| Message ID | 20210909034608.1435-1-yuzenghui@huawei.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | scsi: bsg: Fix device unregistration | expand |
Looks good,
Reviewed-by: Christoph Hellwig <hch@lst.de>
On Thu, Sep 09, 2021 at 11:46:08AM +0800, Zenghui Yu wrote: > We use device_initialize() to take refcount for the device but forget to > put_device() on device teardown, which ends up leaking private data of the > driver core, dev_name(), etc. This is reported by kmemleak at boot time if > we compile kernel with DEBUG_TEST_DRIVER_REMOVE. > > Note that adding the missing put_device() is _not_ sufficient to fix device > unregistration. As we don't provide the .release() method for device, which > turned out to be typically wrong and will be complained loudly by the > driver core. > > Fix both of them. > > Fixes: ead09dd3aed5 ("scsi: bsg: Simplify device registration") > Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> > --- > block/bsg.c | 13 +++++++++++-- > 1 file changed, 11 insertions(+), 2 deletions(-) Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
On 9/8/21 9:46 PM, Zenghui Yu wrote: > We use device_initialize() to take refcount for the device but forget to > put_device() on device teardown, which ends up leaking private data of the > driver core, dev_name(), etc. This is reported by kmemleak at boot time if > we compile kernel with DEBUG_TEST_DRIVER_REMOVE. > > Note that adding the missing put_device() is _not_ sufficient to fix device > unregistration. As we don't provide the .release() method for device, which > turned out to be typically wrong and will be complained loudly by the > driver core. > > Fix both of them. Applied, thanks.
On 9/9/21 2:40 PM, Jens Axboe wrote: > On 9/8/21 9:46 PM, Zenghui Yu wrote: >> We use device_initialize() to take refcount for the device but forget to >> put_device() on device teardown, which ends up leaking private data of the >> driver core, dev_name(), etc. This is reported by kmemleak at boot time if >> we compile kernel with DEBUG_TEST_DRIVER_REMOVE. >> >> Note that adding the missing put_device() is _not_ sufficient to fix device >> unregistration. As we don't provide the .release() method for device, which >> turned out to be typically wrong and will be complained loudly by the >> driver core. >> >> Fix both of them. > > Applied, thanks. Actually, let's move this through the SCSI tree, as the offending patch went that way (and my branches are behind that point).
Jens, > Actually, let's move this through the SCSI tree, as the offending > patch went that way (and my branches are behind that point). Sure. I queued it up...
On Thu, Sep 09, 2021 at 02:42:10PM -0600, Jens Axboe wrote: > Actually, let's move this through the SCSI tree, as the offending patch > went that way (and my branches are behind that point). Btw, should we move bsg.c and bsg-lib.c to drivers/scsi/? They very much are SCSI infrastructure now.
On Thu, Sep 09, 2021 at 11:46:08AM +0800, Zenghui Yu wrote: > We use device_initialize() to take refcount for the device but forget to > put_device() on device teardown, which ends up leaking private data of the > driver core, dev_name(), etc. This is reported by kmemleak at boot time if > we compile kernel with DEBUG_TEST_DRIVER_REMOVE. > > Note that adding the missing put_device() is _not_ sufficient to fix device > unregistration. As we don't provide the .release() method for device, which > turned out to be typically wrong and will be complained loudly by the > driver core. > > Fix both of them. > > Fixes: ead09dd3aed5 ("scsi: bsg: Simplify device registration") > Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> > --- > block/bsg.c | 13 +++++++++++-- > 1 file changed, 11 insertions(+), 2 deletions(-) > +static void bsg_device_release(struct device *dev) > +{ > + struct bsg_device *bd = container_of(dev, struct bsg_device, device); > + > + ida_simple_remove(&bsg_minor_ida, MINOR(bd->device.devt)); > + kfree(bd); > +} > @@ -198,6 +205,7 @@ struct bsg_device *bsg_register_queue(struct request_queue *q, > bd->device.devt = MKDEV(bsg_major, ret); > bd->device.class = bsg_class; > bd->device.parent = parent; > + bd->device.release = bsg_device_release; > dev_set_name(&bd->device, "%s", name); > device_initialize(&bd->device); > > @@ -218,6 +226,7 @@ struct bsg_device *bsg_register_queue(struct request_queue *q, > out_device_del: > cdev_device_del(&bd->cdev, &bd->device); > out_ida_remove: > + put_device(&bd->device); > ida_simple_remove(&bsg_minor_ida, MINOR(bd->device.devt)); > out_kfree: > kfree(bd); Ehh, what about the blatant use-after-free and double-free you just added here? Martin, can this still be dropped from the scsi tree or does it need to be fixed incrementally? Johan
Johan, > Martin, can this still be dropped from the scsi tree or does it need to > be fixed incrementally? I dropped it. Zenghui: Please fix.
On 2021/9/10 20:45, Johan Hovold wrote: > On Thu, Sep 09, 2021 at 11:46:08AM +0800, Zenghui Yu wrote: >> @@ -218,6 +226,7 @@ struct bsg_device *bsg_register_queue(struct request_queue *q, >> out_device_del: >> cdev_device_del(&bd->cdev, &bd->device); >> out_ida_remove: >> + put_device(&bd->device); >> ida_simple_remove(&bsg_minor_ida, MINOR(bd->device.devt)); >> out_kfree: >> kfree(bd); > > Ehh, what about the blatant use-after-free and double-free you just > added here? Yeah, whoops. That's definitely wrong. I'm squash the following fix in this patch. Thanks for the heads up! |diff --git a/block/bsg.c b/block/bsg.c |index c3bb92b9af7e..882f56bff14f 100644 |--- a/block/bsg.c |+++ b/block/bsg.c |@@ -200,7 +200,8 @@ struct bsg_device *bsg_register_queue(struct request_queue *q, | if (ret < 0) { | if (ret == -ENOSPC) | dev_err(parent, "bsg: too many bsg devices\n"); |- goto out_kfree; |+ kfree(bd); |+ return ERR_PTR(ret); | } | bd->device.devt = MKDEV(bsg_major, ret); | bd->device.class = bsg_class; |@@ -213,7 +214,7 @@ struct bsg_device *bsg_register_queue(struct request_queue *q, | bd->cdev.owner = THIS_MODULE; | ret = cdev_device_add(&bd->cdev, &bd->device); | if (ret) |- goto out_ida_remove; |+ goto out_put_device; | | if (q->kobj.sd) { | ret = sysfs_create_link(&q->kobj, &bd->device.kobj, "bsg"); |@@ -225,11 +226,8 @@ struct bsg_device *bsg_register_queue(struct request_queue *q, | | out_device_del: | cdev_device_del(&bd->cdev, &bd->device); |-out_ida_remove: |+out_put_device: | put_device(&bd->device); |- ida_simple_remove(&bsg_minor_ida, MINOR(bd->device.devt)); |-out_kfree: |- kfree(bd); | return ERR_PTR(ret); | } | EXPORT_SYMBOL_GPL(bsg_register_queue);
diff --git a/block/bsg.c b/block/bsg.c index 351095193788..c3bb92b9af7e 100644 --- a/block/bsg.c +++ b/block/bsg.c @@ -165,13 +165,20 @@ static const struct file_operations bsg_fops = { .llseek = default_llseek, }; +static void bsg_device_release(struct device *dev) +{ + struct bsg_device *bd = container_of(dev, struct bsg_device, device); + + ida_simple_remove(&bsg_minor_ida, MINOR(bd->device.devt)); + kfree(bd); +} + void bsg_unregister_queue(struct bsg_device *bd) { if (bd->queue->kobj.sd) sysfs_remove_link(&bd->queue->kobj, "bsg"); cdev_device_del(&bd->cdev, &bd->device); - ida_simple_remove(&bsg_minor_ida, MINOR(bd->device.devt)); - kfree(bd); + put_device(&bd->device); } EXPORT_SYMBOL_GPL(bsg_unregister_queue); @@ -198,6 +205,7 @@ struct bsg_device *bsg_register_queue(struct request_queue *q, bd->device.devt = MKDEV(bsg_major, ret); bd->device.class = bsg_class; bd->device.parent = parent; + bd->device.release = bsg_device_release; dev_set_name(&bd->device, "%s", name); device_initialize(&bd->device); @@ -218,6 +226,7 @@ struct bsg_device *bsg_register_queue(struct request_queue *q, out_device_del: cdev_device_del(&bd->cdev, &bd->device); out_ida_remove: + put_device(&bd->device); ida_simple_remove(&bsg_minor_ida, MINOR(bd->device.devt)); out_kfree: kfree(bd);
We use device_initialize() to take refcount for the device but forget to put_device() on device teardown, which ends up leaking private data of the driver core, dev_name(), etc. This is reported by kmemleak at boot time if we compile kernel with DEBUG_TEST_DRIVER_REMOVE. Note that adding the missing put_device() is _not_ sufficient to fix device unregistration. As we don't provide the .release() method for device, which turned out to be typically wrong and will be complained loudly by the driver core. Fix both of them. Fixes: ead09dd3aed5 ("scsi: bsg: Simplify device registration") Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> --- block/bsg.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)