Skip to contentSkip to conversation

: Fix CDROM skipping in ieee1275 ofdisk iterator.

PATCH In review David Miller 2 participants 5 messages Dec 7, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 : Fix CDROM skipping in ieee1275 ofdisk iterator.

We skip CDROMs in the block device iterator because they simply can't be openned safely unless we absolutely know there is a CDROM drive there and there is in fact a disc in the drive.

Openning it blindly results in either a 10 second hang during the open call, or even worse a crash in OpenFirmware.

My SunBlade100 wouldn't boot a kernel properly because not only does it by default have a "cdrom" alias, it also has two others named "cdrom1" and "cdrom2".

So to cover all cases, match "cdrom" as a 5 character prefix of the alias name in the check.

I tried to surmise other ways to do this more cleanly, such as resolving the path and looking for some device property inside of the resulting node, but there simply isn't anything we can check for.

BTW, if people want to look at some OpenFirmware device tree dumps on just about every sparc64 system type there is a GIT repo where we maintain such things contributed by various sparc64 Linux users at:

git://git.kernel.org/pub/scm/linux/kernel/git/davem/prtconfs.git

There are some 32-bit sparc dumps in there, but the majority are 64-bit systems.

Anyways, here is the fix, please apply thanks!

2009-12-06 David S. Miller <[email protected]>

* disk/ieee1275/ofdisk.c (grub_ofdisk_iterate): Recognize anything even prefixed with 'cdrom' as a cdrom.

=== modified file 'disk/ieee1275/ofdisk.c'

ieee1275/ofdisk.c +1 -1
@@ -107,7 +107,7 @@
 	}
 
       if (! grub_strcmp (alias->type, "block") &&
-	  grub_strcmp (alias->name, "cdrom"))
+	  grub_strncmp (alias->name, "cdrom", 5))
 	ret = hook (alias->name);
       return ret;
     }
#1 lore raw

Committed.

Quoted 5 lines · David Miller

I tried to surmise other ways to do this more cleanly, such as resolving the path and looking for some device property inside of the resulting node, but there simply isn't anything we can check for.

If you find a cleaner way we can commit it later. I was thinking of resolving path to hardware one and then querying hardware directly but it may conflict with OF. What do you think?
How much work would be needed to make ata.mod work on sparc64? Which other drivers we would need to reasonably replace ofdisk (which would still be available as fallback)? In dumps in addition to ide I see protocols FP, ESP, SCSI, FAS, ISP, ULSA, SAS, SoC. ESP looks like SBUS-only so it can be fallbacked to OF considering number of users. I have no idea what FP,FAS, ISP, ULSA, SAS, SoC stand for and even if they are real protocols or just some weird pathnames. As for SCSI we definitely want it and already have scsi.mod but not the encapsulator of SCSI commands

Signature · 2 lines
Regards
Vladimir 'φ-coder/phcoder' Serbinenko
#2 lore raw

From: Vladimir '$B&U(B-coder/phcoder' Serbinenko <[email protected]>
Date: Mon, 07 Dec 2009 12:05:24 +0100

Quoted 10 lines

Committed.

Quoted 5 lines · David Miller

I tried to surmise other ways to do this more cleanly, such as resolving the path and looking for some device property inside of the resulting node, but there simply isn't anything we can check for.

If you find a cleaner way we can commit it later. I was thinking of resolving path to hardware one and then querying hardware directly but it may conflict with OF. What do you think?

I don't think that is wise.

We would then need two mappings, OF --> OS and OS --> OF for device naming. And all of that complexity for what? This silly cdrom issue where a simple strncmp() is solving the problem quite well so far?

There are conventions for alias names created by default by the firmware, so we should be OK with the patch I have written here.

Quoted 1 line

How much work would be needed to make ata.mod work on sparc64?

Probably a lot.

Quoted 2 lines

Which other drivers we would need to reasonably replace ofdisk (which would still be available as fallback)?

I don't think it's worthwhile to break away from ofdisk, it works quite well.

#3 lore raw
Quoted 18 lines · David Miller

From: Vladimir 'φ-coder/phcoder' Serbinenko <[email protected]>
Date: Mon, 07 Dec 2009 12:05:24 +0100

Quoted 9 lines

Committed.

Quoted 6 lines · David Miller

I tried to surmise other ways to do this more cleanly, such as resolving the path and looking for some device property inside of the resulting node, but there simply isn't anything we can check for.

Quoted 4 lines

If you find a cleaner way we can commit it later. I was thinking of resolving path to hardware one and then querying hardware directly but it may conflict with OF. What do you think?

Quoted 4 lines

I don't think that is wise.

We would then need two mappings, OF --> OS and OS --> OF for device naming.

I don't understand why we would need such a mapping. I meant to use full OF path which gives us where device is attached physically and so we can connect to it.

Quoted 9 lines

And all of that complexity for what? This
silly cdrom issue where a simple strncmp() is solving the problem quite well so far?

There are conventions for alias names created by default by the firmware, so we should be OK with the patch I have written here.

The problem is that if user has defined additional aliases we fall into original problem.

Quoted 6 lines
Quoted 2 lines

How much work would be needed to make ata.mod work on sparc64?

Probably a lot.

It already works on MIPS and AFAICT only thing it needs to work is PCI. Do you mean that accessing PCI on sparc is a lot of work?

Signature · 2 lines
Regards
Vladimir 'φ-coder/phcoder' Serbinenko
#4 lore raw

From: Vladimir '$B&U(B-coder/phcoder' Serbinenko <[email protected]>
Date: Mon, 07 Dec 2009 12:52:01 +0100

Quoted 2 lines

It already works on MIPS and AFAICT only thing it needs to work is PCI. Do you mean that accessing PCI on sparc is a lot of work?

Even just calculating the register addresses is a lot of work.

You can't just take the property values for the device as-is, you have to walk up to the root applying 'range' property translations to the register addresses.

And there are special cases and quirks all over the place.

The code to do this in the Linux kernel is several thousand lines of code.