2008-07-08 20:10:39

by San Mehat

Subject: [PATCH 001/001] mmcblk: Fix deadlock when cleanup is called from resume

When CONFIG_UNSAFE_RESUME is used, we may need to cleanup
(and call kthread_stop) when threads are still frozen. To fix this
deadlock, push cleanup off to a work-queue.

Signed-off-by: San Mehat <[email protected]>

---
drivers/mmc/card/queue.c | 36 ++++++++++++++++++++++++++++++++++--
1 files changed, 34 insertions(+), 2 deletions(-)

diff --git a/drivers/mmc/card/queue.c b/drivers/mmc/card/queue.c
index 972738f..908e3cf 100644
--- a/drivers/mmc/card/queue.c
+++ b/drivers/mmc/card/queue.c
@@ -14,6 +14,7 @@
#include <linux/freezer.h>
#include <linux/kthread.h>
#include <linux/scatterlist.h>
+#include <linux/workqueue.h>

#include <linux/mmc/card.h>
#include <linux/mmc/host.h>
@@ -23,6 +24,11 @@

#define MMC_QUEUE_SUSPENDED (1 << 0)

+static void mq_cleanup_work(struct work_struct *work);
+static DECLARE_WORK(mq_cleanup_worker, mq_cleanup_work);
+static struct mmc_queue *cleanup_mq;
+static DEFINE_SPINLOCK(mq_cleanup_lock);
+
/*
* Prepare a MMC request. This just filters out odd stuff.
*/
@@ -217,9 +223,36 @@ int mmc_init_queue(struct mmc_queue *mq, struct mmc_card *card, spinlock_t *lock

void mmc_cleanup_queue(struct mmc_queue *mq)
{
- struct request_queue *q = mq->queue;
unsigned long flags;

+ BUG_ON(!mq);
+
+ /*
+ * The actual cleanup is done from a work_queue because if you
+ * thread_stop() while resuming, you'll deadlock.
+ */
+ spin_lock_irqsave(&mq_cleanup_lock, flags);
+ cleanup_mq = mq;
+ schedule_work(&mq_cleanup_worker);
+ spin_unlock_irqrestore(&mq_cleanup_lock, flags);
+}
+EXPORT_SYMBOL(mmc_cleanup_queue);
+
+static void mq_cleanup_work(struct work_struct *work)
+{
+ struct mmc_queue *mq;
+ unsigned long flags;
+ struct request_queue *q;
+
+ spin_lock_irqsave(&mq_cleanup_lock, flags);
+ mq = cleanup_mq;
+ cleanup_mq = NULL;
+ spin_unlock_irqrestore(&mq_cleanup_lock, flags);
+
+ BUG_ON(!mq);
+
+ q = mq->queue;
+
/* Mark that we should start throwing out stragglers */
spin_lock_irqsave(q->queue_lock, flags);
q->queuedata = NULL;
@@ -246,7 +279,6 @@ void mmc_cleanup_queue(struct mmc_queue *mq)

mq->card = NULL;
}
-EXPORT_SYMBOL(mmc_cleanup_queue);

/**
* mmc_queue_suspend - suspend a MMC request queue

2008-07-09 04:58:29

by Pierre Ossman

Subject: Re: [PATCH 001/001] mmcblk: Fix deadlock when cleanup is called from resume

On Tue, 8 Jul 2008 13:10:11 -0700
[email protected] (San Mehat) wrote:

>
> When CONFIG_UNSAFE_RESUME is used, we may need to cleanup
> (and call kthread_stop) when threads are still frozen. To fix this
> deadlock, push cleanup off to a work-queue.
>
> Signed-off-by: San Mehat <[email protected]>
>

I appreciate patches, but sending me something based on 2.6.21 is
hardly helpful. This particular problem was solved over a year ago in
mainline.

Rgds
--
-- Pierre Ossman

Linux kernel, MMC maintainer http://www.kernel.org
rdesktop, core developer http://www.rdesktop.org

WARNING: This correspondence is being monitored by the
Swedish government. Make sure your server uses encryption
for SMTP traffic and consider using PGP for end-to-end
encryption.