Reusable UIO Platform Driver

These patches implement a reusable UIO platform driver. This driver can be used to export hardware to user space, as long as the device is a) memory mapped and b) equipped with an unique IRQ.

The uio_platform driver gets all information through platform data, including address window information and IRQ number. The driver also supports assigning a contiguous piece of memory to each instance. This may be useful when the exported hardware blocks can bus master but requires physically contiguous memory.

There are not many surprises in the code if you are familiar with UIO, except for the interrupt handling. All UIO kernel drivers that I've seen so far have hardware specific interrupt acknowledge code in the interrupt handler. The uio_platform driver is different.

The interrupt handling code in uio_platform assumes the device is the only user of the assigned interrupt. This may be rare in the PC world but for SuperH almost all interrupt sources are unique. Having an unique interrupt for the device allows the code to use disable_irq() and enable_irq() in kernel space and leave actual interrupt acknowledge to user space. That way we have no hardware specific code in the kernel.

Interrupts are of course serviced in kernel space by the uio_platform driver, but the uio_platform interrupt handler will simply disable the IRQ until next read() or poll() syscall. The uio_platform kernel driver assumes that the user space driver has taken care of acknowledging the interrupt before doing read() and waiting for events again. If no acknowledge has happened then an interrupt will occur again (since it's still pending) and the kernel interrupt handler will disable the IRQ again and unblock the user space process.

The last patch contains SuperH specific code that exports various multimedia acceleration blocks to userspace. The following processors and hardware blocks are exported for now:

sh7343: VPU
sh7722: VPU, VEU
sh7723: VPU, VEU, VEU

If anyone is interested then I have a proof of concept vidix driver for mplayer that is interfacing using UIO to the VEU on a sh7722 to provide accelerated color space conversion and stretching.

[PATCH 01/03] uio: Add enable_irq() callback [PATCH 02/03] uio: Add uio_platform driver
[PATCH 03/03] sh: Export sh7343/sh7722/sh7723 VPU/VEU blocks

---

 arch/sh/kernel/cpu/sh4a/setup-sh7343.c |   32 ++++++
 arch/sh/kernel/cpu/sh4a/setup-sh7722.c |   63 ++++++++++++
 arch/sh/kernel/cpu/sh4a/setup-sh7723.c |   94 ++++++++++++++++++
 drivers/uio/Kconfig                    |   10 +
 drivers/uio/Makefile                   |    1 
 drivers/uio/uio.c                      |    6 +
 drivers/uio/uio_platform.c             |  161 ++++++++++++++++++++++++++++++++
 include/linux/uio_driver.h             |    1 
 include/linux/uio_platform.h           |   10 +
 9 files changed, 378 insertions(+)

Add enable_irq() callback to struct uio_info. This callback is needed by the uio_platform driver so interrupts can be enabled before blocking.

---

 drivers/uio/uio.c          |    6 ++++++
 include/linux/uio_driver.h |    1 +
 2 files changed, 7 insertions(+)

@@ -365,6 +365,9 @@ static unsigned int uio_poll(struct file
 	if (idev->info->irq = UIO_IRQ_NONE)
 		return -EIO;
 
+	if (idev->info->enable_irq)
+		idev->info->enable_irq(idev->info);
+
 	poll_wait(filep, &idev->wait, wait);
 	if (listener->event_count != atomic_read(&idev->event))
 		return POLLIN | POLLRDNORM;

@@ -391,6 +394,9 @@ static ssize_t uio_read(struct file *fil
 	do {
 		set_current_state(TASK_INTERRUPTIBLE);
 
+		if (idev->info->enable_irq)
+			idev->info->enable_irq(idev->info);
+
 		event_count = atomic_read(&idev->event);
 		if (event_count != listener->event_count) {
 			if (copy_to_user(buf, &event_count, count))

@@ -64,6 +64,7 @@ struct uio_info {
 	void			*priv;
 	irqreturn_t (*handler)(int irq, struct uio_info *dev_info);
 	int (*mmap)(struct uio_info *info, struct vm_area_struct *vma);
+	void (*enable_irq)(struct uio_info *info);
 	int (*open)(struct uio_info *info, struct inode *inode);
 	int (*release)(struct uio_info *info, struct inode *inode);
 };

We can most likely use a single uio platform driver if this patch is acceptable. Any NAKs?

Thanks,

/ magnus

Yes. Your approach only allows enabling interrupts, but not disabling them. And I don't like that it is not possible for generic userspace tools to find out if a UIO device has this auto-irq-enabling capability or not.

I just posted a patch that allows enabling _and_ disabling of irqs from userspace by writing 0 or 1 to /dev/uioX. I've CCed you, could you please test? If this doesn't do what you need, please let me know.

Thanks,
Hans

Hi Hans,

Thanks for your effort. I understand that you need to enable and disable interrupts from user space, but that's a bit different from what I want to do. I just want interrupts to be enabled before I do read() or poll().

Also, adding the capability of disabling and enabling interrupts from user space seems a bit error prone to me. Mainly since user space then needs to know that the interrupt handler in kernel space can cope with such changes. Not such a clean interface IMO. OTOH you may need that to cope with some broken hardware.

But why does user space need to know if the auto-irq-enabling function is there or not? If the user is interfacing to the wrong kernel UIO driver with wrong behavior then he has obviously done something wrong. Knowing if auto-irq-enabling is there from user space isn't going to save users from themselves. They can and will mix and match things in all sorts of wrong ways anyway.

I'm sure your patch or the ioctl suggestion both allow re-enabling interrupts from user space. That's great, but both of them add extra syscall overhead compared to my suggestion. They also make the user space interface in user space part of the driver a bit more complicated.

I do understand that you don't want to mess up your UIO kernel callbacks by introducing just merging new ones all the time. OTOH, my patch is just a few lines. Is introducing one extra syscall good enough performance wise?

Thanks for your help!

/ magnus

Am Fri, 23 May 2008 10:24:42 +0900
schrieb "Magnus Damm" <[email protected]>:

Are you sure this works cleanly? You usually do a read immediately after poll, so you might get two interrupts when waiting with select().

No, it doesn't. If the kernel driver doesn't implement the irqcontrol() handler, write attempts will return an error.

All of this talk is _only_ about broken hardware. Decent hardware has seperate IRQ mask and status registers, in which case userspace has no problems at all to deal with several internal interrupt sources of the chip. We only need all this for chips where it is not possible to mask an IRQ through a mappable register or (more often) where acknowledging the interrupt also clears the status register so that userspace has no way of knowing what the source of the interrupt was. The latter applies only to chips with more than one internal irq source.

It doesn't, I just find it nice to have such a flag displayed in lsuio.

Agreed.

UIO deals with two things, device memory and interrupts. We have mmap() for mem and read() for waiting for an irq. This looks clean and logical to me:

1) mmap() => device memory
2) read()/write() => device irqs

This doesn't seem to be a problem, really. This write() is straight forward, without any wait queues etc, so what?

Thanks,
Hans

This patch implements a reusable UIO platform driver. Only memory mapped hardware devices with unique IRQs are supported.

---

Tested using the VEU on a SuperH sh7722.

 drivers/uio/Kconfig          |   10 ++
 drivers/uio/Makefile         |    1 
 drivers/uio/uio_platform.c   |  161 ++++++++++++++++++++++++++++++++++++++++++
 include/linux/uio_platform.h |   10 ++
 4 files changed, 182 insertions(+)

@@ -39,4 +39,14 @@ config UIO_SMX
 
 	  If you compile this as a module, it will be called uio_smx.
 
+config UIO_PLATFORM
+	tristate "Userspace I/O Platform driver"
+	depends on UIO
+	default n
+	help
+	  Reusable userspace IO interface for memory mapped devices
+	  equipped with an unique IRQ. IRQ sharing is not supported.
+
+	  To compile this driver as a module, choose M here: the module
+	  will be called uio_platform.
 endif

@@ -1,3 +1,4 @@
 obj-$(CONFIG_UIO)	+= uio.o
 obj-$(CONFIG_UIO_CIF)	+= uio_cif.o
 obj-$(CONFIG_UIO_SMX)	+= uio_smx.o
+obj-$(CONFIG_UIO_PLATFORM)	+= uio_platform.o

@@ -0,0 +1,10 @@
+#ifndef __LINUX_UIO_PLATFORM_H__
+#define __LINUX_UIO_PLATFORM_H__
+
+struct uio_platform_info {
+	char *name;
+	char *version;
+	unsigned long memsize;
+};
+
+#endif /* __LINUX_UIO_PLATFORM_H__ */

This patch exports the following SuperH hardware to user space:

sh7343: VPU
sh7722: VPU, VEU
sh7723: VPU, VEU, VEU

---

 arch/sh/kernel/cpu/sh4a/setup-sh7343.c |   32 ++++++++++
 arch/sh/kernel/cpu/sh4a/setup-sh7722.c |   63 +++++++++++++++++++++
 arch/sh/kernel/cpu/sh4a/setup-sh7723.c |   94 ++++++++++++++++++++++++++++++++
 3 files changed, 189 insertions(+)

@@ -11,6 +11,37 @@
 #include <linux/init.h>
 #include <linux/serial.h>
 #include <linux/serial_sci.h>
+#include <linux/uio_platform.h>
+
+static struct uio_platform_info vpu_platform_data = {
+	.name = "VPU",
+	.version = "0.0.1",
+	.memsize = 1 << 20,
+};
+
+static struct resource vpu_resources[] = {
+	[0] = {
+		.name	= "VPU",
+		.start	= 0xfe900000,
+		.end	= 0xfe9022ec,
+		.flags	= IORESOURCE_MEM,
+	},
+	[1] = {
+		.start	= 60,
+		.end	= 60,
+		.flags	= IORESOURCE_IRQ,
+	},
+};
+
+static struct platform_device vpu_device = {
+	.name		= "uio-platform",
+	.id		= 0,
+	.dev = {
+		.platform_data	= &vpu_platform_data,
+	},
+	.resource	= vpu_resources,
+	.num_resources	= ARRAY_SIZE(vpu_resources),
+};
 
 static struct plat_sci_port sci_platform_data[] = {
 	{

@@ -33,6 +64,7 @@ static struct platform_device sci_device
 
 static struct platform_device *sh7343_devices[] __initdata = {
 	&sci_device,
+	&vpu_device,
 };
 
 static int __init sh7343_devices_setup(void)

On Tue, May 20, 2008 at 07:51:32PM +0900, Magnus Damm wrote:

Hi Magnus,

Uwe Kleine-Koenig already submitted such a framework:

http://lkml.org/lkml/2008/5/20/94

It's his third version, and it looks good. I presume you didn't know about his work. The main difference is that he leaves interrupt handling to platform code. That might look strange (it did to me first), but it has the advantage that you can put hardware dependent stuff in your board support (which depends on hardware anyway).

Could you have a look at his patch and tell me if that does what you need?

Uwe's approach doesn't have this limitation.

This sounds quite interesting. Unfortunately, I'm not familiar with the SuperH architecture. Could you also do this with Uwe's approach?

I'm about to sign-off Uwe's patch, and we'll possibly have that in mainline soon. I don't mind having a second "generic platform" driver, but you'll need to have good technical arguments.

Thanks,
Hans

Hi Hans!

I was not aware of this driver, thanks for the pointer!

The uio_pdrv driver doesn't do what I need at this point, though I may be able to extend it with the following:
- Interrupt enable/disable code
- Physically contiguous memory support

The interrupt code may be placed in the board/cpu code, but I need to share that code between multiple UIO driver instances. We want to use the same UIO driver for many different processor models and hardware blocks. Extending uio_pdrv driver with a chunk of physically contiguous memory isn't a big deal though.

To be frank, I have my doubts in adding an extra forwarding-only platform layer on top of UIO compared to using uio_register_device() directly from the board code. I like that the platform layer is using struct resource and handles resource ranges for us automatically, but wouldn't it make more sense to extend the UIO core to always use struct resource instead of struct uio_mem? I'd be happy to help out - just point me in the right direction.

True, but the uio_pdrv driver is choosing to not deal with interrupts at all. I'd like to have shared interrupt handling code. With my driver, you just feed it io memory window parameters and an interrupt number and off you go. No need for any callbacks.

I could handle things by extending Uwe's uio_pdrv driver, but I still need the enable_irq() callback. I wonder if it makes sense to let the two drivers coexist side by side, since they are solving different problems. I can rename my driver to uio_pdrv_unique_irq or something, or maybe uio_superh.c. I dislike the latter since my driver doesn't do anything SuperH specific and that I suspect it can be useful for other architectures as well.

I'd like to have this driver upstream as well, and sharing the code with the uio_pdrv driver is one way, but I suspect that adding another reusable layer on top of that driver will just complicate things. The uio-specific part of my patches is less than 200 lines of code. From the top of my head I can think of at least 10 different SuperH hardware devices that can reuse this driver.

Please let me know what you prefer and I'll update the code and repost.

Thanks for your help!

/ magnus

Hello Magnus,

What about adding uio_platform_handler (with a different name) to uio_pdrv.c?
OTOH I don't see why you want to disable the irq. Can you describe the reason?

I wonder how you use that memory. Isn't it just some kind of shared memory? If so, why not use normal shared memory? Do you really need that?

That alone doesn't help. You need a struct device to register a uio device. So a platform device is the straight forward approach.

In my eyes this isn't completly correct. Just the way you specify your handler is a bit different. You can pass a handler via platform data with my driver, too.

BTW, you don't need "depends on UIO" (because it's in a if UIO/endif block) and "default n" (as this is the default anyhow). See also
http://thread.gmane.org/gmane.linux.kernel/663884/focush3097

Best regards
Uwe

Uwe Kleine-König, Software Engineer
Digi International GmbH Branch Breisach, Küferstrasse 8, 79206 Breisach, Germany
Tax: 315/5781/0242 / VAT: DE153662976 / Reg. Amtsgericht Dortmund HRB 13962

Physically contiguous memory is a real requirement, especially for DMA. I'm not sure what's confusing about that?

I got that, yes. The problem is I don't see how you can use it for DMA. The physical address is stored in info->mem[$last].internal_addr and if there is a way to access that variable from user space, I don't see it and would appretiate a hint. Sorry for not expressing my concern more clear at the first go. I hope it's understandable now.

@Magnus: Maybe you can provide the userspace part of the driver? How is that mapping used there?

Best regards
Uwe

Uwe Kleine-König, Software Engineer
Digi International GmbH Branch Breisach, Küferstrasse 8, 79206 Breisach, Germany
Tax: 315/5781/0242 / VAT: DE153662976 / Reg. Amtsgericht Dortmund HRB 13962

[Added Matsubara-san as CC]

Sure, here is a little test program. Have a look at "uio_mem". The "address" member contains the physical address that can be used for bus mastering DMA. Compare that to "iomem" which is the pointer to the virtual memory area in user space.

Hope this helps!

/ magnus

/*
* sh7722 VEU scaling example using UIO - 20080521 Magnus Damm *
* Derived from
* "Simple tick timer through Sky TMU UIO driver" by Katsuya Matsubara */
#include <stdio.h>
#include <unistd.h>
#include <string.h>
#include <stdlib.h>
#include <sys/mman.h>
#include <sys/time.h>
#include <sys/ioctl.h>
#include <fcntl.h>
#include <linux/fb.h>

static int fgets_with_openclose(char *fname, char *buf, size_t maxlen) {
	FILE *fp;

	if ((fp = fopen(fname, "r")) != NULL) {
		fgets(buf, maxlen, fp);
		fclose(fp);
		return strlen(buf);
	} else {
		return -1;
	}
}

struct uio_device {
char *name;
char *path;
int fd;
};

#define MAXUIOIDS 100
#define MAXNAMELEN 256

static int locate_uio_device(char *name, struct uio_device *udp)

{
	char fname[MAXNAMELEN], buf[MAXNAMELEN];
	int uio_id;

	for (uio_id = 0; uio_id < MAXUIOIDS; uio_id++) {
		sprintf(fname, "/sys/class/uio/uio%d/name", uio_id);
		if (fgets_with_openclose(fname, buf, MAXNAMELEN) < 0)
			continue;
		if (strncmp(name, buf, strlen(name)) == 0)
			break;
	}

	if (uio_id >= MAXUIOIDS)
		return -1;

	udp->name = strdup(buf);
	udp->path = strdup(fname);
	udp->path[strlen(udp->path) - 4] = '\0';

	sprintf(buf, "/dev/uio%d", uio_id);
	udp->fd = open(buf, O_RDWR|O_SYNC);

	if (udp->fd < 0) {
		perror("open");
		return -1;
	}

	return 0;
}

struct uio_map {
unsigned long address;
unsigned long size;
void *iomem;
};

static int setup_uio_map(struct uio_device *udp, int nr, struct uio_map *ump)

{
	char fname[MAXNAMELEN], buf[MAXNAMELEN];
 
	sprintf(fname, "%s/maps/map%d/addr", udp->path, nr);
	if (fgets_with_openclose(fname, buf, MAXNAMELEN) <= 0)
		return -1;

	ump->address = strtoul(buf, NULL, 0);

	sprintf(fname, "%s/maps/map%d/size", udp->path, nr);
	if (fgets_with_openclose(fname, buf, MAXNAMELEN) <= 0)
		return -1;

	ump->size = strtoul(buf, NULL, 0);

	ump->iomem = mmap(0, ump->size,
			  PROT_READ|PROT_WRITE, MAP_SHARED,
			  udp->fd, nr * getpagesize());

	if (ump->iomem == MAP_FAILED)
		return -1;

	return 0;
}

struct fb_info {
unsigned long width;
unsigned long height;
unsigned long bpp;
unsigned long line_length;

unsigned long address;
unsigned long size;
};

static int get_fb_info(char *device, struct fb_info *fip)

{
	struct fb_var_screeninfo vinfo;
	struct fb_fix_screeninfo finfo;
	int fd;

	fd = open(device, O_RDWR);
	if (fd < 0) {
		perror("open");
		return -1;
	}

	if (ioctl(fd, FBIOGET_VSCREENINFO, &vinfo) == -1) {
		perror("ioctl(FBIOGET_VSCREENINFO)");
		return -1;
	}

	fip->width = vinfo.xres;
	fip->height = vinfo.yres;
	fip->bpp = vinfo.bits_per_pixel;

	if (ioctl(fd, FBIOGET_FSCREENINFO, &finfo) == -1) {
		perror("ioctl(FBIOGET_FSCREENINFO)");
		return -1;
	}

	fip->address = finfo.smem_start;
	fip->size = finfo.smem_len;
	fip->line_length = finfo.line_length;
	
	close(fd);
	return 0;
}

#define VESTR 0x00 /* start register */
#define VESWR 0x10 /* src: line length */
#define VESSR 0x14 /* src: image size */
#define VSAYR 0x18 /* src: y/rgb plane address */
#define VSACR 0x1c /* src: c plane address */
#define VBSSR 0x20 /* bundle mode register */
#define VEDWR 0x30 /* dst: line length */
#define VDAYR 0x34 /* dst: y/rgb plane address */
#define VDACR 0x38 /* dst: c plane address */
#define VTRCR 0x50 /* transform control */
#define VRFCR 0x54 /* resize scale */
#define VRFSR 0x58 /* resize clip */
#define VENHR 0x5c /* enhance */
#define VFMCR 0x70 /* filter mode */
#define VVTCR 0x74 /* lowpass vertical */
#define VHTCR 0x78 /* lowpass horizontal */
#define VAPCR 0x80 /* color match */
#define VECCR 0x84 /* color replace */
#define VAFXR 0x90 /* fixed mode */
#define VSWPR 0x94 /* swap */
#define VEIER 0xa0 /* interrupt mask */
#define VEVTR 0xa4 /* interrupt event */
#define VSTAR 0xb0 /* status */
#define VBSRR 0xb4 /* reset */

static unsigned long read_reg(struct uio_map *ump, int reg_nr)

{
	  volatile unsigned long *reg = ump->iomem + reg_nr;

	  return *reg;
}

static void write_reg(struct uio_map *ump, unsigned long value, int reg_nr)

{
	  volatile unsigned long *reg = ump->iomem + reg_nr;

	  *reg = value;
}

static void set_scale(struct uio_map *ump, int vertical,
int size_in, int size_out)

{
	unsigned long fixpoint, mant, frac, value;

	/* calculate FRAC and MANT */

	fixpoint = (4096 * (size_in - 1)) / (size_out + 1);
	mant = fixpoint / 4096;
	frac = fixpoint - (mant * 4096);

	if (frac & 0x07) {
		frac &= ~0x07;

		if (size_out > size_in)
			frac -= 8; /* round down if scaling up */
		else
			frac += 8; /* round up if scaling down */
	}

	/* set scale */
	
	value = read_reg(ump, VRFCR);
	
	if (vertical) {
		value &= ~0xffff0000;
  		value |= ((mant << 12) | frac) << 16;
	}
	else {
		value &= ~0xffff;
  		value |= (mant << 12) | frac;
	}
	
	write_reg(ump, value, VRFCR);

	/* set clip */

	value = read_reg(ump, VRFSR);

	if (vertical) {
		value &= ~0xffff0000;
  		value |= size_out << 16;
	}
	else {
		value &= ~0xffff;
  		value |= size_out;
	}
	
	write_reg(ump, value, VRFSR);
}

static int foo(void)

{
	struct fb_info fbi;
	struct uio_device uio_dev;
	struct uio_map uio_mmio, uio_mem;
	unsigned long addr;
	int src_w, src_h, src_bpp;
	int dst_w, dst_h;
	int ret;

	ret = get_fb_info("/dev/fb0", &fbi);
	if (ret < 0)
		return ret;

	printf("using %lux%lu %ld bpp framebuffer at 0x%08lx (0x%08lx).\n",
	       fbi.width, fbi.height, fbi.bpp, fbi.address, fbi.size);

	ret = locate_uio_device("VEU", &uio_dev);
	if (ret < 0)
		return ret;
	
	printf("found matching UIO device at %s\n", uio_dev.path);

	ret = setup_uio_map(&uio_dev, 0, &uio_mmio);
	if (ret < 0)
		return ret;

	ret = setup_uio_map(&uio_dev, 1, &uio_mem);
	if (ret < 0)
		return ret;

	printf("VSTAR = 0x%lx\n", read_reg(&uio_mmio, VSTAR));
	printf("VEVTR = 0x%lx\n", read_reg(&uio_mmio, VEVTR));

	/* reset VEU */

	printf("VBSRR = 0x%lx\n", read_reg(&uio_mmio, VBSRR));
	write_reg(&uio_mmio, 0x100, VBSRR);
	printf("VBSRR(2) = 0x%lx\n", read_reg(&uio_mmio, VBSRR));

	/* destination: get from fb */

	dst_w = fbi.width;
	dst_h = fbi.height;

	/* source bitmap */

	src_w = 32;
	src_h = 32;
	src_bpp = 16;

new_color:

	{
	  struct timeval tv;
	  int k;
	  unsigned char *buf = uio_mem.iomem;

	  gettimeofday(&tv, NULL);

	  for (k = 0; k < (src_w * src_h * (src_bpp / 8)); k++)
		buf[k] = k ^ (tv.tv_sec + tv.tv_usec);
	}

	addr = fbi.address;
	addr += (fbi.bpp / 8) * ((fbi.width - dst_w) / 2);
	addr += fbi.line_length * ((fbi.height - dst_h) / 2);

again:

	write_reg(&uio_mmio, src_w * (src_bpp / 8), VESWR);
	write_reg(&uio_mmio, src_w | (src_h << 16), VESSR);
	write_reg(&uio_mmio, uio_mem.address, VSAYR);
	write_reg(&uio_mmio, 0, VSACR); /* unused in case of RGB */
	write_reg(&uio_mmio, 0, VBSSR); /* not using bundle mode */

	write_reg(&uio_mmio, fbi.line_length, VEDWR);
	write_reg(&uio_mmio, addr, VDAYR);
	write_reg(&uio_mmio, 0, VDACR); /* unused in case of RGB */

	write_reg(&uio_mmio, (6 << 16) | (3 << 8) | 1, VTRCR);

	set_scale(&uio_mmio, 0, src_w, dst_w);
	set_scale(&uio_mmio, 1, src_h, dst_h);

	write_reg(&uio_mmio, 1, VEIER); /* enable interrupt */
	write_reg(&uio_mmio, 1, VESTR);

	/* Wait for an interrupt */
	{
		unsigned long n_pending;
		read(uio_dev.fd, &n_pending, sizeof(u_long));
	}

	write_reg(&uio_mmio, 0x100, VEVTR); /* ack int, write 0 to bit 0 */

	goto again;
	goto new_color;
} 

int main(int argc, char ** argv)

{
	foo();

	return 0;
}

Hello,

Yes it does. I thought the physical address is stored in internal_addr and the virtual in addr, but it's the other way round. Thanks.

Best regards
Uwe

Uwe Kleine-König, Software Engineer
Digi International GmbH Branch Breisach, Küferstrasse 8, 79206 Breisach, Germany
Tax: 315/5781/0242 / VAT: DE153662976 / Reg. Amtsgericht Dortmund HRB 13962

Hi Uwe!

Thanks for your email.

I'm not sure if it will help. What would such function do? Please explain.

Most UIO kernel drivers today contain hardware device specific code to acknowledge interrupts. In other words, most UIO interrupt handlers touches some device specific bits so the interrupt gets deasserted.

My uio_platform driver handles interrupts in a different way. The kernel UIO driver is not aware of the hardware device specific method to acknowledge the interrupt, instead it simply disables the interrupt and notifies user space which instead will acknowledge the interrupt. Next time a read() or poll() call gets made, the interrupt is enabled again.

This allows us to export a hardware device to user space and allow user space to handle interrupts without knowing in kernel space how to acknowledge interrupts.

Yes, I need that to give the exported hardware device some physically contiguous memory for DMA. At this point our hardware is missing IOMMU so physically contiguous memory is needed.

I don't mind that you are using platform devices. Actually, I think platform devices are great. We use them for all sorts of things on the SuperH architecture. I'm trying to suggest that maybe it's a good idea to change the UIO core code to use struct resource instead of struct uio_mem. Or maybe that's not a good idea, I'm not sure.

I don't want to pass any handler. All devices share the same interrupt handler, the only thing that differs between multiple uio_platform devices on one system is memory window information and irq number.

Ah, thanks for pointing that out!

Thank you for your feedback!

/ magnus

Just add irq_disabled to struct uio_platdata and define

	irqreturn_t uio_pdrv_disirq(int irq, struct uio_info *dev_info)
	{
		struct uio_platdata *pdata = container_of(dev_info, struct uio_platdata, uio_info);

		disable_irq(irq);
		pdata->irq_disabled = 1;
		return IRQ_HANDLED;
	}
	EXPORT_SYMBOL(uio_pdrv_disirq);

	void uio_pdrv_enirq(struct uio_info *dev_info)
	{
		...
	}
	EXPORT_SYMBOL(uio_pdrv_enirq);

and then you can do

	info->handler = uio_pdrv_disirq;
	info->enable_irq = uio_pdrv_enirq;

in the arch specific code. I just realize that you need to compile UIO statically then :-(

IMHO something like prep_read_poll is a better name than enable_irq for the new member of uio_info.

OK, got it. The down-side is that you can only get a single interrupt between two calls to read() (or poll()). So you might or might not loose information. And you might run into problems if your device or your interrupt goes berserk as your handler always returns IRQ_HANDLED. With a functional handler you can rely on existing mechanisms in the kernel.

struct resource alone doesn't provide enough information. At least memtype is needed. And you don't need the pointers *parent, *sibling, *child, so in my eyes it's fine to have a dedicated structure for uio.

See above. That would be the cost to share code with "my" driver.

All in all I'm not conviced that it's a good idea to use the irq_disable trick to save acking in kernel space. This doesn't need to stop you doing it that way of course.

Best regards
Uwe

Uwe Kleine-König, Software Engineer
Digi International GmbH Branch Breisach, Küferstrasse 8, 79206 Breisach, Germany
Tax: 315/5781/0242 / VAT: DE153662976 / Reg. Amtsgericht Dortmund HRB 13962

I understand now. Thanks for the clear description.

What about letting the uio_pdrv code override info->handler and info->enable_irq with the above functions if info->handler is NULL? That would be one step closer to a shared driver in my opinion. And it would remove the need for symbol exports and solve the static-compile-only issue.

The physically contiguous memory issue still needs to be solved somehow though. What about using struct resouce flagged as IORESOURCE_DMA to pass the amount of memory that should be allocated?

That's fine.

I agree that I only get a single interrupt, but I'm not agreeing regarding the problems. =)

In my mind, disabling interrupts and acking them from user space only leads to increased interrupt latencies. People may dislike increased interrupt latencies, but if so they shouldn't have their driver in user space. And you may of course choose to ack interrupts in kernel space and queue information there which user space later reads out. But that sounds more like a specialized kernel driver. And that is not what i'm trying to do.

Regarding loosing information, if your hardware device can't cope with long latencies and drops things on the floor then improve your latency, increase buffer size or design better hardware. Also, I don't think the interrupt can go berserk since it will be disabled directly by the interrupt handler.

Maybe the flags member of struct resource together with IORESOURCE_xxx can be used instead of memtype. But there is no point in changing things just for the sake of it, so it is fine as-is in my opinion.

Thank you!

/ magnus

... if both info->handler and info->prep_read_poll are NULL and info->irq >= 0.

I'm not sure that solving that problem in uio_pdrv is the right approach. Other uio drivers might have the same problem, so better allow the userspace driver to allocate some memory in a more generic way?

Assume your irq is stuck at its active level. Normally the irq is then disabled after some time. You can handle that in your userspace driver, but with acking in kernel space and returning IRQ_NONE or IRQ_HANDLED you get it for free. Nevertheless, go on.

Best regards
Uwe

Uwe Kleine-König, Software Engineer
Digi International GmbH Branch Breisach, Küferstrasse 8, 79206 Breisach, Germany
Tax: 315/5781/0242 / VAT: DE153662976 / Reg. Amtsgericht Dortmund HRB 13962

Even better! =)

I don't think there is any generic way for a user space driver to allocate physically contiguous memory. If such way exists then we should use that instead of course. Recommendations anyone?

Ok, so normally if the irq is stuck as asserted then it gets disabled after some time. In my case it gets disabled directly so see it as a feature. =)

Would you like to fold in the irq_handler and irq_enable function in your patch, or would you like me to make a patch that fits on top of your latest version?

Thanks for your help!

/ magnus

I would prefer the latter, because my patch already has acks and is complete as such. Moreover your suggestion needs the irq_enable patch.

Best regards
Uwe

Uwe Kleine-König, Software Engineer
Digi International GmbH Branch Breisach, Küferstrasse 8, 79206 Breisach, Germany
Tax: 315/5781/0242 / VAT: DE153662976 / Reg. Amtsgericht Dortmund HRB 13962