Re: [PATCH] speedy boot from usb devices

Re: [PATCH] speedy boot from usb devices

[Heikki Linnakangas]

David N. Welton wrote:
> where you can get blkdev_wakeup.patch
> 
>         Works like so: whenever a block device comes on line, it
>         signals this fact to a wait queue, so that the init process
>         can stop and wait for slow devices, in particular things such
>         as USB storage devices, which are much slower than IDE
>         devices.  The init process checks the list of available
>         devices and compares it with the desired root device, and if
>         there is a match, proceeds with the initialization process,
>         secure in the knowledge that the device in question has been
>         brought up.  This is useful if one wants to boot quickly from
>         a USB storage device without a trimmed-down kernel, and
>         without going through the whole initrd slog.

I tried the patch, it looks good to me in principle. Thanks! However, it 
didn't work for me :(.

The problem seems to be that the wait is too late. I start my kernel with 
"root=/dev/sda1", and that is parsed to major/minor numbers in the 
name_to_dev_t function which is in turn called by prepare_namespace. The 
call to name_to_dev_t returns 0 because the device isn't attached yet, 
and the mount fails later because of that.

The obvious solution is to move the wait earlier in the chain. Here's a 
modified patch that works for me. I also eliminated the get_blkdevs and 
match_root stuff, and rely on name_to_dev_t succeeding or failing instead.

This is still not perfect. It doesn't work correctly if you try to give 
the major/minor numbers directly in the kernel command line, as 
in "root=8:1". I guess the final solution would be to retry the whole 
mount process from the call to name_to_dev to a successful mount every 
time we get woke up by the wait queue.

- Heikki

[Please CC replies]


diff -ru --exclude='.*' linux-2.6.8-rc2-mm1-orig/drivers/block/genhd.c linux-2.6.8-rc2-mm1/drivers/block/genhd.c
--- linux-2.6.8-rc2-mm1-orig/drivers/block/genhd.c	2004-08-04 23:37:56.000000000 +0300
+++ linux-2.6.8-rc2-mm1/drivers/block/genhd.c	2004-08-04 23:51:09.000000000 +0300
@@ -14,9 +14,14 @@
  #include <linux/slab.h>
  #include <linux/kmod.h>
  #include <linux/kobj_map.h>
+#include <linux/wait.h>
+#include <linux/sched.h>

  #define MAX_PROBE_HASH 255	/* random */

+DECLARE_WAIT_QUEUE_HEAD(disk_wait_h);
+EXPORT_SYMBOL(disk_wait_h);
+
  static struct subsystem block_subsys;

  /*
@@ -197,6 +202,11 @@
  			    disk->minors, NULL, exact_match, exact_lock, disk);
  	register_disk(disk);
  	blk_register_queue(disk);
+
+	/* Wake up queue in init/main.c. */
+	printk("Waking up queue on %s\n",
+	       disk->disk_name);
+	wake_up_interruptible(&disk_wait_h);
  }

  EXPORT_SYMBOL(add_disk);
diff -ru --exclude='.*' linux-2.6.8-rc2-mm1-orig/include/linux/genhd.h linux-2.6.8-rc2-mm1/include/linux/genhd.h
--- linux-2.6.8-rc2-mm1-orig/include/linux/genhd.h	2004-08-04 23:40:29.000000000 +0300
+++ linux-2.6.8-rc2-mm1/include/linux/genhd.h	2004-08-05 00:02:53.543231776 +0300
@@ -16,6 +16,10 @@
  #include <linux/smp.h>
  #include <linux/string.h>
  #include <linux/fs.h>
+#include <linux/wait.h>
+
+/* Wait queue to wait on disks coming online. */
+extern wait_queue_head_t disk_wait_h;

  enum {
  /* These three have identical behaviour; use the second one if DOS FDISK gets
diff -ru --exclude='.*' linux-2.6.8-rc2-mm1-orig/init/do_mounts.c linux-2.6.8-rc2-mm1/init/do_mounts.c
--- linux-2.6.8-rc2-mm1-orig/init/do_mounts.c	2004-06-16 08:19:13.000000000 +0300
+++ linux-2.6.8-rc2-mm1/init/do_mounts.c	2004-08-04 23:07:23.000000000 +0300
@@ -11,6 +11,8 @@
  #include <linux/nfs_fs_sb.h>
  #include <linux/nfs_mount.h>

+#include <linux/genhd.h>
+
  #include "do_mounts.h"

  extern int get_filesystem_list(char * buf);
@@ -390,7 +392,14 @@

  	if (saved_root_name[0]) {
  		root_device_name = saved_root_name;
-		ROOT_DEV = name_to_dev_t(root_device_name);
+
+		/* Here, we wait for the root device to show up, or if it's
+		 * already there, we just go on. */
+		while ((ROOT_DEV = name_to_dev_t(root_device_name)) == 0) {
+			printk("Waiting for root device to wake us up\n");
+			interruptible_sleep_on(&disk_wait_h);
+		}
+
  		if (strncmp(root_device_name, "/dev/", 5) == 0)
  			root_device_name += 5;
  	}

Re: [PATCH] speedy boot from usb devices

[David N. Welton]

Heikki Linnakangas <hlinnaka@iki.fi> writes:

[ Please CC replies to me. ]

> I tried the patch, it looks good to me in principle. Thanks!
> However, it didn't work for me :(.

Yes, I wasn't able to test it on the devices I wanted to untill after
I released it.  Sorry 'bout that.

> The problem seems to be that the wait is too late. I start my kernel
> with "root=/dev/sda1", and that is parsed to major/minor numbers in
> the name_to_dev_t function which is in turn called by
> prepare_namespace. The call to name_to_dev_t returns 0 because the
> device isn't attached yet, and the mount fails later because of
> that.

Right.

> The obvious solution is to move the wait earlier in the
> chain. Here's a modified patch that works for me. I also eliminated
> the get_blkdevs and match_root stuff, and rely on name_to_dev_t
> succeeding or failing instead.

> This is still not perfect. It doesn't work correctly if you try to
> give the major/minor numbers directly in the kernel command line, as
> in "root=8:1". I guess the final solution would be to retry the
> whole mount process from the call to name_to_dev to a successful
> mount every time we get woke up by the wait queue.

We also need to consider 'rdev'ed devices (for lack of a better term)
where there is a "built-in" dev_t, but no name (afaik).

I like your use of name_to_dev_t, that seems maybe more efficient than
what I have even here.  In this version, I consider "built in" root
devices (and the patch works, too:-).

Hrm... maybe the best approach is to create the root_device_name if it
doesn't exist from the ROOT_DEV, if it does, and use that with
name_to_dev_t...

At this point I will probably have to wait till next week to look into
it further.

-- 
David N. Welton
     Personal: http://www.dedasys.com/davidw/
Free Software: http://www.dedasys.com/freesoftware/
   Apache Tcl: http://tcl.apache.org/
       Photos: http://www.dedasys.com/photos/

diff -ru linux-2.6.7/drivers/block/genhd.c hacked/drivers/block/genhd.c
--- linux-2.6.7/drivers/block/genhd.c	2004-06-16 07:19:22.000000000 +0200
+++ hacked/drivers/block/genhd.c	2004-08-04 18:26:32.000000000 +0200
@@ -14,9 +14,14 @@
 #include <linux/slab.h>
 #include <linux/kmod.h>
 #include <linux/kobj_map.h>
+#include <linux/wait.h>
+#include <linux/sched.h>
 
 #define MAX_PROBE_HASH 255	/* random */
 
+DECLARE_WAIT_QUEUE_HEAD(disk_wait_h);
+EXPORT_SYMBOL(disk_wait_h);
+
 static struct subsystem block_subsys;
 
 /*
@@ -57,6 +62,58 @@
 	return len;
 }
 
+/*
+ *-----------------------------------------------------------------------------
+ *
+ * blkdev_is_online --
+ *
+ * 	Checks whether a named device is online or not.  Returns 1 if
+ * 	it is, 0 if not.  name is NULL if we already have a root
+ * 	device 'built in' to the kernel, and we use rootdev to compare
+ * 	with, rather than comparing the strings.
+ *
+ *-----------------------------------------------------------------------------
+ */
+
+int blkdev_is_online(char *name, dev_t rootdev)
+{
+    char buf[BDEVNAME_SIZE];
+    char *devname = NULL;
+    int retval = 0;
+    int major = 0;
+    int minor = 0;
+    struct gendisk *sgp = NULL;
+    struct list_head *p;
+
+    major = MAJOR(rootdev);
+    minor = MINOR(rootdev);
+
+    down_read(&block_subsys.rwsem);
+    list_for_each(p, &block_subsys.kset.list) {
+	sgp = list_entry(p, struct gendisk, kobj.entry);
+	devname = disk_name(sgp, 0, buf);
+	if (name != NULL) {
+	    /* We have a name, use that for the comparison.  */
+	    if (strncmp(devname,
+			name,
+			strlen(devname)) == 0) {
+		retval = 1;
+		break;
+	    }
+	} else {
+	    /* Compare with device numbers. */
+	    if (sgp->major == major && sgp->first_minor == minor) {
+		retval = 1;
+		break;
+	    }
+	}
+    }
+    up_read(&block_subsys.rwsem);
+    return retval;
+}
+
+EXPORT_SYMBOL(blkdev_is_online);
+
 int register_blkdev(unsigned int major, const char *name)
 {
 	struct blk_major_name **n, *p;
@@ -197,6 +254,11 @@
 			    disk->minors, NULL, exact_match, exact_lock, disk);
 	register_disk(disk);
 	blk_register_queue(disk);
+
+	/* Wake up queue in init/main.c. */
+	printk("Waking up queue on %s\n",
+	       disk->disk_name);
+	wake_up_interruptible(&disk_wait_h);
 }
 
 EXPORT_SYMBOL(add_disk);
Only in hacked/drivers/block: genhd.c~
diff -ru linux-2.6.7/include/linux/genhd.h hacked/include/linux/genhd.h
--- linux-2.6.7/include/linux/genhd.h	2004-06-16 07:18:59.000000000 +0200
+++ hacked/include/linux/genhd.h	2004-08-04 18:28:05.000000000 +0200
@@ -16,6 +16,10 @@
 #include <linux/smp.h>
 #include <linux/string.h>
 #include <linux/fs.h>
+#include <linux/wait.h>
+
+/* Wait queue to wait on disks coming online. */
+extern wait_queue_head_t disk_wait_h;
 
 enum {
 /* These three have identical behaviour; use the second one if DOS FDISK gets
@@ -191,6 +195,7 @@
 
 extern void set_device_ro(struct block_device *bdev, int flag);
 extern void set_disk_ro(struct gendisk *disk, int flag);
+extern int blkdev_is_online(char *name, dev_t rootdev);
 
 /* drivers/char/random.c */
 extern void add_disk_randomness(struct gendisk *disk);
Only in hacked/include/linux: genhd.h~
diff -ru linux-2.6.7/init/do_mounts.c hacked/init/do_mounts.c
--- linux-2.6.7/init/do_mounts.c	2004-06-16 07:19:13.000000000 +0200
+++ hacked/init/do_mounts.c	2004-08-04 18:25:49.000000000 +0200
@@ -11,6 +11,8 @@
 #include <linux/nfs_fs_sb.h>
 #include <linux/nfs_mount.h>
 
+#include <linux/genhd.h>
+
 #include "do_mounts.h"
 
 extern int get_filesystem_list(char * buf);
@@ -373,6 +375,34 @@
 			change_floppy("root floppy");
 	}
 #endif
+
+	/* Here, we wait for the root device to show up, or if it's
+	 * already there, we just go on. */
+
+	if (root_device_name != NULL) {
+	    /* Its name has been passed on the command line, so we
+	     * have to check it by name. */
+	    while (blkdev_is_online(root_device_name, 0) == 0) {
+		printk("Waiting for root device %s to wake us up\n",
+		       root_device_name);
+		interruptible_sleep_on(&disk_wait_h);
+	    }
+	    if (saved_root_name[0]) {
+		root_device_name = saved_root_name;
+		ROOT_DEV = name_to_dev_t(root_device_name);
+		if (strncmp(root_device_name, "/dev/", 5) == 0)
+		    root_device_name += 5;
+	    }
+	} else {
+	    /* No name, so we are going to have to check on it by ROOT_DEV. */
+	    while (blkdev_is_online(NULL, ROOT_DEV) == 0) {
+		printk("Waiting for root device %d:%d to wake us up\n",
+		       MAJOR(ROOT_DEV), MINOR(ROOT_DEV));
+		interruptible_sleep_on(&disk_wait_h);
+	    }
+	}
+
+
 	create_dev("/dev/root", ROOT_DEV, root_device_name);
 	mount_block_root("/dev/root", root_mountflags);
 }
@@ -393,6 +423,8 @@
 		ROOT_DEV = name_to_dev_t(root_device_name);
 		if (strncmp(root_device_name, "/dev/", 5) == 0)
 			root_device_name += 5;
+	} else {
+	    root_device_name = NULL;
 	}
 
 	is_floppy = MAJOR(ROOT_DEV) == FLOPPY_MAJOR;
Only in hacked/init: do_mounts.c~

[PATCH] speedy boot from usb devices

[David N. Welton]

[ Please CC replies to me - thanks! ]

Hi,

I gather that there isn't much interest in this idea on behalf of the
'mainstream', because initrd is a far more flexible solution, but I
happen to like the idea of booting quickly from a USB device, without
wasting a bunch of time and space for an initrd, and we're using this
in a product as well, so without further ado, I'll point you at

http://dedasys.com/freesoftware/patches/

where you can get blkdev_wakeup.patch

        Works like so: whenever a block device comes on line, it
        signals this fact to a wait queue, so that the init process
        can stop and wait for slow devices, in particular things such
        as USB storage devices, which are much slower than IDE
        devices.  The init process checks the list of available
        devices and compares it with the desired root device, and if
        there is a match, proceeds with the initialization process,
        secure in the knowledge that the device in question has been
        brought up.  This is useful if one wants to boot quickly from
        a USB storage device without a trimmed-down kernel, and
        without going through the whole initrd slog.

Comments, critiques, suggestions and ideas are all welcome.

Thankyou for your time,
-- 
David N. Welton
     Personal: http://www.dedasys.com/davidw/
Free Software: http://www.dedasys.com/freesoftware/
   Apache Tcl: http://tcl.apache.org/
       Photos: http://www.dedasys.com/photos/

Re: [PATCH] speedy boot from usb devices

[Paulo Marques]

David N. Welton wrote:
> [ Please CC replies to me - thanks! ]
> 
> Hi,
> 
> I gather that there isn't much interest in this idea on behalf of the
> 'mainstream', because initrd is a far more flexible solution, but I
> happen to like the idea of booting quickly from a USB device, without
> wasting a bunch of time and space for an initrd, and we're using this
> in a product as well, so without further ado, I'll point you at
> 
> http://dedasys.com/freesoftware/patches/
> 
> where you can get blkdev_wakeup.patch
> 
>         Works like so: whenever a block device comes on line, it
>         signals this fact to a wait queue, so that the init process
>         can stop and wait for slow devices, in particular things such
>         as USB storage devices, which are much slower than IDE
>         devices.  The init process checks the list of available
>         devices and compares it with the desired root device, and if
>         there is a match, proceeds with the initialization process,
>         secure in the knowledge that the device in question has been
>         brought up.  This is useful if one wants to boot quickly from
>         a USB storage device without a trimmed-down kernel, and
>         without going through the whole initrd slog.
> 
> Comments, critiques, suggestions and ideas are all welcome.

I find this to be very useful. I always found the "sleep for a while 
until the device we want appears" approach very cumbersome.

However, after looking at your patch, it seems that having a 
get_blkdevs() function that alloc's an array of strings, and return it 
to a function that only compares the strings against the name it is 
looking for and drops the array altogether, is a little overkill.

Why not have a simple blkdev_exists(char *name) function in genhd.c, 
call it directly, and drop the match_root_name() function completely?

Maybe I'm missing something, but it seems much simpler...

-- 
Paulo Marques - www.grupopie.com
"In a world without walls and fences who needs windows and gates?"

Re: [PATCH] speedy boot from usb devices

[David N. Welton]

Paulo Marques <pmarques@grupopie.com> writes:

> David N. Welton wrote:

> >         Works like so: whenever a block device comes on line, it
> >         signals this fact to a wait queue, so that the init
> >         process can stop and wait for slow devices, in particular
> >         things such as USB storage devices, which are much slower
> >         than IDE devices.  The init process checks the list of
> >         available devices and compares it with the desired root
> >         device, and if there is a match, proceeds with the
> >         initialization process, secure in the knowledge that the
> >         device in question has been brought up.  This is useful if
> >         one wants to boot quickly from a USB storage device
> >         without a trimmed-down kernel, and without going through
> >         the whole initrd slog.

> I find this to be very useful. I always found the "sleep for a while
> until the device we want appears" approach very cumbersome.

Glad to hear someone likes it.

> However, after looking at your patch, it seems that having a
> get_blkdevs() function that alloc's an array of strings, and return
> it to a function that only compares the strings against the name it
> is looking for and drops the array altogether, is a little overkill.

> Why not have a simple blkdev_exists(char *name) function in genhd.c,
> call it directly, and drop the match_root_name() function
> completely?

Sure, that's probably better.  Maybe "blkdev_is_online"?  I'll see if
I can do it tommorow.

I'm also a bit dubious of having the wait queue floating around as a
global, but don't know the kernel well enough to find it a better
home.

Thanks!
-- 
David N. Welton
     Personal: http://www.dedasys.com/davidw/
Free Software: http://www.dedasys.com/freesoftware/
   Apache Tcl: http://tcl.apache.org/
       Photos: http://www.dedasys.com/photos/

Re: [PATCH] speedy boot from usb devices

[Kalin KOZHUHAROV]

David N. Welton wrote:
> Paulo Marques <pmarques@grupopie.com> writes:
> 
> 
>>David N. Welton wrote:
> 
> 
>>>        Works like so: whenever a block device comes on line, it
>>>        signals this fact to a wait queue, so that the init
>>>        process can stop and wait for slow devices, in particular
>>>        things such as USB storage devices, which are much slower
>>>        than IDE devices.  The init process checks the list of
>>>        available devices and compares it with the desired root
>>>        device, and if there is a match, proceeds with the
>>>        initialization process, secure in the knowledge that the
>>>        device in question has been brought up.  This is useful if
>>>        one wants to boot quickly from a USB storage device
>>>        without a trimmed-down kernel, and without going through
>>>        the whole initrd slog.
>>I find this to be very useful. I always found the "sleep for a while
>>until the device we want appears" approach very cumbersome.

I hope it works. I tried to hack some delay on my own, but without success :-(
Will try this patch.

> Glad to hear someone likes it.
> 
> 
>>However, after looking at your patch, it seems that having a
>>get_blkdevs() function that alloc's an array of strings, and return
>>it to a function that only compares the strings against the name it
>>is looking for and drops the array altogether, is a little overkill.
>>Why not have a simple blkdev_exists(char *name) function in genhd.c,
>>call it directly, and drop the match_root_name() function
>>completely?
> 
> 
> Sure, that's probably better.  Maybe "blkdev_is_online"?  I'll see if
> I can do it tommorow.

Waiting :-)
I am planing to run a Gentoo-based small dist off a 128MB USB flash.

> I'm also a bit dubious of having the wait queue floating around as a
> global, but don't know the kernel well enough to find it a better
> home.

Thumbs up!

Kalin.

-- 
 || ~~~~~~~~~~~~~~~~~~~~~~ ||
(  ) http://ThinRope.net/ (  )
 || ______________________ ||