bsg referencing bus driver module

Commit Message

Anatoliy Glagolev April 20, 2018, 10:44 p.m. UTC

> This patch isn't applyable because your mailer has changed all the tabs
> to spaces.
> 
> I also think there's no need to do it this way.  I think what we need
> is for fc_bsg_remove() to wait until the bsg queue is drained.  It does
> look like the author thought this happened otherwise the code wouldn't
> have the note.  If we fix it that way we can do the same thing in all
> the other transport classes that use bsg (which all have a similar
> issue).
> 
> James
> 

Thanks, James. Sorry about the tabs; re-sending.

On fc_bsg_remove()...: are you suggesting to implement the whole fix
in scsi_transport_fc.c? That would be nice, but I do not see how that
is possible. Even with the queue drained bsg still holds a reference
to the Scsi_Host via bsg_class_device; bsg_class_device itself is
referenced on bsg_open and kept around while a user-mode process keeps
a handle to bsg.
Even if we somehow implement the waiting the call may be stuck
forever if the user-mode process keeps the handle.
I think handling it via a rererence to the module is more consistent
with the way things are done in Linux. You suggested the approach
youself back in "Waiting for scsi_host_template release" discussion.


From df939b80d02bf37b21efaaef8ede86cfd39b0cb8 Mon Sep 17 00:00:00 2001
From: Anatoliy Glagolev <[e-mail]>
Date: Thu, 19 Apr 2018 15:06:06 -0600
Subject: [PATCH] bsg referencing parent module

Fixing a bug when bsg layer holds the last reference to device
when the device's module has been unloaded. Upon dropping the
reference the device's release function may touch memory of
the unloaded module.
---
 block/bsg-lib.c                  | 24 ++++++++++++++++++++++--
 block/bsg.c                      | 29 ++++++++++++++++++++++++++---
 drivers/scsi/scsi_transport_fc.c |  8 ++++++--
 include/linux/bsg-lib.h          |  4 ++++
 include/linux/bsg.h              |  5 +++++
 5 files changed, 63 insertions(+), 7 deletions(-)

Comments

James Bottomley April 22, 2018, 7:47 a.m. UTC | #1

On Fri, 2018-04-20 at 16:44 -0600, Anatoliy Glagolev wrote:
>  
> > This patch isn't applyable because your mailer has changed all the
> > tabs to spaces.
> > 
> > I also think there's no need to do it this way.  I think what we
> > need is for fc_bsg_remove() to wait until the bsg queue is
> > drained.  It does look like the author thought this happened
> > otherwise the code wouldn't have the note.  If we fix it that way
> > we can do the same thing in all the other transport classes that
> > use bsg (which all have a similar issue).
> > 
> > James
> > 
> 
> Thanks, James. Sorry about the tabs; re-sending.
> 
> On fc_bsg_remove()...: are you suggesting to implement the whole fix
> in scsi_transport_fc.c?

Yes, but it's not just scsi_transport_fc, scsi_transport_sas has the
same issue.  I think it's probably just the one liner addition of
blk_drain_queue() that fixes this.  There should probably be a block
primitive that does the correct queue reference dance and calls
blk_cleanup_queue() and blk_drain_queue() in order.

>  That would be nice, but I do not see how that
> is possible. Even with the queue drained bsg still holds a reference
> to the Scsi_Host via bsg_class_device; bsg_class_device itself is
> referenced on bsg_open and kept around while a user-mode process
> keeps a handle to bsg.

Once you've called bsg_unregister_queue(), the queue will be destroyed
and the reference released once the last job is drained, meaning the
user can keep the bsg device open, but it will just return errors
because of the lack of queue.  This scenario allows removal to proceed
without being held hostage by open devices.

> Even if we somehow implement the waiting the call may be stuck
> forever if the user-mode process keeps the handle.

No it won't: after blk_cleanup_queue(), the queue is in bypass mode: no
requests queued after this do anything other than complete with error,
so they never make it into SCSI.

> I think handling it via a rererence to the module is more consistent
> with the way things are done in Linux. You suggested the approach
> youself back in "Waiting for scsi_host_template release" discussion.

That was before I analyzed the code paths.  Module release is tricky,
because the module exit won't be called until the references drop to
zero, so you have to be careful about not creating a situation where
module exit never gets called and module exit code should force stuff
to detach and wait for the forcing to complete to make up for the
reference circularity problem.  If you do it purely by refcounting, the
module actually may never release (that's why scsi_remove_host works
the way it does, for instance).

James

Anatoliy Glagolev April 23, 2018, 6:38 p.m. UTC | #2

Thanks, James. The idea of cutting communications with Scsi_Host at
bsg_unregister_queue(..) time and leaving bsg_class_device to
its own fate makes a lot of sense, conceptually. But there are
implementation issues that are difficult to work around.

bsg.c creates bsg_class_device and takes a reference to Scsi_Host
at bsg_register_queue(..) time. The reference is dropped at
bsg_class_device's release(..) function. If the driver implementing
Scsi_Host template is not around we crash.
We could move the reference drop from bsg_class_device's release(..)
function to bsg_unregister_queue(..). That would be a small change in
bsg.c. But bsg.c sets Scsi_Host as the parent of bsg_class_device's
device. We cannot have a device around with a dangling parent.
A device's parent cannot be changed dynamically. Not setting
the device's parent at creation may affect software relying
on bsg_class_device - Scsi_Host child-parent relations.

It looks like I am out of options. Do you have suggestions on
how to work around Scsi_Host being bsg_class_device's parent?

Patch

diff --git a/block/bsg-lib.c b/block/bsg-lib.c
index fc2e5ff..90c28fd 100644
--- a/block/bsg-lib.c
+++ b/block/bsg-lib.c
@@ -309,6 +309,25 @@  struct request_queue *bsg_setup_queue(struct device *dev, const char *name,
 		bsg_job_fn *job_fn, int dd_job_size,
 		void (*release)(struct device *))
 {
+	return bsg_setup_queue_ex(dev, name, job_fn, dd_job_size, release,
+		NULL);
+}
+EXPORT_SYMBOL_GPL(bsg_setup_queue);
+
+/**
+ * bsg_setup_queue - Create and add the bsg hooks so we can receive requests
+ * @dev: device to attach bsg device to
+ * @name: device to give bsg device
+ * @job_fn: bsg job handler
+ * @dd_job_size: size of LLD data needed for each job
+ * @release: @dev release function
+ * @dev_module: @dev's module
+ */
+struct request_queue *bsg_setup_queue_ex(struct device *dev, const char *name,
+		bsg_job_fn *job_fn, int dd_job_size,
+		void (*release)(struct device *),
+		struct module *dev_module)
+{
 	struct request_queue *q;
 	int ret;
 
@@ -331,7 +350,8 @@  struct request_queue *bsg_setup_queue(struct device *dev, const char *name,
 	blk_queue_softirq_done(q, bsg_softirq_done);
 	blk_queue_rq_timeout(q, BLK_DEFAULT_SG_TIMEOUT);
 
-	ret = bsg_register_queue(q, dev, name, &bsg_transport_ops, release);
+	ret = bsg_register_queue_ex(q, dev, name, &bsg_transport_ops, release,
+		dev_module);
 	if (ret) {
 		printk(KERN_ERR "%s: bsg interface failed to "
 		       "initialize - register queue\n", dev->kobj.name);
@@ -343,4 +363,4 @@  struct request_queue *bsg_setup_queue(struct device *dev, const char *name,
 	blk_cleanup_queue(q);
 	return ERR_PTR(ret);
 }
-EXPORT_SYMBOL_GPL(bsg_setup_queue);
+EXPORT_SYMBOL_GPL(bsg_setup_queue_ex);
diff --git a/block/bsg.c b/block/bsg.c
index defa06c..6920c5b 100644
--- a/block/bsg.c
+++ b/block/bsg.c
@@ -750,7 +750,8 @@  static struct bsg_device *__bsg_get_device(int minor, struct request_queue *q)
 	return bd;
 }
 
-static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file)
+static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file,
+	struct bsg_class_device **pbcd)
 {
 	struct bsg_device *bd;
 	struct bsg_class_device *bcd;
@@ -766,6 +767,7 @@  static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file)
 
 	if (!bcd)
 		return ERR_PTR(-ENODEV);
+	*pbcd = bcd;
 
 	bd = __bsg_get_device(iminor(inode), bcd->queue);
 	if (bd)
@@ -781,22 +783,34 @@  static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file)
 static int bsg_open(struct inode *inode, struct file *file)
 {
 	struct bsg_device *bd;
+	struct bsg_class_device *bcd;
 
-	bd = bsg_get_device(inode, file);
+	bd = bsg_get_device(inode, file, &bcd);
 
 	if (IS_ERR(bd))
 		return PTR_ERR(bd);
 
 	file->private_data = bd;
+	if (bcd->parent_module) {
+		if (!try_module_get(bcd->parent_module)) {
+			bsg_put_device(bd);
+			return -ENODEV;
+		}
+	}
 	return 0;
 }
 
 static int bsg_release(struct inode *inode, struct file *file)
 {
+	int ret;
 	struct bsg_device *bd = file->private_data;
+	struct module *parent_module = bd->queue->bsg_dev.parent_module;
 
 	file->private_data = NULL;
-	return bsg_put_device(bd);
+	ret = bsg_put_device(bd);
+	if (parent_module)
+		module_put(parent_module);
+	return ret;
 }
 
 static __poll_t bsg_poll(struct file *file, poll_table *wait)
@@ -922,6 +936,14 @@  int bsg_register_queue(struct request_queue *q, struct device *parent,
 		const char *name, const struct bsg_ops *ops,
 		void (*release)(struct device *))
 {
+	return bsg_register_queue_ex(q, parent, name, ops, release, NULL);
+}
+
+int bsg_register_queue_ex(struct request_queue *q, struct device *parent,
+		const char *name, const struct bsg_ops *ops,
+		void (*release)(struct device *),
+		struct module *parent_module)
+{
 	struct bsg_class_device *bcd;
 	dev_t dev;
 	int ret;
@@ -958,6 +980,7 @@  int bsg_register_queue(struct request_queue *q, struct device *parent,
 	bcd->parent = get_device(parent);
 	bcd->release = release;
 	bcd->ops = ops;
+	bcd->parent_module = parent_module;
 	kref_init(&bcd->ref);
 	dev = MKDEV(bsg_major, bcd->minor);
 	class_dev = device_create(bsg_class, parent, dev, NULL, "%s", devname);
diff --git a/drivers/scsi/scsi_transport_fc.c b/drivers/scsi/scsi_transport_fc.c
index be3be0f..f21f7d2 100644
--- a/drivers/scsi/scsi_transport_fc.c
+++ b/drivers/scsi/scsi_transport_fc.c
@@ -3772,17 +3772,21 @@  static int fc_bsg_dispatch(struct bsg_job *job)
 	struct fc_internal *i = to_fc_internal(shost->transportt);
 	struct request_queue *q;
 	char bsg_name[20];
+	struct module *shost_module = NULL;
 
 	fc_host->rqst_q = NULL;
 
 	if (!i->f->bsg_request)
 		return -ENOTSUPP;
 
+	if (shost->hostt)
+		shost_module = shost->hostt->module;
+
 	snprintf(bsg_name, sizeof(bsg_name),
 		 "fc_host%d", shost->host_no);
 
-	q = bsg_setup_queue(dev, bsg_name, fc_bsg_dispatch, i->f->dd_bsg_size,
-			NULL);
+	q = bsg_setup_queue_ex(dev, bsg_name, fc_bsg_dispatch,
+		i->f->dd_bsg_size, NULL, shost_module);
 	if (IS_ERR(q)) {
 		dev_err(dev,
 			"fc_host%d: bsg interface failed to initialize - setup queue\n",
diff --git a/include/linux/bsg-lib.h b/include/linux/bsg-lib.h
index 28a7ccc..232c855 100644
--- a/include/linux/bsg-lib.h
+++ b/include/linux/bsg-lib.h
@@ -74,6 +74,10 @@  void bsg_job_done(struct bsg_job *job, int result,
 struct request_queue *bsg_setup_queue(struct device *dev, const char *name,
 		bsg_job_fn *job_fn, int dd_job_size,
 		void (*release)(struct device *));
+struct request_queue *bsg_setup_queue_ex(struct device *dev, const char *name,
+		bsg_job_fn *job_fn, int dd_job_size,
+		void (*release)(struct device *),
+		struct module *dev_module);
 void bsg_job_put(struct bsg_job *job);
 int __must_check bsg_job_get(struct bsg_job *job);
 
diff --git a/include/linux/bsg.h b/include/linux/bsg.h
index 0c7dd9c..0e7c26c 100644
--- a/include/linux/bsg.h
+++ b/include/linux/bsg.h
@@ -23,11 +23,16 @@  struct bsg_class_device {
 	struct kref ref;
 	const struct bsg_ops *ops;
 	void (*release)(struct device *);
+	struct module *parent_module;
 };
 
 int bsg_register_queue(struct request_queue *q, struct device *parent,
 		const char *name, const struct bsg_ops *ops,
 		void (*release)(struct device *));
+int bsg_register_queue_ex(struct request_queue *q, struct device *parent,
+		const char *name, const struct bsg_ops *ops,
+		void (*release)(struct device *),
+		struct module *parent_module);
 int bsg_scsi_register_queue(struct request_queue *q, struct device *parent);
 void bsg_unregister_queue(struct request_queue *q);
 #else