Skip to contentSkip to conversation

sh: extend INTC with optional ioremap() support

PATCH In review Magnus Damm 2 participants 3 messages Feb 10, 2010 SuperH

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sh: extend INTC with optional ioremap() support

From: Magnus Damm <[email protected]>

Add the member "io_window" to intc_desc to allow passing a physical memory window to the INTC code. The INTC code will ioremap the window if present.

Required to support INTCS on SH-Mobile G-series.

---

 drivers/sh/intc.c       |   14 ++++++++++++++
 include/linux/sh_intc.h |    3 +++
 2 files changed, 17 insertions(+)
drivers/sh/intc.c +14 -0
@@ -45,6 +45,8 @@ struct intc_handle_int {
 
 struct intc_desc_int {
 	struct list_head list;
+	unsigned long virt_base;
+	unsigned long phys_base;
 	struct sys_device sysdev;
 	pm_message_t state;
 	unsigned long *reg;
@@ -425,6 +427,9 @@ static unsigned int __init intc_get_reg(
 {
 	unsigned int k;
 
+	address -= d->phys_base;
+	address += d->virt_base;
+
 	for (k = 0; k < d->nr_reg; k++) {
 		if (d->reg[k] = address)
 			return k;
@@ -774,6 +779,9 @@ static unsigned int __init save_reg(stru
 				    unsigned int smp)
 {
 	if (value) {
+		value -= d->phys_base;
+		value += d->virt_base;
+
 		d->reg[cnt] = value;
 #ifdef CONFIG_SMP
 		d->smp[cnt] = smp;
@@ -800,6 +808,12 @@ void __init register_intc_controller(str
 	INIT_LIST_HEAD(&d->list);
 	list_add(&d->list, &intc_list);
 
+	if (desc->io_window) {
+		d->phys_base = desc->io_window->start;
+		d->virt_base = (unsigned long)ioremap_nocache(d->phys_base,
+					       resource_size(desc->io_window));
+	}
+
 	d->nr_reg = hw->mask_regs ? hw->nr_mask_regs * 2 : 0;
 	d->nr_reg += hw->prio_regs ? hw->nr_prio_regs * 2 : 0;
 	d->nr_reg += hw->sense_regs ? hw->nr_sense_regs : 0;
include/linux/sh_intc.h +3 -0
@@ -1,6 +1,8 @@
 #ifndef __SH_INTC_H
 #define __SH_INTC_H
 
+#include <linux/ioport.h>
+
 typedef unsigned char intc_enum;
 
 struct intc_vect {
@@ -71,6 +73,7 @@ struct intc_hw_desc {
 
 struct intc_desc {
 	char *name;
+	struct resource *io_window;
 	intc_enum force_enable;
 	struct intc_hw_desc hw;
 };
#1 lore raw
Quoted 18 lines · Magnus Damm

From: Magnus Damm <[email protected]>

Add the member "io_window" to intc_desc to allow passing a physical memory window to the INTC code. The INTC code will ioremap the window if present.

Required to support INTCS on SH-Mobile G-series.

---

 drivers/sh/intc.c       |   14 ++++++++++++++
 include/linux/sh_intc.h |    3 +++
 2 files changed, 17 insertions(+)

drivers/sh/intc.c struct intc_handle_int { · line 52

Quoted 8 lines
drivers/sh/intc.c +2 -0
@@ -45,6 +45,8 @@ struct intc_handle_int {
 
 struct intc_desc_int {
 	struct list_head list;
+	unsigned long virt_base;
+	unsigned long phys_base;
 	struct sys_device sysdev;
 	pm_message_t state;
 	unsigned long *reg;

phys_base can be a phys_addr_t, while virt_base wants to be a void __iomem *.

drivers/sh/intc.c void __init register_intc_controller(str · line 816

Quoted 10 lines
drivers/sh/intc.c +6 -0
@@ -800,6 +808,12 @@ void __init register_intc_controller(str
 	INIT_LIST_HEAD(&d->list);
 	list_add(&d->list, &intc_list);
 
+	if (desc->io_window) {
+		d->phys_base = desc->io_window->start;
+		d->virt_base = (unsigned long)ioremap_nocache(d->phys_base,
+					       resource_size(desc->io_window));
+	}
+

Error handling?

include/linux/sh_intc.h struct intc_hw_desc { · line 79

Quoted 10 lines
include/linux/sh_intc.h +1 -0
@@ -71,6 +73,7 @@ struct intc_hw_desc {
 
 struct intc_desc {
 	char *name;
+	struct resource *io_window;
 	intc_enum force_enable;
 	struct intc_hw_desc hw;
 };

io_window is pretty much completely lacking in meaning. Why does this platform want to pass in a resource that gets remapped while others do not?

This all looks terribly fragile. You're basically counting on the fact that all addresses are going to be covered by this window, which seems like a pretty lofty assumption given how some of the INTC2 registers were mapped out. It's also pretty ambiguous, given that this is obviously a an MMIO resource yet is named like a PIO one.

Special casing the remapping is really just crap. Each of the register ranges wants to be remapped, while on SH given that these are in the P4 space nothing of much note will happen. You are better off throwing in void __iomem * pointers for each of the register ranges that you are passing in and permit the data structures to use offsets in to that. This can be done incrementally for the SH parts, since remapping isn't going to change the underlying address.

#2 lore raw
In reply to Magnus Damm
Quoted 20 lines · Paul Mundt
Quoted 19 lines · Magnus Damm

From: Magnus Damm <[email protected]>

Add the member "io_window" to intc_desc to allow passing a physical memory window to the INTC code. The INTC code will ioremap the window if present.

Required to support INTCS on SH-Mobile G-series.

---

 drivers/sh/intc.c       |   14 ++++++++++++++
 include/linux/sh_intc.h |    3 +++

 2 files changed, 17 insertions(+)
drivers/sh/intc.c    2010-02-10 19:31:14.000000000 +0900 renamed from drivers/sh/intc.c +0 -0
@@ -45,6 +45,8 @@ struct intc_handle_int {
 
Quoted 10 lines

drivers/sh/intc.c    2010-02-10 19:31:14.000000000 +0900 struct intc_handle_int { · line 52

Quoted 7 lines
drivers/sh/intc.c    2010-02-10 19:31:14.000000000 +0900 renamed from drivers/sh/intc.c +2 -0
  struct intc_desc_int {
       struct list_head list;
+     unsigned long virt_base;
+     unsigned long phys_base;
       struct sys_device sysdev;
       pm_message_t state;
       unsigned long *reg;

phys_base can be a phys_addr_t, while virt_base wants to be a void __iomem *.

Right, I almost converted the code to use void __iomem *.

Quoted 11 lines

drivers/sh/intc.c    2010-02-10 19:31:14.000000000 +0900 void __init register_intc_controller(str · line 816

Quoted 10 lines
drivers/sh/intc.c    2010-02-10 19:31:14.000000000 +0900 renamed from drivers/sh/intc.c +6 -0
@@ -800,6 +808,12 @@ void __init register_intc_controller(str
       INIT_LIST_HEAD(&d->list);
       list_add(&d->list, &intc_list);
 
+     if (desc->io_window) {
+             d->phys_base = desc->io_window->start;
+             d->virt_base = (unsigned long)ioremap_nocache(d->phys_base,
+                                            resource_size(desc->io_window));
+     }
+

Error handling?

There is basically no error handling in the INTC code except some BUG()s. Reworking the code a bit to get better error handling sounds like a good plan.

Quoted 4 lines
Quoted 4 lines
include/linux/sh_intc.h      2010-02-10 19:31:14.000000000 +0900 renamed from include/linux/sh_intc.h +0 -0
@@ -71,6 +73,7 @@ struct intc_hw_desc {
 
Quoted 10 lines

include/linux/sh_intc.h      2010-02-10 19:31:14.000000000 +0900 struct intc_hw_desc { · line 79

Quoted 6 lines
include/linux/sh_intc.h      2010-02-10 19:31:14.000000000 +0900 renamed from include/linux/sh_intc.h +1 -0
  struct intc_desc {
       char *name;
+     struct resource *io_window;
       intc_enum force_enable;
       struct intc_hw_desc hw;
  };

io_window is pretty much completely lacking in meaning. Why does this platform want to pass in a resource that gets remapped while others do not?

Yes. I'm open to naming suggestions, but I don't think that's what you're after. On G-series the INTCA block is covered by the 1:1 mapped virt:phys, so it does not need to be ioremapped. The INTCS block OTOH, it's not covered by any 1:1 mapping and therefore needs ioremap() badly.

Quoted 5 lines

This all looks terribly fragile. You're basically counting on the fact that all addresses are going to be covered by this window, which seems like a pretty lofty assumption given how some of the INTC2 registers were mapped out. It's also pretty ambiguous, given that this is obviously a an MMIO resource yet is named like a PIO one.

In some cases we may have multiple separate interrupt controllers that happen to be covered by a single INTC table. It's of course possible to have multiple ranges, but since it's not directly required for INTCS support I thought I'd go with the simplest possible solution to begin with. If the user is accessing registers outside the window then it's a programming error - very similar to anyone accessing outside ioremapped memory areas. Just don't. I guess you'd like to see some range checking in the cold paths.

Quoted 7 lines

Special casing the remapping is really just crap. Each of the register ranges wants to be remapped, while on SH given that these are in the P4 space nothing of much note will happen. You are better off throwing in void __iomem * pointers for each of the register ranges that you are passing in and permit the data structures to use offsets in to that. This can be done incrementally for the SH parts, since remapping isn't going to change the underlying address.

So how exactly do you propose that we change the INTC data structures? Interrupt controller code from the days before INTC used base addresses and offsets if I'm not mistaken, and that was both hard to read and quite error prone. I'm of course open to improve the INTC code, but I don't see any obvious way to extend the tables as they are today.

/ magnus