Skip to contentSkip to conversation

sd: fix cache flushing on module removal (and individual device removal)

PATCH In review James Bottomley 2 participants 4 messages Aug 31, 2006 – Sep 1, 2006 SCSI

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sd: fix cache flushing on module removal (and individual device removal)

The fix isn't actually in sd: it's in scsi_device_get(). I modified it to allow devices to be returned in SDEV_CANCEL, but not SDEV_DEL. This means that the device_remove_driver, which occurs in device_del() in scsi_remove_device() after the device has gone into SDEV_CANCEL is now effective at flushing the cache.

James

drivers/scsi/scsi.c +1 -1
index 94df671..3294863 100644
@@ -851,7 +851,7 @@ EXPORT_SYMBOL(scsi_track_queue_full);
  */
 int scsi_device_get(struct scsi_device *sdev)
 {
-	if (sdev->sdev_state == SDEV_DEL || sdev->sdev_state == SDEV_CANCEL)
+	if (sdev->sdev_state == SDEV_DEL)
 		return -ENXIO;
 	if (!get_device(&sdev->sdev_gendev))
 		return -ENXIO;
#1 lore raw
Quoted 5 lines · James Bottomley

The fix isn't actually in sd: it's in scsi_device_get(). I modified it to allow devices to be returned in SDEV_CANCEL, but not SDEV_DEL. This means that the device_remove_driver, which occurs in device_del() in scsi_remove_device() after the device has gone into SDEV_CANCEL is now effective at flushing the cache.

Actually, lets try that again, but with the correct patch.

James

drivers/scsi/scsi.c +9 -6
Index: linux-2.6/drivers/scsi/scsi.c
===================================================================
@@ -835,14 +835,14 @@ EXPORT_SYMBOL(scsi_track_queue_full);
  */
 int scsi_device_get(struct scsi_device *sdev)
 {
-	if (sdev->sdev_state == SDEV_DEL || sdev->sdev_state == SDEV_CANCEL)
+	if (sdev->sdev_state == SDEV_DEL)
 		return -ENXIO;
 	if (!get_device(&sdev->sdev_gendev))
 		return -ENXIO;
-	if (!try_module_get(sdev->host->hostt->module)) {
-		put_device(&sdev->sdev_gendev);
-		return -ENXIO;
-	}
+	/* We can fail this if we're doing SCSI operations
+	 * from module exit (like cache flush) */
+	try_module_get(sdev->host->hostt->module);
+
 	return 0;
 }
 EXPORT_SYMBOL(scsi_device_get);
@@ -857,7 +857,10 @@ EXPORT_SYMBOL(scsi_device_get);
  */
 void scsi_device_put(struct scsi_device *sdev)
 {
-	module_put(sdev->host->hostt->module);
+	/* The module refcount will be zero if scsi_device_get()
+	 * was called from a module removal routine */
+	if (likely(module_refcount(sdev->host->hostt->module) != 0))
+		module_put(sdev->host->hostt->module);
 	put_device(&sdev->sdev_gendev);
 }
 EXPORT_SYMBOL(scsi_device_put);
#2 lore raw

James Bottomley wrote:
[...]

Quoted 24 lines
drivers/scsi/scsi.c +5 -5
@@ -835,14 +835,14 @@ EXPORT_SYMBOL(scsi_track_queue_full);
  */
 int scsi_device_get(struct scsi_device *sdev)
 {
-	if (sdev->sdev_state == SDEV_DEL || sdev->sdev_state == SDEV_CANCEL)
+	if (sdev->sdev_state == SDEV_DEL)
 		return -ENXIO;
 	if (!get_device(&sdev->sdev_gendev))
 		return -ENXIO;
-	if (!try_module_get(sdev->host->hostt->module)) {
-		put_device(&sdev->sdev_gendev);
-		return -ENXIO;
-	}
+	/* We can fail this if we're doing SCSI operations
+	 * from module exit (like cache flush) */
+	try_module_get(sdev->host->hostt->module);
+
 	return 0;
 }
 EXPORT_SYMBOL(scsi_device_get);
@@ -857,7 +857,10 @@ EXPORT_SYMBOL(scsi_device_get);
  */

drivers/scsi/scsi.c EXPORT_SYMBOL(scsi_device_get); · line 866

Quoted 10 lines
drivers/scsi/scsi.c +4 -1
 void scsi_device_put(struct scsi_device *sdev)
 {
-	module_put(sdev->host->hostt->module);
+	/* The module refcount will be zero if scsi_device_get()
+	 * was called from a module removal routine */
+	if (likely(module_refcount(sdev->host->hostt->module) != 0))
+		module_put(sdev->host->hostt->module);
 	put_device(&sdev->sdev_gendev);
 }
 EXPORT_SYMBOL(scsi_device_put);

Somehow the (void)try_module_get(...) looks dangerous to me. Is it really safe to always ignore failures to get the module? Why would we want to ignore failures? Couldn't there be border cases where a module_getter/_putter in a concurrent code path disturbs scsi_device_get/_put's underlying assumptions?

Signature · 3 lines
Stefan Richter
-=====-=-==- =--= ----=
http://arcgraph.de/sr/
#3 lore raw
Quoted 5 lines · Stefan Richter

Somehow the (void)try_module_get(...) looks dangerous to me. Is it really safe to always ignore failures to get the module? Why would we want to ignore failures? Couldn't there be border cases where a module_getter/_putter in a concurrent code path disturbs scsi_device_get/_put's underlying assumptions?

As long as we don't do spurious module_puts, yes. However, there looks to be another nasty module race that's orthogonal to this, in that the final scsi_host_put() of a module doesn't necessarily wait for the host actually to be released, so it's possible to free the host template when the module exit finishes and still have a partially functioning host.

James