A follow-up on earlier discussions: [PATCH] bsg referencing bus driver module https://www.spinics.net/lists/linux-scsi/msg119631.html [PATCH] Waiting for scsi_host_template release https://www.spinics.net/lists/linux-scsi/msg119432.html All these discussions are attempts to fix a crash after SCSI transport driver unload if a user-mode process holds a handle in BSG layer towards the unloaded driver via SCSI mid-layer: [16834.636216,07] Call Trace: ... scsi_proc_hostdir_rm [16834.641944,07] [<ffffffff8141723f>] scsi_host_dev_release+0x3f/0x130 [16834.647740,07] [<ffffffff813e4f82>] device_release+0x32/0xa0 [16834.653423,07] [<ffffffff812dc6c7>] kobject_cleanup+0x77/0x190 [16834.659002,07] [<ffffffff812dc585>] kobject_put+0x25/0x50 [16834.664430,07] [<ffffffff813e5277>] put_device+0x17/0x20 [16834.669740,07] [<ffffffff812d0334>] bsg_kref_release_function+0x24/0x30 [16834.675007,07] [<ffffffff812d14a6>] bsg_release+0x166/0x1d0 [16834.680148,07] [<ffffffff8119ba2b>] __fput+0xcb/0x1d0 [16834.685156,07] [<ffffffff8119bb6e>] ____fput+0xe/0x10 [16834.690017,07] [<ffffffff81077476>] task_work_run+0x86/0xb0 [16834.694781,07] [<ffffffff81057043>] exit_to_usermode_loop+0x6b/0x9a [16834.699466,07] [<ffffffff81002875>] syscall_return_slowpath+0x55/0x60 [16834.704110,07] [<ffffffff8172d615>] int_ret_from_sys_call+0x25/0x9f The latest input from earlier discussions was to cut off access to the unloaded driver at bsg_unregister_queue time by calling blk_cleanup_queue. If we do that we still have to release the reference to the parent device (otherwise we crash with the same stack). The next logical step is, rather than maintaining a "part-time" reference to be dropped early, we discard referencing completely. Discarding the reference turns out to be the only thing needed to fix the problem: all transport drivers already do blk_cleanup_queue before releasing their reference to the parent device. >From 7eaa9b43f0b99766b1d197044eb4d2e549d11a24 Mon Sep 17 00:00:00 2001 From: Anatoliy Glagolev <glagolig@xxxxxxxxx> Date: Wed, 16 May 2018 17:03:09 -0600 Subject: [PATCH] Bsg referencing parent device Signed-off-by: Anatoliy Glagolev <glagolig@xxxxxxxxx> Bsg holding reference to a parent device may result in a crash if user closes a bsg file handle after the parent device driver has unloaded. Holding a reference is not really needed: parent device must exist between bsg_register_queue and bsg_unregister_queue. Before the device goes away the caller does blk_cleanup_queue so that all in-flight requests to the device are gone and all new requests cannot pass beyond the queue. The queue itself is a refcounted object and it will stay alive with a bsg file. --- block/bsg.c | 37 +++++++++++++++++-------------------- 1 file changed, 17 insertions(+), 20 deletions(-) diff --git a/block/bsg.c b/block/bsg.c index defa06c..f9e4b91 100644 --- a/block/bsg.c +++ b/block/bsg.c @@ -650,18 +650,6 @@ static struct bsg_device *bsg_alloc_device(void) return bd; } -static void bsg_kref_release_function(struct kref *kref) -{ - struct bsg_class_device *bcd = - container_of(kref, struct bsg_class_device, ref); - struct device *parent = bcd->parent; - - if (bcd->release) - bcd->release(bcd->parent); - - put_device(parent); -} - static int bsg_put_device(struct bsg_device *bd) { int ret = 0, do_free; @@ -694,7 +682,6 @@ static int bsg_put_device(struct bsg_device *bd) kfree(bd); out: - kref_put(&q->bsg_dev.ref, bsg_kref_release_function); if (do_free) blk_put_queue(q); return ret; @@ -760,8 +747,6 @@ static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file) */ mutex_lock(&bsg_mutex); bcd = idr_find(&bsg_minor_idr, iminor(inode)); - if (bcd) - kref_get(&bcd->ref); mutex_unlock(&bsg_mutex); if (!bcd) @@ -772,8 +757,6 @@ static struct bsg_device *bsg_get_device(struct inode *inode, struct file *file) return bd; bd = bsg_add_device(inode, bcd->queue, file); - if (IS_ERR(bd)) - kref_put(&bcd->ref, bsg_kref_release_function); return bd; } @@ -902,6 +885,8 @@ static long bsg_ioctl(struct file *file, unsigned int cmd, unsigned long arg) void bsg_unregister_queue(struct request_queue *q) { + struct device *parent; + void (*release)(struct device *dev); struct bsg_class_device *bcd = &q->bsg_dev; if (!bcd->class_dev) @@ -913,8 +898,21 @@ void bsg_unregister_queue(struct request_queue *q) sysfs_remove_link(&q->kobj, "bsg"); device_unregister(bcd->class_dev); bcd->class_dev = NULL; - kref_put(&bcd->ref, bsg_kref_release_function); + parent = bcd->parent; + release = bcd->release; + bcd->parent = NULL; + bcd->release = NULL; mutex_unlock(&bsg_mutex); + + /* + * The caller of bsg_[un]register_queue must hold a reference + * to the parent device between ..register.. and ..unregister.. + * so we do not maintain a refcount to parent here. + * Note: the caller is expected to blk_cleanup_queue(q) + * before releasing the reference to the parent device. + */ + if (release) + release(parent); } EXPORT_SYMBOL_GPL(bsg_unregister_queue); @@ -955,10 +953,9 @@ int bsg_register_queue(struct request_queue *q, struct device *parent, bcd->minor = ret; bcd->queue = q; - bcd->parent = get_device(parent); + bcd->parent = parent; bcd->release = release; bcd->ops = ops; - kref_init(&bcd->ref); dev = MKDEV(bsg_major, bcd->minor); class_dev = device_create(bsg_class, parent, dev, NULL, "%s", devname); if (IS_ERR(class_dev)) { -- 1.9.1