Skip to contentSkip to conversation

drivers/ide/pci/alim15x3.c SMP fix

PATCH In review Ingo Molnar 3 participants 12 messages Sep 1, 2005 – Oct 17, 2005

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drivers/ide/pci/alim15x3.c SMP fix

is this the right way to fix the UP assumption below?

Ingo

drivers/ide/pci/alim15x3.c +8 -8
Index: linux/drivers/ide/pci/alim15x3.c
===================================================================
@@ -323,7 +323,7 @@ static void ali15x3_tune_drive (ide_driv
 		if (r_clc >= 16)
 			r_clc = 0;
 	}
-	local_irq_save(flags);
+	spin_lock_irqsave(&ide_lock, flags);
 	
 	/* 
 	 * PIO mode => ATA FIFO on, ATAPI FIFO off
@@ -345,7 +345,7 @@ static void ali15x3_tune_drive (ide_driv
 	
 	pci_write_config_byte(dev, port, s_clc);
 	pci_write_config_byte(dev, port+drive->select.b.unit+2, (a_clc << 4) | r_clc);
-	local_irq_restore(flags);
+	spin_unlock_irqrestore(&ide_lock, flags);
 
 	/*
 	 * setup   active  rec
@@ -601,7 +601,7 @@ static unsigned int __devinit init_chips
 	}
 #endif  /* defined(DISPLAY_ALI_TIMINGS) && defined(CONFIG_PROC_FS) */
 
-	local_irq_save(flags);
+	spin_lock_irqsave(&ide_lock, flags);
 
 	if (m5229_revision < 0xC2) {
 		/*
@@ -614,7 +614,7 @@ static unsigned int __devinit init_chips
 		 * clear bit 7
 		 */
 		pci_write_config_byte(dev, 0x4b, tmpbyte & 0x7F);
-		local_irq_restore(flags);
+		spin_unlock_irqrestore(&ide_lock, flags);
 		return 0;
 	}
 
@@ -639,7 +639,7 @@ static unsigned int __devinit init_chips
 	 * 0:0.0 so if we didn't find one we know what is cooking.
 	 */
 	if (north && north->vendor != PCI_VENDOR_ID_AL) {
-		local_irq_restore(flags);
+		spin_unlock_irqrestore(&ide_lock, flags);
 	        return 0;
 	}
 
@@ -662,7 +662,7 @@ static unsigned int __devinit init_chips
 			pci_write_config_byte(isa_dev, 0x79, tmpbyte | 0x02);
 		}
 	}
-	local_irq_restore(flags);
+	spin_unlock_irqrestore(&ide_lock, flags);
 	return 0;
 }
 
@@ -686,7 +686,7 @@ static unsigned int __devinit ata66_ali1
 	unsigned long flags;
 	u8 tmpbyte;
 
-	local_irq_save(flags);
+	spin_lock_irqsave(&ide_lock, flags);
 
 	if (m5229_revision >= 0xC2) {
 		/*
@@ -736,7 +736,7 @@ static unsigned int __devinit ata66_ali1
 
 	pci_write_config_byte(dev, 0x53, tmpbyte);
 
-	local_irq_restore(flags);
+	spin_unlock_irqrestore(&ide_lock, flags);
 
 	return(ata66);
 }
#1 lore raw
Quoted 8 lines · Ingo Molnar

is this the right way to fix the UP assumption below?

Ingo

Index: linux/drivers/ide/pci/alim15x3.c
[snip]

OK. The reported boot WARNING seems to be over now. Tested on the offended laptop ([email protected]/UP, PCI chipset: ALi M1533) with 2.6.13-rt3, where the suggested patch on drivers/ide/pci/alim15x3.c seems to fix the burp. All seems to be working fine, still ;)

Thanks.

Signature · 2 lines
rncbc aka Rui Nuno Capela
[email protected]
#2 lore raw
Quoted 7 lines · Rui Nuno Capela
Quoted 6 lines · Ingo Molnar

is this the right way to fix the UP assumption below?

Ingo

Quoted 7 lines
Quoted 2 lines

Index: linux/drivers/ide/pci/alim15x3.c
[snip]

OK. The reported boot WARNING seems to be over now. Tested on the offended laptop ([email protected]/UP, PCI chipset: ALi M1533) with 2.6.13-rt3, where the suggested patch on drivers/ide/pci/alim15x3.c seems to fix the burp. All seems to be working fine, still ;)

just to make sure the original point gets across: the warning is only in the -rt tree, and it pinpoints potential SMP bugs. Does your box do hyperthreading? If yes then this could be a live (but probably mostly harmless) SMP bug. Maybe the whole IRQ disabling is unnecessary? __devinit is mostly serialized, so i'm not sure there's any protection needed against parallel IRQs?

Ingo

#3 lore raw
Quoted 4 lines · Ingo Molnar
Quoted 1 line · Rui Nuno Capela

OK. The reported boot WARNING seems to be over now. Tested on the offended laptop ([email protected]/UP, PCI chipset: ALi M1533) with 2.6.13-rt3, where the suggested patch on drivers/ide/pci/alim15x3.c seems to fix the burp. All seems to be working fine, still ;)

just to make sure the original point gets across: the warning is only in the -rt tree, and it pinpoints potential SMP bugs. Does your box do hyperthreading?

Nope. The [email protected] is from pre-HT age, and the kernel is being built for UP.

Bye now.

Signature · 2 lines
rncbc aka Rui Nuno Capela
[email protected]
#4 lore raw
In reply to Ingo Molnar
Quoted 1 line · Ingo Molnar

is this the right way to fix the UP assumption below?

Probably not. The ide_lock may already be held (randomly depending on the code path) at the point we retune a drive on error. Actually you probably crash before this anyway.

The ALi code in question was written knowing the system would be uniprocessor and making various related assumptions. You also have to get this locking right just to make it more fun - loading the timings for one channel while another is doing I/O corrupts your data silently in some cases. Fixing the ide_lock to be consistent in usage when the tuning calls are made (ie fix the reset path and other offenders) might be possible and would make using ide_lock ok, but it would still be wrong with pre-emption and/or SMP.

There is currently no sane locking mechanism to enforce or implement this in the IDE layer. Welcome to hell, please leave your brain at the door.

Alan

#5 lore raw
Quoted 7 lines · Alan Cox
Quoted 1 line · Ingo Molnar

is this the right way to fix the UP assumption below?

Probably not. The ide_lock may already be held (randomly depending on the code path) at the point we retune a drive on error. Actually you probably crash before this anyway.

Quoted 8 lines

The ALi code in question was written knowing the system would be uniprocessor and making various related assumptions. You also have to get this locking right just to make it more fun - loading the timings for one channel while another is doing I/O corrupts your data silently in some cases. Fixing the ide_lock to be consistent in usage when the tuning calls are made (ie fix the reset path and other offenders) might be possible and would make using ide_lock ok, but it would still be wrong with pre-emption and/or SMP.

so perhaps part of the solution would be to do the initialization under the IDE lock, via the patch below? It boots fine on my box so the basic codepaths seem to be OK. Then the retuning codepaths need to be checked to make sure they are holding the IDE lock.

Ingo

drivers/ide/setup-pci.c +5 -0
Index: linux/drivers/ide/setup-pci.c
===================================================================
@@ -665,8 +665,11 @@ static int do_ide_setup_pci_device(struc
 {
 	static ata_index_t ata_index = { .b = { .low = 0xff, .high = 0xff } };
 	int tried_config = 0;
+	unsigned long flags;
 	int pciirq, ret;
 
+	spin_lock_irqsave(&ide_lock, flags);
+
 	ret = ide_setup_pci_controller(dev, d, noisy, &tried_config);
 	if (ret < 0)
 		goto out;
@@ -721,6 +724,8 @@ static int do_ide_setup_pci_device(struc
 	*index = ata_index;
 	ide_pci_setup_ports(dev, d, pciirq, index);
 out:
+	spin_unlock_irqrestore(&ide_lock, flags);
+
 	return ret;
 }
 
#6 lore raw
Quoted 4 lines · Ingo Molnar

so perhaps part of the solution would be to do the initialization under the IDE lock, via the patch below? It boots fine on my box so the basic codepaths seem to be OK. Then the retuning codepaths need to be checked to make sure they are holding the IDE lock.

The initialisation can take several seconds. Some controller paths may also sleep so you would need to do a full audit first.

Alan

#7 lore raw

tsc_c3_compensate undefined since patch-2.6.13-rt13

In reply to Ingo Molnar

Ingo,

I'll take this late opportunity to report something that have been looking suspicious since 2.6.13-rt13, inclusive, about this symbol of tsc_c3_compensate being undefined and causing some noise on all kernel builds since then.

To put things in brief, here follows a small exchange that took place on the linux-audio-user list, regarding this thingie. Apparently for Mark, it was a kernel build showstopper.

Quoted 34 lines · Mark Knecht

Hi there,
I have a newish AMD64 machine. NForce4 chipset. Asus A8N-E
motherboard. PCI-Express 16x graphics, etc. No matter what I try I've been constantly stopped from building a 2.6.13 kernel with Ingo's rt14 patch.

Here's the error message:

AS arch/x86_64/lib/copy_user.o

  AS      arch/x86_64/lib/csum-copy.o
  CC      arch/x86_64/lib/csum-partial.o
  CC      arch/x86_64/lib/csum-wrappers.o
 CC      arch/x86_64/lib/dec_and_lock.o
 CC      arch/x86_64/lib/delay.o
  AS      arch/x86_64/lib/getuser.o
  AS      arch/x86_64/lib/putuser.o
 AS      arch/x86_64/lib/thunk.o
  CC      arch/x86_64/lib/usercopy.o
 AR      arch/x86_64/lib/lib.a
 GEN     .version
  CHK     include/linux/compile.h
  UPD     include/linux/compile.h
  CC      init/version.o
  LD      init/built-in.o
  LD      .tmp_vmlinux1

drivers/built-in.o(.text+0x24acc): In function `acpi_processor_idle':
: undefined reference to `tsc_c3_compensate'
make: *** [.tmp_vmlinux1] Error 1
lightning linux #

The 2.6.13 kernel builds fine before I apply the patch but fails afterward.

I do not find the error message

Quoted 9 lines

undefined reference to `tsc_c3_compensate'

in Google so maybe it's just me and my kernel config. I've attached the config file zipped. I've tried some obvious stuff like completely disabling ACPI, etc., but I haven't found anything that works yet.

Thanks in advance for any ideas. This has held up my new AMD64 machine being useful for a few weeks now.

I'm spotting a very similar message when building 2.6.13.x-rt14, but as a module linkage warning, not a fatal build error.

Maybe that's because I try to configure _everything_ I can as a module, not as built-in. As this has been just a warning on the module install phase I got along and all seems to boot and run fine -- except for some acpi stuff e.g. I've lost thermal zone sensor on my laptop, but that didn't look like a big deal.

Yep, that's it... I found the exact messages in my attic, although from an erlier build:

WARNING: /lib/modules/2.6.13.1-rt13.0mdk/kernel/drivers/char/hangcheck-timer.ko needs unknown symbol do_monotonic_clock
WARNING: /lib/modules/2.6.13.1-rt13.0mdk/kernel/drivers/acpi/processor.ko needs unknown symbol tsc_c3_compensate

How about reporting to Ingo and/or the lkml? As you can see its not an AMD64 issue, because I don't have such a beast here.

Bye.

Signature · 2 lines
rncbc aka Rui Nuno Capela
[email protected]
#8 lore raw

Re: tsc_c3_compensate undefined since patch-2.6.13-rt13

Quoted 10 lines · Rui Nuno Capela

Ingo,

I'll take this late opportunity to report something that have been looking suspicious since 2.6.13-rt13, inclusive, about this symbol of tsc_c3_compensate being undefined and causing some noise on all kernel builds since then.

To put things in brief, here follows a small exchange that took place on the linux-audio-user list, regarding this thingie. Apparently for Mark, it was a kernel build showstopper.

thanks for the reminder!

Quoted 6 lines

WARNING:
/lib/modules/2.6.13.1-rt13.0mdk/kernel/drivers/char/hangcheck-timer.ko needs unknown symbol do_monotonic_clock
WARNING:
/lib/modules/2.6.13.1-rt13.0mdk/kernel/drivers/acpi/processor.ko needs unknown symbol tsc_c3_compensate

back then i fixed do_monotonic_clock, but forgot to export tsc_c3_compensate. I have fixed this in my tree, and have uploaded the 2.6.14-rc3-rt3 patch. Does it build without warnings for you now?

Ingo

#9 lore raw

Re: tsc_c3_compensate undefined since patch-2.6.13-rt13

Quoted 8 lines · Ingo Molnar
Quoted 2 lines · Rui Nuno Capela

WARNING: /lib/modules/2.6.13.1-rt13.0mdk/kernel/drivers/char/hangcheck-timer.ko needs unknown symbol do_monotonic_clock
WARNING: /lib/modules/2.6.13.1-rt13.0mdk/kernel/drivers/acpi/processor.ko needs unknown symbol tsc_c3_compensate

back then i fixed do_monotonic_clock, but forgot to export tsc_c3_compensate. I have fixed this in my tree, and have uploaded the 2.6.14-rc3-rt3 patch. Does it build without warnings for you now?

OK. Just built 2.6.14-rc3-rt4 and it got thru. Thermal-zone sensor is back in town :)

Tks.

Signature · 2 lines
rncbc aka Rui Nuno Capela
[email protected]
#10 lore raw

Re: tsc_c3_compensate undefined since patch-2.6.13-rt13

In reply to Ingo Molnar
Quoted 10 lines · Ingo Molnar
Quoted 5 lines · Rui Nuno Capela

Ingo,

I'll take this late opportunity to report something that have been looking suspicious since 2.6.13-rt13, inclusive, about this symbol of tsc_c3_compensate being undefined and causing some noise on all kernel builds since then.

To put things in brief, here follows a small exchange that took place on the linux-audio-user list, regarding this thingie. Apparently for Mark, it was a kernel build showstopper.

Quoted 9 lines

thanks for the reminder!

Quoted 2 lines

WARNING: /lib/modules/2.6.13.1-rt13.0mdk/kernel/drivers/char/hangcheck-timer.ko needs unknown symbol do_monotonic_clock
WARNING: /lib/modules/2.6.13.1-rt13.0mdk/kernel/drivers/acpi/processor.ko needs unknown symbol tsc_c3_compensate

back then i fixed do_monotonic_clock, but forgot to export tsc_c3_compensate. I have fixed this in my tree, and have uploaded the 2.6.14-rc3-rt3 patch. Does it build without warnings for you now?

As it seems, tsc_c3_compensate isn't being defined again, as of patch-2.6.14-rc4-rt6 (-rt4 was ok).

Cheers.
--
rncbc aka Rui Nuno Capela
[email protected]

#11 lore raw

Re: tsc_c3_compensate undefined since patch-2.6.13-rt13

Quoted 2 lines · Rui Nuno Capela

As it seems, tsc_c3_compensate isn't being defined again, as of patch-2.6.14-rc4-rt6 (-rt4 was ok).

oops - export went MIA during the clockevents merge. Have put it back again. (will be in -rt7)

Ingo