libata: fix locking for kmap_atomic

Change locking surrounding kmap_atomic from local_irqsave to local_irqsave_nort. This fixes issues with PREEMPT_RT.

--- linux-2.6.25.10/drivers/ata/libata-core.c 2008-07-03 05:46:47.000000000
+0200
+++ linux-2.6.25.10-rt7/drivers/ata/libata-core.c 2008-07-12
23:59:33.132140258 +0200
@@ -5157,14 +5157,14 @@
unsigned long flags;

/* FIXME: use a bounce buffer */
- local_irq_save(flags);
+ local_irq_save_nort(flags);
buf = kmap_atomic(page, KM_IRQ0);

/* do the actual data transfer */ ap->ops->data_xfer(qc->dev, buf + offset, qc->sect_size, do_write);

kunmap_atomic(buf, KM_IRQ0);
- local_irq_restore(flags);
+ local_irq_restore_nort(flags);
} else {
buf = page_address(page);
ap->ops->data_xfer(qc->dev, buf + offset, qc->sect_size, do_write);
@@ -5294,14 +5294,14 @@
unsigned long flags;

/* FIXME: use bounce buffer */
- local_irq_save(flags);
+ local_irq_save_nort(flags);
buf = kmap_atomic(page, KM_IRQ0);

/* do the actual data transfer */
consumed = ap->ops->data_xfer(dev, buf + offset, count, rw);

kunmap_atomic(buf, KM_IRQ0);
- local_irq_restore(flags);
+ local_irq_restore_nort(flags);
} else {
buf = page_address(page);
consumed = ap->ops->data_xfer(dev, buf + offset, count, rw);

I assume this is a -rt kernel patch.

Sunday, 13 of July 2008 10:26:30 Andrew Morton napisał(a):

Sorry, this was for -rt only.

[   17.012011] BUG: sleeping function called from invalid context IRQ-14(5732) 
at arch/x86/mm/highmem_32.c:8
[   17.012011] in_atomic():0 [00000000], irqs_disabled():1
[   17.012011] Pid: 5732, comm: IRQ-14 Not tainted 2.6.25.10-rtXXX #11
[   17.012011]  [<c0120fc4>] __might_sleep+0xf1/0xf8
[   17.012011]  [<c011c035>] kmap+0x47/0x5a
[   17.012011]  [<c033e728>] ata_hsm_move+0x3d7/0x657
[   17.012011]  [<c0342fd6>] ata_interrupt+0x14e/0x1cb
[   17.012011]  [<c015c72c>] handle_IRQ_event+0x4e/0xd1
[   17.012011]  [<c015d266>] do_irqd+0x126/0x224
[   17.012011]  [<c015d140>] ? do_irqd+0x0/0x224
[   17.012011]  [<c013ae9d>] kthread+0x3b/0x62
[   17.012011]  [<c013ae62>] ? kthread+0x0/0x62
[   17.012011]  [<c0108287>] kernel_thread_helper+0x7/0x10
[   17.012011]  =======================

The subject says kmap_atomic, but this is kmap. It definitely makes no sense to call kmap in an IRQ, regardless of the locking. There seems to be a larger structural problem here.

J

Sunday, 13 of July 2008 19:46:29 Jeremy Fitzhardinge napisał(a):

--- linux-2.6.25.8-rt7.orig/include/asm-x86/highmem.h 2008-06-23
19:12:47.000000000 -0400
+++ linux-2.6.25.8-rt7/include/asm-x86/highmem.h 2008-06-23
19:13:58.000000000 -0400

+/*
+ * on PREEMPT_RT kmap_atomic() is a wrapper that uses kmap():
+ */
+#ifdef CONFIG_PREEMPT_RT
+# define kmap_atomic_prot(page, type, prot) ({ pagefault_disable();
kmap(page); })

+# define kmap_atomic(page, type)       ({ pagefault_disable(); kmap(page); })
+# define kmap_atomic_pfn(pfn, type)    kmap(pfn_to_page(pfn))
+# define kunmap_atomic(kvaddr, type)   do { pagefault_enable(); 

kunmap_virt(kvaddr); } while(0)
+# define kmap_atomic_to_page(kvaddr) kmap_to_page(kvaddr)
+#else
+# define kmap_atomic_prot(page, type, prot) __kmap_atomic_prot(page, type,
prot)

+# define kmap_atomic(page, type)       __kmap_atomic(page, type)
+# define kmap_atomic_pfn(pfn, type)    __kmap_atomic_pfn(pfn, type)
+# define kunmap_atomic(kvaddr, type)   __kunmap_atomic(kvaddr, type)
+# define kmap_atomic_to_page(kvaddr)   __kmap_atomic_to_page(kvaddr)

+#endif
+

OK, I guess the "structural problem" is RT ;)

J

The right way to fix this is to use a bounce buffer. I sent Jeff patches to do that a couple of times. I'm not sure what has happened to them since ? So NAK this, we should apply the patch to actually fix the bug instead.

Jeff ?

Alan

Sunday, 13 of July 2008 12:25:43 Alan Cox napisał(a):

Sent patch was for:

[   16.787021] BUG: sleeping function called from invalid context IRQ-14(5730) 
at arch/x86/mm/highmem_32.c:8
[   16.787021] in_atomic():0 [00000000], irqs_disabled():1
[   16.787021] Pid: 5730, comm: IRQ-14 Not tainted 2.6.25.10-rtXXX #12
[   16.787021]  [<c0120fc4>] __might_sleep+0xf1/0xf8
[   16.787021]  [<c011c035>] kmap+0x47/0x5a
[   16.787021]  [<c03442b2>] ata_scsi_rbuf_get+0x23/0x31
[   16.787021]  [<c0345975>] atapi_qc_complete+0x252/0x2c5
[   16.787021]  [<c012edee>] ? irq_exit+0x7b/0x7e
[   16.787021]  [<c0109727>] ? do_IRQ+0x6e/0x87
[   16.787021]  [<c033d041>] __ata_qc_complete+0x8e/0x93
[   16.787021]  [<c033ddfb>] ata_qc_complete+0x182/0x188
[   16.787021]  [<c033e188>] ata_hsm_qc_complete+0xa1/0xa5
[   16.787021]  [<c033e8fd>] ata_hsm_move+0x5dd/0x628
[   16.787021]  [<c0342f76>] ata_interrupt+0x14e/0x1cb
[   16.787021]  [<c015c72c>] handle_IRQ_event+0x4e/0xd1
[   16.787021]  [<c015d266>] do_irqd+0x126/0x224
[   16.787021]  [<c015d140>] ? do_irqd+0x0/0x224
[   16.787021]  [<c013ae9d>] kthread+0x3b/0x62
[   16.787021]  [<c013ae62>] ? kthread+0x0/0x62
[   16.787021]  [<c0108287>] kernel_thread_helper+0x7/0x10
[   16.787021]  =======================

Fine - but the right fix is still to apply the patches that use a bounce buffer, otherwise you might paper over the warning but you'll trash rt performance. It needs fixing anyway especially with Tejun's changes to the ATAPI logic to use PIO more often

Alan