Re: 2.6.20-rc6-mm1

[Andrew Morton <akpm@osdl.org>]

On Mon, 29 Jan 2007 23:27:27 -0800
Andrew Morton <akpm@osdl.org> wrote:

> On Sun, 28 Jan 2007 11:25:42 +0100
> Jiri Slaby <jirislaby@gmail.com> wrote:
> 
> > Andrew Morton napsal(a):
> > > Temporarily at
> > > 
> > > 	http://userweb.kernel.org/~akpm/2.6.20-rc6-mm1/
> > 
> > I'm still seeing this during bootup:
> > BUG: at /home/l/latest/xxx/arch/i386/mm/highmem.c:52 kmap_atomic()
> >   [<c0104ffb>] show_trace_log_lvl+0x1a/0x30
> >   [<c01056b5>] show_trace+0x12/0x14
> >   [<c010573c>] dump_stack+0x16/0x18
> >   [<c011b24f>] kmap_atomic+0x16c/0x20e
> >   [<c01b279e>] ntfs_end_buffer_async_read+0x18e/0x2ed
> >   [<c0183352>] end_bio_bh_io_sync+0x26/0x3f
> >   [<c0184dd4>] bio_endio+0x37/0x62
> >   [<c01cda43>] __end_that_request_first+0x224/0x444
> >   [<c01cdc6b>] end_that_request_chunk+0x8/0xa
> >   [<c024507d>] scsi_end_request+0x1f/0xc7
> >   [<c02451e6>] scsi_io_completion+0x7b/0x33a
> >   [<c024aa5d>] sd_rw_intr+0x23/0x1ab
> >   [<c02415ed>] scsi_finish_command+0x42/0x47
> >   [<c0245a38>] scsi_softirq_done+0x64/0xcf
> >   [<c01cf9e6>] blk_done_softirq+0x54/0x62
> >   [<c0127bc5>] __do_softirq+0x75/0xde
> >   [<c0127c69>] do_softirq+0x3b/0x3d
> >   [<c0127ee6>] irq_exit+0x3b/0x3d
> >   [<c0106804>] do_IRQ+0x51/0x8d
> >   [<c0104a5f>] common_interrupt+0x23/0x28
> >   [<c010241a>] cpu_idle+0x80/0xc3
> >   [<c010140d>] rest_init+0x23/0x36
> >   [<c045ea59>] start_kernel+0x3a5/0x43c
> >   [<00000000>] 0x0
> >   =======================
> > 
> > I.e. KM_BIO_SRC_IRQ through softirq path.
> > 
> 
> argh.
> 
> ntfs_end_buffer_async_read() doesn't know whether it will be called from
> hardirq or from softirq context: it depends upon the underlying driver.
> 
> In this case, if the CPU running ntfs_end_buffer_async_read() is
> interrupted by IO completion against a different disk controller and that
> completion handler uses KM_BIO_SRC_IRQ (as it is allowed to do), it will
> trash ntfs_end_buffer_async_read()'s atomic kmap and unpleasing things will
> ensue.
> 
> I guess a suitable fix here is to protect that kmap with
> local_irq_save/restore.
> 
> I wonder where else we have that bug?

Actually, this isn't related to softirq-vs-hardirq.  Most interrupt
handlers are interruptible, so the rule is simply that KM_BIO_SRC_IRQ must
always be taken under local_irq_disable().

A quick scan indicates that the following files might be buggy in this
regard:

drivers/mmc/wbsd.c
drivers/mmc/at91_mci.c
drivers/mmc/sdhci.c
drivers/scsi/scsi_lib.c when called from stex.c
fs/ntfs/aops.c

Happily, KM_BIO_DST_IRQ has no users and can presumably be removed.


Fixes for stex and ntfs follow.


From: Andrew Morton <akpm@osdl.org>

The KM_BIO_SRC_IRQ kmap slot requires local irq protection.

Cc: James Bottomley <James.Bottomley@steeleye.com>
Cc: Ed Lin <ed.lin@promise.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
---

 drivers/scsi/stex.c |    8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff -puN drivers/scsi/stex.c~stex-kmap_atomic-atomicity-fix drivers/scsi/stex.c
--- a/drivers/scsi/stex.c~stex-kmap_atomic-atomicity-fix
+++ a/drivers/scsi/stex.c
@@ -459,15 +459,19 @@ static void stex_internal_copy(struct sc
 		*count = cmd->request_bufflen;
 	lcount = *count;
 	while (lcount) {
+		unsigned long flags = flags;	/* Suppress uninit warning */
+
 		len = lcount;
 		s = (void *)src;
 		if (cmd->use_sg) {
 			size_t offset = *count - lcount;
 			s += offset;
+			local_irq_save(flags);
 			base = scsi_kmap_atomic_sg(cmd->request_buffer,
 				sg_count, &offset, &len);
 			if (base == NULL) {
 				*count -= lcount;
+				local_irq_restore(flags);
 				return;
 			}
 			d = base + offset;
@@ -480,8 +484,10 @@ static void stex_internal_copy(struct sc
 			memcpy(s, d, len);
 
 		lcount -= len;
-		if (cmd->use_sg)
+		if (cmd->use_sg) {
 			scsi_kunmap_atomic_sg(base);
+			local_irq_restore(flags);
+		}
 	}
 }
 
_



From: Andrew Morton <akpm@osdl.org>

The KM_BIO_SRC_IRQ kmap slot requires local irq protection.

Cc: Anton Altaparmakov <aia21@cantab.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
---

 fs/ntfs/aops.c |    6 ++++++
 1 file changed, 6 insertions(+)

diff -puN fs/ntfs/aops.c~ntfs-kmap_atomic-atomicity-fix fs/ntfs/aops.c
--- a/fs/ntfs/aops.c~ntfs-kmap_atomic-atomicity-fix
+++ a/fs/ntfs/aops.c
@@ -88,14 +88,17 @@ static void ntfs_end_buffer_async_read(s
 		if (unlikely(file_ofs + bh->b_size > init_size)) {
 			u8 *kaddr;
 			int ofs;
+			unsigned long flags;
 
 			ofs = 0;
 			if (file_ofs < init_size)
 				ofs = init_size - file_ofs;
+			local_irq_save(flags);
 			kaddr = kmap_atomic(page, KM_BIO_SRC_IRQ);
 			memset(kaddr + bh_offset(bh) + ofs, 0,
 					bh->b_size - ofs);
 			kunmap_atomic(kaddr, KM_BIO_SRC_IRQ);
+			local_irq_restore(flags);
 			flush_dcache_page(page);
 		}
 	} else {
@@ -138,16 +141,19 @@ static void ntfs_end_buffer_async_read(s
 		u8 *kaddr;
 		unsigned int i, recs;
 		u32 rec_size;
+		unsigned long flags;
 
 		rec_size = ni->itype.index.block_size;
 		recs = PAGE_CACHE_SIZE / rec_size;
 		/* Should have been verified before we got here... */
 		BUG_ON(!recs);
+		local_irq_save(flags);
 		kaddr = kmap_atomic(page, KM_BIO_SRC_IRQ);
 		for (i = 0; i < recs; i++)
 			post_read_mst_fixup((NTFS_RECORD*)(kaddr +
 					i * rec_size), rec_size);
 		kunmap_atomic(kaddr, KM_BIO_SRC_IRQ);
+		local_irq_restore(flags);
 		flush_dcache_page(page);
 		if (likely(page_uptodate && !PageError(page)))
 			SetPageUptodate(page);
_