Skip to contentSkip to conversation

musb bogosity

PATCH Paul Mundt 2 participants 6 messages Aug 8, 2008

Subsystems (7)
lore
Share & tools
lore

Paul Mundt

author
#0 lore raw

The musb code currently kills the SH randconfig build (and this will hose any platform that also selects HAVE_CLK -- a PPC or AVR32 randconfig would have also hit this eventually), as can be seen here:

http://kisskb.ellerman.id.au/kisskb/buildresult/41095/

This initial failure comes from the fact that musb handily uses "special" I/O routines that it handily wraps, after having failed at grepping for other users. Why this driver isn't using ioread/writeXX_rep() is beyond me, as that's the portable interface we have for doing precisely this sort of thing, without this bizarre PIO/MMIO wrapper munging that isn't even going to work on most platforms.

With that ifdef in place, it's on to the next integral build failure:

drivers/usb/musb/musb_core.c: In function 'fifo_setup':
drivers/usb/musb/musb_core.c:1122: error: 'MUSB_C_RAM_BITS' undeclared (first use in this function) drivers/usb/musb/musb_core.c:1122: error: (Each undeclared identifier is reported only once drivers/usb/musb/musb_core.c:1122: error: for each function it appears in.) drivers/usb/musb/musb_core.c: In function 'ep_config_from_table':
drivers/usb/musb/musb_core.c:1246: error: 'MUSB_C_RAM_BITS' undeclared (first use in this function) drivers/usb/musb/musb_core.c: In function 'musb_remove':
drivers/usb/musb/musb_core.c:2131: warning: unused variable 'ctrl_base'
make[1]: *** [drivers/usb/musb/musb_core.o] Error 1
make: *** [drivers/usb/musb/musb_core.o] Error 2

...

$ git grep MUSB_C_RAM_BITS drivers/usb/musb
drivers/usb/musb/musb_core.c:#define DYN_FIFO_SIZE (1<<(MUSB_C_RAM_BITS+2))
drivers/usb/musb/tusb6010.h:#define MUSB_C_RAM_BITS 12

...

the comment above MUSB_C_RAM_BITS seems to suggest that it's entirely dependent on the chip, so moving it in to musb_core.c wouldn't be terribly productive. Which leaves us in a situation where the core is dependent on arbitrary driver definitions, while the driver itself is obviously only dependent on the core. Has anyone actually tested this with TUSB6010 support disabled?

A quick grep suggests that blackfin is also going to get bitten by this, so simply tossing a depends on (ARM && BROKEN) in wouldn't help matters either.

---

drivers/usb/musb/musb_io.h +1 -1
index 6bbedae..d0f812a 100644
@@ -37,7 +37,7 @@
 
 #include <linux/io.h>
 
-#ifndef	CONFIG_ARM
+#if !defined(CONFIG_ARM) && !defined(CONFIG_SUPERH)
 static inline void readsl(const void __iomem *addr, void *buf, int len)
 	{ insl((unsigned long)addr, buf, len); }
 static inline void readsw(const void __iomem *addr, void *buf, int len)
#1 lore raw
Quoted 50 lines · ext Paul Mundt

The musb code currently kills the SH randconfig build (and this will hose any platform that also selects HAVE_CLK -- a PPC or AVR32 randconfig would have also hit this eventually), as can be seen here:

http://kisskb.ellerman.id.au/kisskb/buildresult/41095/

This initial failure comes from the fact that musb handily uses "special" I/O routines that it handily wraps, after having failed at grepping for other users. Why this driver isn't using ioread/writeXX_rep() is beyond me, as that's the portable interface we have for doing precisely this sort of thing, without this bizarre PIO/MMIO wrapper munging that isn't even going to work on most platforms.

With that ifdef in place, it's on to the next integral build failure:

drivers/usb/musb/musb_core.c: In function 'fifo_setup':
drivers/usb/musb/musb_core.c:1122: error: 'MUSB_C_RAM_BITS' undeclared (first use in this function) drivers/usb/musb/musb_core.c:1122: error: (Each undeclared identifier is reported only once drivers/usb/musb/musb_core.c:1122: error: for each function it appears in.) drivers/usb/musb/musb_core.c: In function 'ep_config_from_table':
drivers/usb/musb/musb_core.c:1246: error: 'MUSB_C_RAM_BITS' undeclared (first use in this function) drivers/usb/musb/musb_core.c: In function 'musb_remove':
drivers/usb/musb/musb_core.c:2131: warning: unused variable 'ctrl_base'
make[1]: *** [drivers/usb/musb/musb_core.o] Error 1
make: *** [drivers/usb/musb/musb_core.o] Error 2

...

$ git grep MUSB_C_RAM_BITS drivers/usb/musb
drivers/usb/musb/musb_core.c:#define DYN_FIFO_SIZE (1<<(MUSB_C_RAM_BITS+2))
drivers/usb/musb/tusb6010.h:#define MUSB_C_RAM_BITS 12

...

the comment above MUSB_C_RAM_BITS seems to suggest that it's entirely dependent on the chip, so moving it in to musb_core.c wouldn't be terribly productive. Which leaves us in a situation where the core is dependent on arbitrary driver definitions, while the driver itself is obviously only dependent on the core. Has anyone actually tested this with TUSB6010 support disabled?

A quick grep suggests that blackfin is also going to get bitten by this, so simply tossing a depends on (ARM && BROKEN) in wouldn't help matters either.

---

drivers/usb/musb/musb_io.h +0 -0
index 6bbedae..d0f812a 100644

drivers/usb/musb/musb_io.h line 43

Quoted 8 lines
drivers/usb/musb/musb_io.h +1 -1
@@ -37,7 +37,7 @@
 
 #include <linux/io.h>
 
-#ifndef	CONFIG_ARM
+#if !defined(CONFIG_ARM) && !defined(CONFIG_SUPERH)
 static inline void readsl(const void __iomem *addr, void *buf, int len)
 	{ insl((unsigned long)addr, buf, len); }
 static inline void readsw(const void __iomem *addr, void *buf, int len)

The right way to fix this would be by removing all those defines and make that configuration come from platform_data. I have a patch for that, I'll send it to Greg and try to be sure it gets applied. I've been using it quite a while and it seems stable.

Signature · 1 line
balbi
author
#2 lore raw
Quoted 8 lines · Felipe Balbi
Quoted 7 lines · ext Paul Mundt

This initial failure comes from the fact that musb handily uses "special" I/O routines that it handily wraps, after having failed at grepping for other users. Why this driver isn't using ioread/writeXX_rep() is beyond me, as that's the portable interface we have for doing precisely this sort of thing, without this bizarre PIO/MMIO wrapper munging that isn't even going to work on most platforms.

[snip]

Quoted 13 lines
Quoted 13 lines

A quick grep suggests that blackfin is also going to get bitten by this, so simply tossing a depends on (ARM && BROKEN) in wouldn't help matters either.

---

drivers/usb/musb/musb_io.h +0 -0
index 6bbedae..d0f812a 100644
@@ -37,7 +37,7 @@
 
 #include <linux/io.h>
 
Quoted 11 lines

drivers/usb/musb/musb_io.h line 43

Quoted 5 lines
drivers/usb/musb/musb_io.h +1 -1
-#ifndef	CONFIG_ARM
+#if !defined(CONFIG_ARM) && !defined(CONFIG_SUPERH)
 static inline void readsl(const void __iomem *addr, void *buf, int len)
 	{ insl((unsigned long)addr, buf, len); }
 static inline void readsw(const void __iomem *addr, void *buf, int len)

The right way to fix this would be by removing all those defines and make that configuration come from platform_data. I have a patch for that, I'll send it to Greg and try to be sure it gets applied. I've been using it quite a while and it seems stable.

Great. The fact the I/O routines are broken and that this doesn't build on anything but ARM still remains a problem. Your generic ifdefs also seem to be using __raw_xxx() as MMIO accessors while using the xxx() variants as PIO accessors, irregardless of whether the platform sets NO_IOPORT or not. If you aren't going to use generic routines, then don't bother removing your architecture dependency in Kconfig. As it is, your driver's assumptions on I/O accesses are simply bogus, which is precisely why we have these portable interfaces in the first place.

#3 lore raw
In reply to Felipe Balbi
Quoted 1 line · Felipe Balbi
Quoted 1 line

other users. Why this driver isn't using ioread/writeXX_rep() is beyond

I'll try to fix io routines asap. From your next mail I could see that it's really bogus :-p

Quoted 10 lines
Quoted 10 lines

drivers/usb/musb/musb_core.c: In function 'fifo_setup':
drivers/usb/musb/musb_core.c:1122: error: 'MUSB_C_RAM_BITS' undeclared (first use in this function) drivers/usb/musb/musb_core.c:1122: error: (Each undeclared identifier is reported only once drivers/usb/musb/musb_core.c:1122: error: for each function it appears in.) drivers/usb/musb/musb_core.c: In function 'ep_config_from_table':
drivers/usb/musb/musb_core.c:1246: error: 'MUSB_C_RAM_BITS' undeclared (first use in this function) drivers/usb/musb/musb_core.c: In function 'musb_remove':
drivers/usb/musb/musb_core.c:2131: warning: unused variable 'ctrl_base'
make[1]: *** [drivers/usb/musb/musb_core.o] Error 1
make: *** [drivers/usb/musb/musb_core.o] Error 2

Below is the patch that fixes it.

I don't have sh cross-compilers, so if you could try it out and be sure it works, I'd be glad.

==== cut here ====

Quoted 1 line

From 0bbc416e8170c4169608f7496b8d67be3828d76b Mon Sep 17 00:00:00 2001

From: Felipe Balbi <[email protected]>
Date: Fri, 8 Aug 2008 11:39:41 +0300
Subject: [PATCH] usb: musb: pass configuration specifics via pdata

Use platform_data to pass musb configuration-specific details to musb driver.

This patch will prevent that other platforms selecting HAVE_CLK and enabling musb won't break tree building.

The other parts of it will come when linux-omap merge up more omap2/3 board-files.

---
 arch/arm/mach-omap2/usb-tusb6010.c |    1 -
 drivers/usb/musb/musb_core.c       |   37 ++++-----
 drivers/usb/musb/musb_core.h       |   14 +---
 drivers/usb/musb/tusb6010.h        |  169 ------------------------------------
 include/linux/usb/musb.h           |   38 +++++++-
 5 files changed, 51 insertions(+), 208 deletions(-)
arch/arm/mach-omap2/usb-tusb6010.c +0 -1
index 80bb42e..8fee64c 100644
@@ -317,7 +317,6 @@ tusb6010_setup_interface(struct musb_hdrc_platform_data *data,
 		printk(error, 6, status);
 		return -ENODEV;
 	}
-	data->multipoint = 1;
 	tusb_device.dev.platform_data = data;
 
 	/* REVISIT let the driver know what DMA channels work */
drivers/usb/musb/musb_core.c +16 -21
index 462586d..d68ec6d 100644
@@ -990,12 +990,6 @@ static void musb_shutdown(struct platform_device *pdev)
  * We don't currently use dynamic fifo setup capability to do anything
  * more than selecting one of a bunch of predefined configurations.
  */
-#ifdef MUSB_C_DYNFIFO_DEF
-#define	can_dynfifo()	1
-#else
-#define	can_dynfifo()	0
-#endif
-
 #if defined(CONFIG_USB_TUSB6010) || \
 	defined(CONFIG_ARCH_OMAP2430) || defined(CONFIG_ARCH_OMAP34XX)
 static ushort __initdata fifo_mode = 4;
@@ -1008,8 +1002,6 @@ module_param(fifo_mode, ushort, 0);
 MODULE_PARM_DESC(fifo_mode, "initial endpoint configuration");
 
 
-#define DYN_FIFO_SIZE (1<<(MUSB_C_RAM_BITS+2))
-
 enum fifo_style { FIFO_RXTX, FIFO_TX, FIFO_RX } __attribute__ ((packed));
 enum buf_mode { BUF_SINGLE, BUF_DOUBLE } __attribute__ ((packed));
 
@@ -1119,11 +1111,12 @@ fifo_setup(struct musb *musb, struct musb_hw_ep  *hw_ep,
 
 	c_size = size - 3;
 	if (cfg->mode == BUF_DOUBLE) {
-		if ((offset + (maxpacket << 1)) > DYN_FIFO_SIZE)
+		if ((offset + (maxpacket << 1)) >
+				(1 << (musb->config->ram_bits + 2)))
 			return -EMSGSIZE;
 		c_size |= MUSB_FIFOSZ_DPB;
 	} else {
-		if ((offset + maxpacket) > DYN_FIFO_SIZE)
+		if ((offset + maxpacket) > (1 << (musb->config->ram_bits + 2)))
 			return -EMSGSIZE;
 	}
 
@@ -1219,13 +1212,13 @@ static int __init ep_config_from_table(struct musb *musb)
 	/* assert(offset > 0) */
 
 	/* NOTE:  for RTL versions >= 1.400 EPINFO and RAMINFO would
-	 * be better than static MUSB_C_NUM_EPS and DYN_FIFO_SIZE...
+	 * be better than static musb->config->num_eps and DYN_FIFO_SIZE...
 	 */
 
 	for (i = 0; i < n; i++) {
 		u8	epn = cfg->hw_ep_num;
 
-		if (epn >= MUSB_C_NUM_EPS) {
+		if (epn >= musb->config->num_eps) {
 			pr_debug("%s: invalid ep %d\n",
 					musb_driver_name, epn);
 			continue;
@@ -1242,8 +1235,8 @@ static int __init ep_config_from_table(struct musb *musb)
 
 	printk(KERN_DEBUG "%s: %d/%d max ep, %d/%d memory\n",
 			musb_driver_name,
-			n + 1, MUSB_C_NUM_EPS * 2 - 1,
-			offset, DYN_FIFO_SIZE);
+			n + 1, musb->config->num_eps * 2 - 1,
+			offset, (1 << (musb->config->ram_bits + 2)));
 
 #ifdef CONFIG_USB_MUSB_HDRC_HCD
 	if (!musb->bulk_ep) {
@@ -1270,7 +1263,7 @@ static int __init ep_config_from_hw(struct musb *musb)
 
 	/* FIXME pick up ep0 maxpacket size */
 
-	for (epnum = 1; epnum < MUSB_C_NUM_EPS; epnum++) {
+	for (epnum = 1; epnum < musb->config->num_eps; epnum++) {
 		musb_ep_select(mbase, epnum);
 		hw_ep = musb->endpoints + epnum;
 
@@ -1424,14 +1417,14 @@ static int __init musb_core_init(u16 musb_type, struct musb *musb)
 	musb->epmask = 1;
 
 	if (reg & MUSB_CONFIGDATA_DYNFIFO) {
-		if (can_dynfifo())
+		if (musb->config->dyn_fifo)
 			status = ep_config_from_table(musb);
 		else {
 			ERR("reconfigure software for Dynamic FIFOs\n");
 			status = -ENODEV;
 		}
 	} else {
-		if (!can_dynfifo())
+		if (!musb->config->dyn_fifo)
 			status = ep_config_from_hw(musb);
 		else {
 			ERR("reconfigure software for static FIFOs\n");
@@ -1788,7 +1781,8 @@ static void musb_irq_work(struct work_struct *data)
  */
 
 static struct musb *__init
-allocate_instance(struct device *dev, void __iomem *mbase)
+allocate_instance(struct device *dev,
+		struct musb_hdrc_config *config, void __iomem *mbase)
 {
 	struct musb		*musb;
 	struct musb_hw_ep	*ep;
@@ -1820,8 +1814,9 @@ allocate_instance(struct device *dev, void __iomem *mbase)
 	musb->mregs = mbase;
 	musb->ctrl_base = mbase;
 	musb->nIrq = -ENODEV;
+	musb->config = config;
 	for (epnum = 0, ep = musb->endpoints;
-			epnum < MUSB_C_NUM_EPS;
+			epnum < musb->config->num_eps;
 			epnum++, ep++) {
 
 		ep->musb = musb;
@@ -1929,7 +1924,7 @@ bad_config:
 	}
 
 	/* allocate */
-	musb = allocate_instance(dev, ctrl);
+	musb = allocate_instance(dev, plat->config, ctrl);
 	if (!musb)
 		return -ENOMEM;
 
@@ -1987,7 +1982,7 @@ bad_config:
 	musb_generic_disable(musb);
 
 	/* setup musb parts of the core (especially endpoints) */
-	status = musb_core_init(plat->multipoint
+	status = musb_core_init(plat->config->multipoint
 			? MUSB_CONTROLLER_MHDRC
 			: MUSB_CONTROLLER_HDRC, musb);
 	if (status < 0)
drivers/usb/musb/musb_core.h +2 -12
index 90035c1..eade46d 100644
@@ -56,18 +56,6 @@ struct musb_ep;
 #include "musb_debug.h"
 #include "musb_dma.h"
 
-#ifdef CONFIG_USB_MUSB_SOC
-/*
- * Get core configuration from a header converted (by cfg_conv)
- * from the Verilog config file generated by the core config utility
- *
- * For now we assume that header is provided along with other
- * arch-specific files.  Discrete chips will need a build tweak.
- * So will using AHB IDs from silicon that provides them.
- */
-#include <asm/arch/hdrc_cnf.h>
-#endif
-
 #include "musb_io.h"
 #include "musb_regs.h"
 
@@ -440,6 +428,8 @@ struct musb {
 	struct usb_gadget_driver *gadget_driver;	/* its driver */
 #endif
 
+	struct musb_hdrc_config	*config;
+
 #ifdef MUSB_CONFIG_PROC_FS
 	struct proc_dir_entry *proc_entry;
 #endif
drivers/usb/musb/tusb6010.h +0 -169

173 lines not on this page · raw message

include/linux/usb/musb.h +33 -5
index d325a0d..630962c 100644
@@ -19,6 +19,36 @@ enum musb_mode {
 
 struct clk;
 
+struct musb_hdrc_eps_bits {
+	const char	name[16];
+	u8		bits;
+};
+
+struct musb_hdrc_config {
+	/* MUSB configuration-specific details */
+	unsigned	multipoint:1;	/* multipoint device */
+	unsigned	dyn_fifo:1;	/* supports dynamic fifo sizing */
+	unsigned	soft_con:1;	/* soft connect required */
+	unsigned	utm_16:1;	/* utm data witdh is 16 bits */
+	unsigned	big_endian:1;	/* true if CPU uses big-endian */
+	unsigned	mult_bulk_tx:1;	/* Tx ep required for multbulk pkts */
+	unsigned	mult_bulk_rx:1;	/* Rx ep required for multbulk pkts */
+	unsigned	high_iso_tx:1;	/* Tx ep required for HB iso */
+	unsigned	high_iso_rx:1;	/* Rx ep required for HD iso */
+	unsigned	dma:1;		/* supports DMA */
+	unsigned	vendor_req:1;	/* vendor registers required */
+
+	u8		num_eps;	/* number of endpoints _with_ ep0 */
+	u8		dma_channels;	/* number of dma channels */
+	u8		dyn_fifo_size;	/* dynamic size in bytes */
+	u8		vendor_ctrl;	/* vendor control reg width */
+	u8		vendor_stat;	/* vendor status reg witdh */
+	u8		dma_req_chan;	/* bitmask for required dma channels */
+	u8		ram_bits;	/* ram address size */
+
+	struct musb_hdrc_eps_bits *eps_bits;
+};
+
 struct musb_hdrc_platform_data {
 	/* MUSB_HOST, MUSB_PERIPHERAL, or MUSB_OTG */
 	u8		mode;
@@ -38,16 +68,14 @@ struct musb_hdrc_platform_data {
 	/* (HOST or OTG) msec/2 after VBUS on till power good */
 	u8		potpgt;
 
-	/* TBD:  chip defaults should probably go someplace else,
-	 * e.g. number of tx/rx endpoints, etc
-	 */
-	unsigned	multipoint:1;
-
 	/* Power the device on or off */
 	int		(*set_power)(int state);
 
 	/* Turn device clock on or off */
 	int		(*set_clock)(struct clk *clock, int is_on);
+
+	/* MUSB configuration-specific details */
+	struct musb_hdrc_config	*config;
 };
 
 
Signature · 4 lines
1.6.0.rc1.71.gfba5

-- 
balbi
author Acked-by
#4 lore raw
Quoted 17 lines · Felipe Balbi
Quoted 1 line · Felipe Balbi
Quoted 1 line

other users. Why this driver isn't using ioread/writeXX_rep() is beyond

I'll try to fix io routines asap. From your next mail I could see that it's really bogus :-p

Quoted 10 lines
Quoted 10 lines

drivers/usb/musb/musb_core.c: In function 'fifo_setup':
drivers/usb/musb/musb_core.c:1122: error: 'MUSB_C_RAM_BITS' undeclared (first use in this function) drivers/usb/musb/musb_core.c:1122: error: (Each undeclared identifier is reported only once drivers/usb/musb/musb_core.c:1122: error: for each function it appears in.) drivers/usb/musb/musb_core.c: In function 'ep_config_from_table':
drivers/usb/musb/musb_core.c:1246: error: 'MUSB_C_RAM_BITS' undeclared (first use in this function) drivers/usb/musb/musb_core.c: In function 'musb_remove':
drivers/usb/musb/musb_core.c:2131: warning: unused variable 'ctrl_base'
make[1]: *** [drivers/usb/musb/musb_core.o] Error 1
make: *** [drivers/usb/musb/musb_core.o] Error 2

Quoted 5 lines

Below is the patch that fixes it.

I don't have sh cross-compilers, so if you could try it out and be sure it works, I'd be glad.

Yes, that fixes that problem at least, thanks.

Now the I/O bits are the only outstanding issue, as I still need the CONFIG_SUPERH test in musb_io.h to get around conflicting definitions.

Presently ARM and SH are the only ones that define the readsl() variants anyways, so stubbing in an extra ifdef there is fine as a temporary measure, but it's not a long-term solution. The only reason SH defines them at all is because we ran in to this exact same problem the last time an ARM driver tried passing itself off as generic ;-)

Quoted 10 lines

==== cut here ====

From 0bbc416e8170c4169608f7496b8d67be3828d76b Mon Sep 17 00:00:00 2001
From: Felipe Balbi <[email protected]>
Date: Fri, 8 Aug 2008 11:39:41 +0300
Subject: [PATCH] usb: musb: pass configuration specifics via pdata

Use platform_data to pass musb configuration-specific details to musb driver.

Quoted 7 lines

This patch will prevent that other platforms selecting HAVE_CLK and enabling musb won't break tree building.

The other parts of it will come when linux-omap merge up more omap2/3 board-files.

#5 lore raw
Quoted 1 line · ext Paul Mundt

Yes, that fixes that problem at least, thanks.

Cool :-)

Quoted 2 lines

Now the I/O bits are the only outstanding issue, as I still need the CONFIG_SUPERH test in musb_io.h to get around conflicting definitions.

I think we're gonna need that for a while, can you send a proper patch changing that ifndef ??

Feel free to put my Acked-by: Felipe Balbi <[email protected]> if you wish.

Quoted 5 lines

Presently ARM and SH are the only ones that define the readsl() variants anyways, so stubbing in an extra ifdef there is fine as a temporary measure, but it's not a long-term solution. The only reason SH defines them at all is because we ran in to this exact same problem the last time an ARM driver tried passing itself off as generic ;-)

eheh... But this actually is a generic driver, we're only missing a pci bus glue to use the musb chip on the opt cards, that would be really nice :-)

Signature · 1 line
balbi