2.5.63: 'Debug: sleeping function called from illegal context
Quoted 10 lines · Alan
Quoted 4 lines · Dr. David Alan Gilbert
Hi,
2.5.63 had a good go at trying to boot for me; the only error during
boot was 'Debug: sleeping function called from illegal context atmm/slab.c:1617' during the IDE startup.Known problem. Its a bug in the
request_irqcode on x86. IDE just happens to be a victim of it.
I would call it an IDE bug: request_irq is not atomic, there are dozends of kmalloc(GFP_KERNEL) calls in the implementation. And on some archs it might be impossible to implement it without sleeping - e.g. on a NUMA arch an IPI to the right node could be necessary. What about fixing ide? If ide can't handle early interrupts, then it can use
disable_irq(); request_irq(); spin_lock_irq(&ide_lock); do_setup(); spin_unlock_irq(&ide_lock); enable_irq();
I've attached a proposal - what do you think?
--
Manfred
// $Header$
// Kernel Version:
// VERSION = 2 // PATCHLEVEL = 5 // SUBLEVEL = 63 // EXTRAVERSION =
include/linux/ide.h +13 -0
@@ -1730,6 +1730,19 @@extern int ide_set_xfer_rate (ide_drive_t *drive,u8 rate);extern spinlock_t ide_lock; +extern struct semaphore ide_cfg_sem; +/* +* Structure locking: +* +* ide_cfg_sem and ide_lock together protect changes to +* ide_hwif_t->{next,hwgroup} +* ide_drive_t->next +* +* ide_hwgroup_t->busy: ide_lock +* ide_hwgroup_t->hwif: ide_lock +* ide_hwif_t->mate: constant, no locking +* ide_drive_t->hwif: constant, no locking +*/ #define local_irq_set (flags) do { local_save_flags((flags)); local_irq_enable(); } while (0)
drivers/ide/ide-probe.c +17 -8
@@ -1032,16 +1032,17 @@*/ static int init_irq (ide_hwif_t *hwif) { -unsigned long flags;unsigned int index; +int do_enable_irq;ide_hwgroup_t *hwgroup, *new_hwgroup;ide_hwif_t *match =NULL ;/* Allocate the buffer and potentially sleep first */ new_hwgroup =kmalloc (sizeof (ide_hwgroup_t),GFP_KERNEL ); - -spin_lock_irqsave (&ide_lock, flags); +BUG_ON (in_interrupt ()); +BUG_ON (irqs_disabled ()); +down (&ide_cfg_sem); hwif->hwgroup =NULL ;#if MAX_HWIFS >1 /*
@@ -1078,7 +1079,7 @@ }else { hwgroup = new_hwgroup;if (!hwgroup) { -spin_unlock_irqrestore (&ide_lock, flags); +up (&ide_cfg_sem);return 1 ; }memset (hwgroup,0 ,sizeof (ide_hwgroup_t));
@@ -1095,6 +1096,7 @@/* * Allocate the irq, if not already obtained for another hwif */ + do_enable_irq =0 ;if (!match || match->irq != hwif->irq) {int sa =SA_INTERRUPT ;#if defined(__mc68000__) || defined(CONFIG_APUS )
@@ -1112,17 +1114,23 @@/* clear nIEN */ hwif->OUTB (0x08 , hwif->io_ports[IDE_CONTROL_OFFSET ]); +disable_irq (hwif->irq);if (request_irq (hwif->irq,&ide_intr,sa,hwif->name,hwgroup)) {if (!match)kfree (hwgroup); -spin_unlock_irqrestore (&ide_lock, flags); +enable_irq (&hwif->irq); +up (&ide_cfg_sem);return 1 ; } + do_enable_irq =1 ; }/* * Everything is okay, so link us into the hwgroup */ +spin_lock_irq (&ide_lock); +if (do_enable_irq) +enable_irq (hwif->irq); hwif->hwgroup = hwgroup; hwif->next = hwgroup->hwif->next; hwgroup->hwif->next = hwif;
@@ -1135,9 +1143,9 @@ hwgroup->drive = drive; drive->next = hwgroup->drive->next; hwgroup->drive->next = drive; -spin_unlock_irqrestore (&ide_lock, flags); +spin_unlock_irq (&ide_lock);ide_init_queue (drive); -spin_lock_irqsave (&ide_lock, flags); +spin_lock_irq (&ide_lock); }if (!hwgroup->hwif) {
@@ -1148,7 +1156,7 @@ }/* all CPUs; safe now that hwif->hwgroup is set up */ -spin_unlock_irqrestore (&ide_lock, flags); +spin_unlock_irq (&ide_lock);#if !defined(__mc68000__) && !defined(CONFIG_APUS ) && !defined(__sparc__) printk("%s at 0x%03lx-0x%03lx,0x%03lx on irq %d", hwif->name,
@@ -1168,6 +1176,7 @@printk (" (%sed with %s)" , hwif->sharing_irq ?"shar" :"serializ" , match->name);printk ("\n" ); +up (&ide_cfg_sem);return 0 ; }
drivers/ide/ide.c +9 -5
@@ -177,6 +177,7 @@static int system_bus_speed;/* holds what we think is VESA/PCI bus speed */ static int initializing;/* set while initializing built-in drivers */ +DECLARE_MUTEX (ide_cfg_sem); spinlock_t ide_lock __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;static int ide_scan_direction;/* THIS was formerly 2.2.x pci=reverse */
@@ -584,13 +585,15 @@ide_hwif_t *hwif, *g;ide_hwgroup_t *hwgroup;int irq_count =0 , unit, i; -unsigned long flags;ide_hwif_t old_hwif;if (index >=MAX_HWIFS )BUG (); -spin_lock_irqsave (&ide_lock, flags); +BUG_ON (in_interrupt ()); +BUG_ON (irqs_disabled ()); +down (&ide_cfg_sem); +spin_lock_irq (&ide_lock); hwif = &ide_hwifs[index];if (!hwif->present) goto abort;
@@ -605,7 +608,7 @@ } hwif->present =0 ; -spin_unlock_irqrestore (&ide_lock, flags); +spin_unlock_irq (&ide_lock);for (unit =0 ; unit <MAX_DRIVES ; ++unit) { drive = &hwif->drives[unit];
@@ -618,7 +621,6 @@#ifdef CONFIG_PROC_FS destroy_proc_ide_drives (hwif);#endif -spin_lock_irqsave (&ide_lock, flags); hwgroup = hwif->hwgroup;/*
@@ -633,6 +635,7 @@if (irq_count ==1 )free_irq (hwif->irq, hwgroup); +spin_lock_irq (&ide_lock);/* * Note that we only release the standard ports, * and do not even try to handle any extra ports
@@ -813,7 +816,8 @@ hwif->hwif_data = old_hwif.hwif_data; abort: -spin_unlock_irqrestore (&ide_lock, flags); +spin_unlock_irq (&ide_lock); +up (&ide_cfg_sem); }EXPORT_SYMBOL (ide_unregister);
drivers/ide/ide-io.c +2 -2
@@ -745,8 +745,8 @@/* for atari only: POSSIBLY BROKEN HERE(?) */ ide_get_lock (ide_intr, hwgroup); -/* necessary paranoia: ensure IRQs are masked on local CPU */ -local_irq_disable (); +/* caller must own ide_lock */ +BUG_ON (!irqs_disabled ());while (!hwgroup->busy) { hwgroup->busy =1 ;
Quoted 4 lines · Manfred Spraul
sleeping - e.g. on a NUMA arch an IPI to the right node could be necessary. What about fixing ide? If ide can't handle early interrupts, then it can use
disable_irq();
disable_irq(n) isn't something all platforms seem to define before the IRQ is requested, In addition disable_irq() is not supported on some systems the IDE code supports. I have to have local CPU interrupts disabled on the CPU at the point the IRQ becomes live.
The former we could probably solve with "SA_DISABLE" to request an IRQ that is disabled, but I'm not sure how we deal with platforms which plain and simply do not have 'mask irq' functionality but are using a level triggered interrupt.