mmap slower than ioctl

mandoray · 09-18-2014, 08:09 AM

I have written a driver that needs to access a PCI card's registers rapidly from a user application. This card provides a time stamp that is synchronized with other systems. The driver provides a mechanism to do this via an ioctl call, or by allowing the user to do a mmap and read the registers directly.

It appears that the ioctl call is faster and more consistent than reading the registers directly.

Any ideas why this is so?

smallpond · 09-20-2014, 07:43 AM

Seems unlikely. Mmap is just creating a mapping in the process page tables to the PCI address. Both the kernel and the process have to do the same slow PCI access. The ioctl just has the extra call into the kernel. Is your CPU under heavy load at the time? Have you tried changing your process priority?

mandoray · 09-22-2014, 05:37 AM

It does seem like the mmaped I/O should be quicker than the ioctl, that's why I asked to see if others had seen this. The machine has 16 cores and nothing but Linux and a few background services running. It's pretty much a dedicated development machine. I am not sure what Linux does when PCI memory space is mapped to a process. Perhaps there is a page fault every time it is accessed.

jpollard · 09-22-2014, 06:58 AM

It may depend on the amount of memory actualy being mapped - mmap requires the addresses to be on a page boundary, and of a page size.

Not sure what happens when it is not, as the kernel may be emulating the access via page faults.