Message ID | 20150318000652.GJ31346@atomide.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
> And we also need to populate the tables along the lines of 27b7d5f3cc49 > ("bus: omap_l3_noc: Add AM4372 interconnect error data"). I think intercon data should be in DT rather than some hardcoded table. > Do the below ranges match your JTAG results? Yup. Based a memory dump I had done earlier, combined with the high degree of similarity with centaurus and filling in the remaining blanks with my best guess yields: 44000000 host L3F 44000100 target 01 dmm port 0 44000200 target 02 dmm port 1 44000300 target 03 iva 0 sl2 ? 44000400 target 04 iva 1 sl2 ? 44000500 target 05 dsp sdma 44000600 target 06 expansion port 44000700 target 07 edma tc 0 44000800 target 08 edma tc 1 44000900 target 09 edma tc 2 44000a00 target 0a edma tc 3 44000b00 target 0b edma cc 44000c00 target 23 system mmu 44000d00 target 25 iva 2 sl2 ? 44000e00 flagmux l3f 44000f00 flagmux top 44001000 statcoll 44001400 statcoll 44001800 statcoll 44001c00 bw-regulator 44001d00 bw-regulator 44001e00 bw-regulator 44001f00 bw-regulator 44002000 bw-regulator 44002100 bw-regulator 44002200 bw-regulator 44002300 bw-regulator 44002400 bw-regulator 44400000 host L3M 44400100 target 0d ducati 44400200 target 0e sgx 44400300 target 0f pcie 44400400 target 10 ocmc ram 44400500 target 11 vcp 44400600 target 12 iva config 44400700 target 13 iss 44400800 target 14 tppss 44400900 target 15 l4hs port 0 44400a00 target 16 secss 44400b00 target 24 l4hs port 1 44400c00 target 26 mmc 2 44400d00 target 1f l3_instr 44400e00 flagmux L3M 44401000 statcoll 44800000 host L3S 44800100 target 18 vlynq ? 44800200 target 19 mcbsp 44800300 target 1a hdmi 44800400 target 1b l4fw 44800500 target 1c l4ls port 0 44800600 target 1d l4ls port 1 44800700 target 1e gpmc 44800800 target 20 mcasp 0 44800900 target 21 mcasp 1 44800a00 target 22 mcasp 2 44800b00 target 27 usb 44800c00 flagmux L3S Note BTW that centaurus (in contrast with netra and subarctic) actually has a pretty good interconnect chapter with full register maps (except for the L3 firewals). > That got me wondering if we can actually scan that data based > on the ranges below as target is 00130001 and flagmux 00370001. > We would be missing the names, but that would be still less data > to pile up in the kernel :) If some module is disabled, then it > should never produce errors so it seems safe to scan the data.. Note that reading invalid addresses while scanning does yield bus errors you'd need to trap. There can be gaps, and some modules can be larger than 0x100 (e.g. statcoll is variable in size), but other than that it's just a matter of reading words in increments of 0x100 and match with one of the known types: // 13'0001 target // 1a'0001 host // 2c'0001 bw_limiter // 2d'0001 rate_adapter // 31'0001 bw_regulator // 37'0001 flagmux // 38'0001 pwr_discon // 3a'0001 statcoll The statcoll content should read as zeros after reset, so no risk in accidently detecting a module inside it. > Can we dig even more info? Much can be auto-detected indeed, although it may be preferred to use auto-detection to generate the data, assign missing names, and then have other users use that data file. If you know where the FlexNOC L3 intercon registers themselves are located (called the "service network" in Vayu), then you can scan the L3 with the following steps: 1. Detect all FlexNOC components as mentioned above. The components we care about for now are Target NIUs and Flagmuxes, and optionally the Host to decode errors while scanning the service network itself. In this step you learn the association Target NIU regbase <-> Target NIU ID since this is indicated in one of the regs. Make sure error logging is enabled for all targets. 2. Disable all L3 Target NIUs. You will temporarily lose access to all peripherals, OCMC ram, and on subarctic also DDR memory (I think on devices with a DMM you'd retain access to DDR memory). Note that the service network itself isn't considered an L3 target. 3. Probe the whole address space with reads or writes, while trapping data aborts (or use posted writes, those won't generate a data abort but only signal the L3 error irq). The smallest L3 ranges I've seen so far were 64 KB, so scanning the whole address space would require 2^16 accesses. Shouldn't take too long. Ranges which are not routed to L3 targets (e.g. the service networks themselves, or MPUSS-local peripherals) should be manually excluded. 4. For each access, examine the Flagmux modules and Target NIU error logs. (Then clear the error log for the next access) You should now learn: a. which target was reached at that address b. to which target-local address it was mapped c. which bit of which flagmux is used for errors from this target d. to which bit of the top flagmux this flagmux connects Note that usually a Target NIU will have its regs in the service network of the L3 it belongs to, and report errors to the associated Flagmux. Centaurus is a counter-example since for some odd reason its DMM Target NIUs regs moved to the L3M service network, even though they attach to the L3F and report errors to the L3F flagmux. Distinguishing the top flagmux from the others, or even more general flagmux topologies, can also be deduced programmatically but I'm not sure that's worth the effort. The target which seems to be reachable at many many address ranges, none of which documented as containing any peripheral, is your "error target" where unroutable packets are sent to to generate an error reply. 5. Reenable all target NIUs. 6. Identify the L3 targets. This will be semi-automatic at best, since there's no generally reliable way to do this. Not all peripherals have a highlander-style revision register, and even the ones that do sometimes have it at a non-zero offset or failed to pick a unique function code to identify the module. (Or split it into two 16-bit halves in consecutive 32-bit registers... nicely done, omap-i2c) One interesting option is keeping a database of L3 targets across multiple devices and taking advantage of design reuse: if you find a target whose NIU ID and mapped memory ranges match those of a specific target in another device, then there's a good chance they are in fact the same. This is highly visible in centaurus vs subarctic for example. 7. Some targets will be other interconnects; proceed with inspecting those. (Fortunately the L4 intercons are easier to scan). Matthijs -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
* Matthijs van Duin <matthijsvanduin@gmail.com> [150318 01:33]: > > And we also need to populate the tables along the lines of 27b7d5f3cc49 > > ("bus: omap_l3_noc: Add AM4372 interconnect error data"). > > I think intercon data should be in DT rather than some hardcoded table. Yeah so it seems. > > Do the below ranges match your JTAG results? > > Yup. Based a memory dump I had done earlier, combined with the high > degree of similarity with centaurus and filling in the remaining > blanks with my best guess yields: > > 44000000 host L3F > 44000100 target 01 dmm port 0 > 44000200 target 02 dmm port 1 > 44000300 target 03 iva 0 sl2 ? > 44000400 target 04 iva 1 sl2 ? > 44000500 target 05 dsp sdma > 44000600 target 06 expansion port > 44000700 target 07 edma tc 0 > 44000800 target 08 edma tc 1 > 44000900 target 09 edma tc 2 > 44000a00 target 0a edma tc 3 > 44000b00 target 0b edma cc > 44000c00 target 23 system mmu > 44000d00 target 25 iva 2 sl2 ? > 44000e00 flagmux l3f > 44000f00 flagmux top > 44001000 statcoll Oh there's a gap here and it goes all the way to 2400.. Is there a gap on am335x too? > 44001400 statcoll > 44001800 statcoll > 44001c00 bw-regulator > 44001d00 bw-regulator > 44001e00 bw-regulator > 44001f00 bw-regulator > 44002000 bw-regulator > 44002100 bw-regulator > 44002200 bw-regulator > 44002300 bw-regulator > 44002400 bw-regulator > > 44400000 host L3M > 44400100 target 0d ducati > 44400200 target 0e sgx > 44400300 target 0f pcie > 44400400 target 10 ocmc ram > 44400500 target 11 vcp > 44400600 target 12 iva config > 44400700 target 13 iss > 44400800 target 14 tppss > 44400900 target 15 l4hs port 0 > 44400a00 target 16 secss > 44400b00 target 24 l4hs port 1 > 44400c00 target 26 mmc 2 > 44400d00 target 1f l3_instr > 44400e00 flagmux L3M > 44401000 statcoll > > 44800000 host L3S > 44800100 target 18 vlynq ? > 44800200 target 19 mcbsp > 44800300 target 1a hdmi > 44800400 target 1b l4fw > 44800500 target 1c l4ls port 0 > 44800600 target 1d l4ls port 1 > 44800700 target 1e gpmc > 44800800 target 20 mcasp 0 > 44800900 target 21 mcasp 1 > 44800a00 target 22 mcasp 2 > 44800b00 target 27 usb > 44800c00 flagmux L3S OK that's great, thanks for the data! > Note BTW that centaurus (in contrast with netra and subarctic) > actually has a pretty good interconnect chapter with full register > maps (except for the L3 firewals). OK that's good to hear. > > That got me wondering if we can actually scan that data based > > on the ranges below as target is 00130001 and flagmux 00370001. > > We would be missing the names, but that would be still less data > > to pile up in the kernel :) If some module is disabled, then it > > should never produce errors so it seems safe to scan the data.. > > Note that reading invalid addresses while scanning does yield bus > errors you'd need to trap. There can be gaps, and some modules can be > larger than 0x100 (e.g. statcoll is variable in size), but other than > that it's just a matter of reading words in increments of 0x100 and > match with one of the known types: > // 13'0001 target > // 1a'0001 host > // 2c'0001 bw_limiter > // 2d'0001 rate_adapter > // 31'0001 bw_regulator > // 37'0001 flagmux > // 38'0001 pwr_discon > // 3a'0001 statcoll > > The statcoll content should read as zeros after reset, so no risk in > accidently detecting a module inside it. > > > Can we dig even more info? > > Much can be auto-detected indeed, although it may be preferred to use > auto-detection to generate the data, assign missing names, and then > have other users use that data file. > > If you know where the FlexNOC L3 intercon registers themselves are > located (called the "service network" in Vayu), then you can scan the > L3 with the following steps: > > 1. Detect all FlexNOC components as mentioned above. The components we > care about for now are Target NIUs and Flagmuxes, and optionally the > Host to decode errors while scanning the service network itself. In > this step you learn the association Target NIU regbase <-> Target NIU > ID since this is indicated in one of the regs. Make sure error > logging is enabled for all targets. > > 2. Disable all L3 Target NIUs. You will temporarily lose access to all > peripherals, OCMC ram, and on subarctic also DDR memory (I think on > devices with a DMM you'd retain access to DDR memory). Note that the > service network itself isn't considered an L3 target. Hmm well we could that in the kernel using the SRAM code as long as it does not reset the devices. > 3. Probe the whole address space with reads or writes, while trapping > data aborts (or use posted writes, those won't generate a data abort > but only signal the L3 error irq). The smallest L3 ranges I've seen > so far were 64 KB, so scanning the whole address space would require > 2^16 accesses. Shouldn't take too long. Ranges which are not routed to > L3 targets (e.g. the service networks themselves, or MPUSS-local > peripherals) should be manually excluded. OK the driver would have to be enabled with interrupts while scanning. > 4. For each access, examine the Flagmux modules and Target NIU error > logs. (Then clear the error log for the next access) You should now > learn: > a. which target was reached at that address > b. to which target-local address it was mapped > c. which bit of which flagmux is used for errors from this target > d. to which bit of the top flagmux this flagmux connects > > Note that usually a Target NIU will have its regs in the service > network of the L3 it belongs to, and report errors to the associated > Flagmux. Centaurus is a counter-example since for some odd reason its > DMM Target NIUs regs moved to the L3M service network, even though > they attach to the L3F and report errors to the L3F flagmux. > > Distinguishing the top flagmux from the others, or even more general > flagmux topologies, can also be deduced programmatically but I'm not > sure that's worth the effort. > > The target which seems to be reachable at many many address ranges, > none of which documented as containing any peripheral, is your "error > target" where unroutable packets are sent to to generate an error > reply. > > 5. Reenable all target NIUs. > > 6. Identify the L3 targets. This will be semi-automatic at best, since > there's no generally reliable way to do this. Not all peripherals have > a highlander-style revision register, and even the ones that do > sometimes have it at a non-zero offset or failed to pick a unique > function code to identify the module. (Or split it into two 16-bit > halves in consecutive 32-bit registers... nicely done, omap-i2c) > > One interesting option is keeping a database of L3 targets across > multiple devices and taking advantage of design reuse: if you find a > target whose NIU ID and mapped memory ranges match those of a specific > target in another device, then there's a good chance they are in fact > the same. This is highly visible in centaurus vs subarctic for > example. Yes sounds like that would be probably the best way to go eventually. > 7. Some targets will be other interconnects; proceed with inspecting > those. (Fortunately the L4 intercons are easier to scan). Thanks for all the info :) Regards, Tony -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
> Matthijs van Duin <matthijsvanduin@gmail.com> wrote: > 44400500 target 11 vcp > 44800100 target 18 vlynq ? scrap these guesses: I don't think Netra has a vcp module, and I had only filled in vlynq based on it being the only remaining target NIU. Once the memory ranges for the targets have been determined they should be easy to identify conclusively. On 18 March 2015 at 17:54, Tony Lindgren <tony@atomide.com> wrote: > Oh there's a gap here and it goes all the way to 2400.. Is there > a gap on am335x too? There are indeed. See my work-in-progress spreadsheet on its L3 interconnect: http://goo.gl/miyCQw As with the memory map, there are some "ghost" modules which I'm reasonably sure don't exist, even though they have a target NIU and clkctrl registers. I haven't yet checked which initiator the sole bandwidth regulator on subarctic is attached to. Plausible options are SGX, the LCD controller, or one of the EDMA TCs. I also just noticed how *big* its statistics collectors are. o.O For comparison, the ones on netra and centaurus are 0x200 (statcoll 0 and 3) or 0x400 (statcoll 1 and 2). >> 2. Disable all L3 Target NIUs. You will temporarily lose access to all >> peripherals, OCMC ram, and on subarctic also DDR memory (I think on >> devices with a DMM you'd retain access to DDR memory). Note that the >> service network itself isn't considered an L3 target. > > Hmm well we could that in the kernel using the SRAM code as long as > it does not reset the devices. If you run with MMU disabled then you'd have to run out of the 64K internal SRAM, and catch data aborts since all writes will be non-posted. With MMU enabled, locking code and MMU translation tables into the (much larger) L2 cache makes more sense. I *think* disabling the NIU only blocks access to the module and doesn't affect the module operationally. I'd be inclined to put external RAM into self-refresh during the procedure just to be sure, in case disabling NIUs results in clock gating (which would otherwise be hard to notice). > OK the driver would have to be enabled with interrupts while scanning. Polling makes more sense, especially since nothing else can be going on at the same time anyway. I still think a separate tool that scrapes as much data as possible from the target and pours it into DT form makes more sense than doing it in a kernel driver, especially since discovered targets may often still need manual identification. Also, probing discovered targets requires you enable the clock(s) they need and if necessary release local reset. Auto-discovery of these associations may also be possible, but probably would become a rather lenghy procedure. A separate tool can also leave a record of what it's doing in SRAM and enable the watchdog to be able to recover from probes that lock up the system (and log a warning about them). Also, on HS devices, it would surprise me if you're allowed to access any interconnect registers at all. Matthijs -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
--- a/arch/arm/boot/dts/am33xx.dtsi +++ b/arch/arm/boot/dts/am33xx.dtsi @@ -105,7 +105,10 @@ * the whole bus hierarchy. */ ocp { - compatible = "simple-bus"; + compatible = "ti,am335-l3-noc", "simple-bus"; + reg = <0x44000000 0x1200 + 0x44800000 0x1000>; + interrupts = <9 10>; #address-cells = <1>; #size-cells = <1>; ranges; --- a/arch/arm/boot/dts/dm816x.dtsi +++ b/arch/arm/boot/dts/dm816x.dtsi @@ -58,8 +58,10 @@ * the whole bus hierarchy. */ ocp { - compatible = "ti,omap4-l3-noc", "simple-bus"; - reg = <0x44000000 0x10000>; + compatible = "ti,dm816-l3-noc", "simple-bus"; + reg = <0x44000000 0x1200 + 0x44400000 0x1200 + 0x44800000 0x1000>; interrupts = <9 10>; #address-cells = <1>; #size-cells = <1>;