxxxxx Subject: Re: Question on unmapping of PCIe mapped regions Hi.. On Feb 13, 2008 10:07 PM, Priya Suryanarayanan <priya.suryanarayanan@xxxxxxxxxx> wrote: > > > > > I own a Gigabit Ethernet driver which is being tested on a PCI Express PC > running Linux 2.6.18.1. > > > > For normal traffic (ping, ftp, telnet, WWW), there seems to be no problem. > When regression-testing with netperf, the system hangs randomly. After a lot > of debugging, I've discovered that this event coincides with the device > registers suddenly showing 0xffffffff. Debugging with Vmetro indicates that > the device registers do not really contain this value. Therefore, somehow > the space has got unmapped from the system virtual address space. I try to offer some possibilities from my own thinking: a. maybe the driver does something like zero copy ...in normal cases it works OK, but in some rare corner cases (read: huge load), card data transfer is not fast enough, so it hit race condition where it tries to copy data from user space while the page itself is going unmapped b. it is simply another race condition, but not related to user space. skbuff is released sooner while still the payload is still being copied to card's memory. I might be offering silly thoughts here.... regards, Mulyadi. -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx Please read the FAQ at http://kernelnewbies.org/FAQ
