On Tuesday August 22, dxm@clouds.melbourne.sgi.com wrote:
>
> System: SGI 1400 SMP w/ sym53c8xx SCSI.
> Kernel: 2.4.0-test5
>
> I've been hitting a SMP deadlock caused by creating traffic
> on a ramdisk at the same time as syncing. The backtraces show
> the locks causing the deadlock. I don't know how to fix this
> one myself, so here's the info in case it helps...
>
> CPU A
>
> insert_into_queues <takes lru_list_lock>
> getblk
> rd_request
> generic_unplug_device <takes io_request_lock>
> __wait_on_buffer
> unmap_buffer
> block_flushpage
> truncate_inode_pages
> vmtruncate
> inode_setattr
> notify_change
> do_truncate
> open_namei
> filp_open
> sys_open
> system_call
>
> CPU B
>
> sym53c8xx_intr <takes io_request_lock>
> handle_IRQ_event
> <scsi interrupt here>
> sync_buffers <takes lru_list_lock>
> fsync_dev
> sys_sync
> system_call
>
> Regards,
My feeling is that the "wrong" that is happening is that getblk is
being called inside the io_request_lock.
This can be fixed quite easily by changing rd to not use a request
queue at all, but passing each buffer directly from ll_rw_block
through generic_make_request into rd.c, instead of queuing them up in
a request queue. After all, rd does *not* need an elevator algorithm.
The following patch should do this, but is totally untested. I don't
even know that it compiles.
Let me know how it goes if you try it.
NeilBrown
--- drivers/block/rd.c 2000/08/22 11:58:30 1.1
+++ drivers/block/rd.c 2000/08/22 12:06:49
@@ -194,50 +194,48 @@
* 19-JAN-1998 Richard Gooch <rgooch@atnf.csiro.au> Added devfs support
*
*/
-static void rd_request(request_queue_t * q)
+static int rd_make_request(request_queue_t * q, int rw, struct buffer_head *sbh)
{
unsigned int minor;
unsigned long offset, len;
struct buffer_head *rbh;
- struct buffer_head *sbh;
-repeat:
- INIT_REQUEST;
- minor = MINOR(CURRENT->rq_dev);
+ minor = MINOR(sbh->b_rdev);
+
+ if (minor >= NUM_RAMDISKS)
+ goto fail;
- if (minor >= NUM_RAMDISKS) {
- end_request(0);
- goto repeat;
- }
- offset = CURRENT->sector << 9;
- len = CURRENT->current_nr_sectors << 9;
+ offset = sbh->b_rsector << 9;
+ len = sbh->b_size;
- if ((offset + len) > rd_length[minor]) {
- end_request(0);
- goto repeat;
- }
+ if ((offset + len) > rd_length[minor])
+ goto fail;
- if ((CURRENT->cmd != READ) && (CURRENT->cmd != WRITE)) {
- printk(KERN_INFO "RAMDISK: bad command: %d\n", CURRENT->cmd);
- end_request(0);
- goto repeat;
- }
-
- sbh = CURRENT->bh;
- rbh = getblk(sbh->b_dev, sbh->b_blocknr, sbh->b_size);
- if (CURRENT->cmd == READ) {
+ if (rw==READA)
+ rw=READ;
+ if ((rw != READ) && (rw != WRITE)) {
+ printk(KERN_INFO "RAMDISK: bad command: %d\n", rw);
+ goto fail;
+ }
+
+ sbh = bh;
+ rbh = getblk(sbh->b_rdev, sbh->b_rsector*(sbh->b_size>>9), sbh->b_size);
+ if (rw == READ) {
if (sbh != rbh)
- memcpy(CURRENT->buffer, rbh->b_data, rbh->b_size);
+ memcpy(sbh->b_data, rbh->b_data, rbh->b_size);
} else
if (sbh != rbh)
- memcpy(rbh->b_data, CURRENT->buffer, rbh->b_size);
+ memcpy(rbh->b_data, sbh->b_data, rbh->b_size);
mark_buffer_protected(rbh);
brelse(rbh);
- end_request(1);
- goto repeat;
+ sbh->b_end_io(sbh,1);
+ return 0;
+ fail:
+ sbh->b_end_io(sbh,0);
+ return 0;
}
static int rd_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
@@ -378,7 +376,6 @@
devfs_unregister (devfs_handle);
unregister_blkdev( MAJOR_NR, "ramdisk" );
- blk_cleanup_queue(BLK_DEFAULT_QUEUE(MAJOR_NR));
hardsect_size[MAJOR_NR] = NULL;
blksize_size[MAJOR_NR] = NULL;
blk_size[MAJOR_NR] = NULL;
@@ -403,7 +400,7 @@
return -EIO;
}
- blk_init_queue(BLK_DEFAULT_QUEUE(MAJOR_NR), &rd_request);
+ blk_queue_make_request(BLK_DEFAULT_QUEUE(MAJOR_NR), &rd_make_request);
for (i = 0; i < NUM_RAMDISKS; i++) {
/* rd_size is given in kB */
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Wed Aug 23 2000 - 21:00:06 EST