From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andrew Bradford Date: Fri, 29 May 2015 14:27:46 -0400 Subject: [U-Boot] [PATCH] x86: baytrail: PCI is not always mapped to end of ram In-Reply-To: References: <1432321778-23960-1-git-send-email-andrew@bradfordembedded.com> <20150526121719.GA3952@rhino.kp.kodak.com> <20150527115308.GA4398@rhino.kp.kodak.com> Message-ID: <20150529182746.GA4380@rhino.kp.kodak.com> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: u-boot@lists.denx.de Hi Bin, On 05/29 13:00, Bin Meng wrote: > Hi Andrew, > > On Wed, May 27, 2015 at 7:53 PM, Andrew Bradford > wrote: > > Hi Bin, > > > > On 05/27 12:21, Bin Meng wrote: > >> Hi Andrew, > >> > >> On Tue, May 26, 2015 at 8:17 PM, Andrew Bradford > >> wrote: > >> > Hi Bin, > >> > > >> > On 05/23 23:50, Bin Meng wrote: > >> >> +Simon. > >> >> > >> >> Hi Andrew, > >> >> > >> >> On Sat, May 23, 2015 at 3:09 AM, wrote: > >> >> > From: Andrew Bradford > >> >> > > >> >> > PCI on Intel Baytrail is mapped to 0x80000000, which is not always at > >> >> > the end of SDRAM, such as when running with 4 GiB of SDRAM. The PCI bus > >> >> > memory mapping must stay within low memory and so when running with > > >> >> > 2 GiB of SDRAM, there is a hole in the SDRAM between 2 GiB and 4 GiB for > >> >> > memory mapped IO, such as PCI. > >> >> > >> >> Are you saying that if we mount 4GB DDR DIMM on the MinnowMax board, > >> >> the Intel FSP will only put 0~2G as system RAM space, and leave 2G~4G > >> >> as PCI address and other I/Os? > >> > > >> > Yes. If you mount 4 GiB of SDRAM onto an E3800 processor, then physical > >> > addresses from 0 to just below 2 GiB will be SDRAM (as per the HOBs) and > >> > also from 4 GiB to 6 GiB (also verified via the HOBs). The space from 2 > >> > GiB to 4 GiB will be mapped as various other regions. > >> > >> Ah, that's exactly the information I want :) > >> > >> > If you see section 4.1.1.1 (page 71 in the January 2015, Revision 3.6) > >> > E3800 datasheet, it shows that from 2 GiB to 4 GiB is mapped for PCI, > >> > Abort Page, Local APIC, and the Boot Vector. There's a lot of space in > >> > this area which appears unused, so I'm unsure as to why the area is so > >> > large. > >> > > >> > I have an Intel Valley Island board with E3825 and a 4 GiB SODIMM. I'm > >> > working on getting patches ready for this board but found that if I > >> > enabled all 4 GiB of SDRAM that the PCI bus would not function > >> > correctly. With this patch then the PCI bus functions (I'm able to do > >> > network operations with the RTL8111 Ethernet adapter). > >> > > >> >> I see from minnowmax.h, the PCI address starts from 0xd0000000. > >> >> > >> >> #define CONFIG_PCI_MEM_BUS 0xd0000000 > >> >> #define CONFIG_PCI_MEM_PHYS CONFIG_PCI_MEM_BUS > >> >> #define CONFIG_PCI_MEM_SIZE 0x10000000 > >> >> > >> >> #define CONFIG_PCI_PREF_BUS 0xc0000000 > >> >> #define CONFIG_PCI_PREF_PHYS CONFIG_PCI_PREF_BUS > >> >> #define CONFIG_PCI_PREF_SIZE 0x10000000 > >> > > >> > I see that hose->regions+0 is set to CONFIG_PCI_MEM_BUS and > >> > hose->regions+2 is set to CONFIG_PCI_PREF_BUS. However I'm modifying > >> > hose->regions+3. So the values from minnowmax.h *are* being used. I'm > >> > not yet that familiar with PCI configuration, so likely I'm not fully > >> > understanding how u-boot sets this up. > >> > > >> > >> The regions+3 is used by the inbound access, eg: PCI device access to > >> system memory. > >> > >> > Possibly my address of 0x80000000 is not correct, even though it works > >> > for me. But 0x80000000 is where it was being placed before, since it > >> > was going at the end of SDRAM (2GiB on minnowmax). > >> > > >> > >> You understanding is correct. We should only open 2GiB inbound memory > >> window for PCI devices. > > > > But, if you have less than 2 GiB of memory, my patch in theory *could* > > break things, right?. It seems like a better approach would be to limit > > the size to the lesser of 0x80000000 and gd->ram_size. Does that make > > sense? > > > > I think 2GB is a safe value and won't break things. Region 0 and > region 3 should not overlap. > > >> > If I artificially limit the amount of SDRAM by setting the fsp > >> > configuration to memory-down and then setting the DRAM configuration > >> > values such that I mimmic 1 GiB or 2 GiB of SDRAM, having my patch still > >> > provides access to the PCI bus, so with my patch there should be no > >> > adverse affects on E3800 systems that have less than 4 GiB of SDRAM. > >> > But without my patch, when running with >=4 GiB of SDRAM, PCI accesses > >> > end up returning "pci_hose_bus_to_phys: invalid physical address" > >> > errors. > > Can you add some printf to show all of the pci_hose_bus_to_phys() > parameters' value here when 4GB RAM is mounted? I'd like to understand > how the message "pci_hose_bus_to_phys: invalid physical address" is > produced. Patch of my changes to enable reporting of physical addresses being used looks like: diff --git a/drivers/pci/pci_common.c b/drivers/pci/pci_common.c index b9ff23f..3babcb7 100644 --- a/drivers/pci/pci_common.c +++ b/drivers/pci/pci_common.c @@ -205,7 +205,7 @@ int __pci_hose_bus_to_phys(struct pci_controller *hose, return 0; } } - + printf("__pci_hose_bus_to_phys() failed!\n"); return 1; } @@ -237,6 +237,7 @@ phys_addr_t pci_hose_bus_to_phys(struct pci_controller *hose, if (ret) puts("pci_hose_bus_to_phys: invalid physical address\n"); + printf("bus_addr: 0x%x \t flags: 0x%lx \t phys_addr: %llx\n", bus_addr, flags, phys_addr); return phys_addr; } When I only configure 2GB of RAM, this prints bus and physical addresses in the 0x7adc0000 range when interfacing to the RTL8111. Everything works as expected, and I get only 1 copy of my "failed" message for each printing of my bus_addr, flags, and phys_addr data. When I switch to 4 GB of RAM configured, now __pci_hose_bus_to_phys() will not perform the modification to the physical address that gets passed to it. Then I see accesses to phys_addr to 0x7adc0000 range but phys_addresses are printed as 0 and I get 2 copies of my "failed" message for each printing of my bus_addr, flags, and phys_addr data. Ethernet does not work in this case. When I change to 1 GB of RAM, everything works and the phys and bus addresses are in the 0x3adc0000 range and I only get 1 "failed" message for each printing of bus_addr, flags, and phys_addr data. Just like in the 2GB case. Once I apply my patch to change region+3 to 0x80000000, then both 1GB and 2 GB work the same as before, but now with 4 GB things work and the output from my new printf()s matches the 2GB operation with phys and bus addresses in the 0x7adc0000 range. Sorry I'm relaying information rather than giving direct output from the board, but I don't yet have a UART setup on my Valley Island from within u-boot, I'm doing all my work via the vesa console and a USB keyboard. Thanks, Andrew