[PATCH 03/15] ARM: tegra: use fixed PCI i/o mapping

(cutting down CCs for Tegra-specific discussion)

On 07/06/2012 02:36 PM, Stephen Warren wrote:

On 07/06/2012 02:16 PM, Stephen Warren wrote:

quoted

On 07/06/2012 02:11 PM, Rob Herring wrote:

quoted

On 07/06/2012 02:44 PM, Stephen Warren wrote:

quoted

On 07/06/2012 12:40 PM, Rob Herring wrote:

quoted

From: Rob Herring <redacted>

Move tegra PCI to fixed i/o mapping and remove io.h.

Thierry, since you're the Tegra PCIe expert right now, could you please
test and/or comment on this.

I did try testing this on next-20120705 on TrimSlice (i.e. the
PCIe-based Ethernet controller), but found that PCIe has stopped working
there due to "resource collisions". I know this used to work fairly
recently, since I tested it when I added the PCIe initialization call to
board-dt-tegra20.c. The PCIe messages are:

This is with my change and it works currently without?

Sorry, no, it's broken even without your change. Hence, I can't test the
impact of your change. Well, I saw the same failure with your patches
too, but that isn't really conclusive testing of your change:-)

Aha. The PCIe problem only shows up when booting TrimSlice using DT (in
next-20120705 or Tegra's for-next branch).

When booting using board files, PCIe works, both without and with your
patch, on top of next-20120705.

So, your patches are tested on Tegra and still working.

I haven't root-caused it, but I have found what changed that triggered
the problem:

When I first wrote the patch to board-dt-tegra20.c that brought DT
booting up to feature parity with non-DT boot, IIRC, I wrote a
late_initcall() to run whatever code I added. I'm sure I tested
TrimSlice PCIe when booting using DT then, since that was the whole
point of the patch.

Later, I revised the patch to sit on top of Shawn Guo's .init_late
machine descriptor patch, and ran the code from there instead. However,
this appears to have broken PCIe on Trimslice; I must have screwed up
the testing of that change. I don't know why this causes a behavior
difference though; perhaps some resource acquisition race condition.

But just reverting that change doesn't fix the problem; I need to switch
to a *subsys*_initcall() instead (which is in fact what the Trimslice
board file uses for PCIe init). In other words:

diff --git a/arch/arm/mach-tegra/board-dt-tegra20.c

b/arch/arm/mach-tegra/board-dt-tegra20.c
index 70a19a9..74aefaf 100644

--- a/arch/arm/mach-tegra/board-dt-tegra20.c
+++ b/arch/arm/mach-tegra/board-dt-tegra20.c

@@ -145,19 +145,20 @@ static struct {
 #endif
 };

-static void __init tegra_dt_init_late(void)
+static int __init tegra_dt_init_late(void)
 {
 	int i;

-	tegra_init_late();
-
 	for (i = 0; i < ARRAY_SIZE(board_init_funcs); i++) {
 		if (of_machine_is_compatible(board_init_funcs[i].machine)) {
 			board_init_funcs[i].init();
 			break;
 		}
 	}
+
+	return 0;
 }
+subsys_initcall(tegra_dt_init_late);

 static const char *tegra20_dt_board_compat[] = {
 	"nvidia,tegra20",

@@ -171,7 +172,7 @@ DT_MACHINE_START(TEGRA_DT, "nVidia Tegra20

(Flattened Device Tree)")
 	.handle_irq	= gic_handle_irq,
 	.timer		= &tegra_timer,
 	.init_machine	= tegra_dt_init,
-	.init_late	= tegra_dt_init_late,
+	.init_late	= tegra_init_late,
 	.restart	= tegra_assert_system_reset,
 	.dt_compat	= tegra20_dt_board_compat,
 MACHINE_END