the following patch was just integrated into master: commit 7b22d84d55386ee422fa77c1d8bce9dddeaaa231 Author: Timothy Pearson tpearson@raptorengineeringinc.com Date: Fri Aug 28 19:52:05 2015 -0500
drivers/pc80: Add optional spinlock for nvram CBFS access
When enabling the IOMMU on certain systems dmesg is spammed with I/O page faults like the following: AMD-Vi: Event logged [IO_PAGE_FAULT device=00:14.0 domain=0x000a address=0x000000fdf9103300 flags=0x0030]
Decoding the faulting address: 0x000000fdf9103300 fdf91x Hypertransport system management region 33 SysMgtCmd (System Management Command) = 0x33 3 Base Command Type = 0x3: STPCLK (Stop Clock request) 3 SMAF (System Management Action Field) = [3:1] = 0x1 1 Signal State Bit Map = [0] = 0x1
Therefore, the error appears to be triggered by an upstream C1E request.
This was eventually traced to concurrent access to the SP5100's SPI Flash controller by multiple APs during startup. Calls to the nvram read functions get_option and read_option call CBFS functions, which in turn make near-simultaneous requests to the SPI Flash controller, thus placing the SP5100 in an invalid state. This limitation is not documented in any public AMD errata, and was only discovered through considerable debugging effort.
Change-Id: I4e61b1ab767b1b7958ac7c1cf20eee41d2261bef Signed-off-by: Timothy Pearson tpearson@raptorengineeringinc.com Reviewed-on: https://review.coreboot.org/12061 Tested-by: build bot (Jenkins) Reviewed-by: Martin Roth martinroth@google.com
See https://review.coreboot.org/12061 for details.
-gerrit