| Message ID | 20190516233034.16407-4-parav@mellanox.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | vfio/mdev: Improve vfio/mdev core module | expand |
On Thu, 16 May 2019 18:30:34 -0500 Parav Pandit <parav@mellanox.com> wrote: > In following sequences, child devices created while removing mdev parent > device can be left out, or it may lead to race of removing half > initialized child mdev devices. > > issue-1: > -------- > cpu-0 cpu-1 > ----- ----- > mdev_unregister_device() > device_for_each_child() > mdev_device_remove_cb() > mdev_device_remove() > create_store() > mdev_device_create() [...] > device_add() > parent_remove_sysfs_files() > > /* BUG: device added by cpu-0 > * whose parent is getting removed > * and it won't process this mdev. > */ > > issue-2: > -------- > Below crash is observed when user initiated remove is in progress > and mdev_unregister_driver() completes parent unregistration. > > cpu-0 cpu-1 > ----- ----- > remove_store() > mdev_device_remove() > active = false; > mdev_unregister_device() > parent device removed. > [...] > parents->ops->remove() > /* > * BUG: Accessing invalid parent. > */ > > This is similar race like create() racing with mdev_unregister_device(). > > BUG: unable to handle kernel paging request at ffffffffc0585668 > PGD e8f618067 P4D e8f618067 PUD e8f61a067 PMD 85adca067 PTE 0 > Oops: 0000 [#1] SMP PTI > CPU: 41 PID: 37403 Comm: bash Kdump: loaded Not tainted 5.1.0-rc6-vdevbus+ #6 > Hardware name: Supermicro SYS-6028U-TR4+/X10DRU-i+, BIOS 2.0b 08/09/2016 > RIP: 0010:mdev_device_remove+0xfa/0x140 [mdev] > Call Trace: > remove_store+0x71/0x90 [mdev] > kernfs_fop_write+0x113/0x1a0 > vfs_write+0xad/0x1b0 > ksys_write+0x5a/0xe0 > do_syscall_64+0x5a/0x210 > entry_SYSCALL_64_after_hwframe+0x49/0xbe > > Therefore, mdev core is improved as below to overcome above issues. > > Wait for any ongoing mdev create() and remove() to finish before > unregistering parent device using refcount and completion. > This continues to allow multiple create and remove to progress in > parallel for different mdev devices as most common case. > At the same time guard parent removal while parent is being access by > create() and remove callbacks. > > Code is simplified from kref to use refcount as unregister_device() has > to wait anyway for all create/remove to finish. > > While removing mdev devices during parent unregistration, there isn't > need to acquire refcount of parent device, hence code is restructured > using mdev_device_remove_common() to avoid it. > > Fixes: 7b96953bc640 ("vfio: Mediated device Core driver") > Signed-off-by: Parav Pandit <parav@mellanox.com> > --- > drivers/vfio/mdev/mdev_core.c | 86 ++++++++++++++++++++------------ > drivers/vfio/mdev/mdev_private.h | 6 ++- > 2 files changed, 60 insertions(+), 32 deletions(-) I'm still not quite happy with this patch. I think most of my dislike comes from how you are using a member called 'refcount' vs. what I believe a refcount actually is. See below. > > diff --git a/drivers/vfio/mdev/mdev_core.c b/drivers/vfio/mdev/mdev_core.c > index 0bef0cae1d4b..ca33246c1dc3 100644 > --- a/drivers/vfio/mdev/mdev_core.c > +++ b/drivers/vfio/mdev/mdev_core.c > @@ -78,34 +78,41 @@ static struct mdev_parent *__find_parent_device(struct device *dev) > return NULL; > } > > -static void mdev_release_parent(struct kref *kref) > +static bool mdev_try_get_parent(struct mdev_parent *parent) > { > - struct mdev_parent *parent = container_of(kref, struct mdev_parent, > - ref); > - struct device *dev = parent->dev; > - > - kfree(parent); > - put_device(dev); > + if (parent) > + return refcount_inc_not_zero(&parent->refcount); > + return false; > } > > -static struct mdev_parent *mdev_get_parent(struct mdev_parent *parent) > +static void mdev_put_parent(struct mdev_parent *parent) > { > - if (parent) > - kref_get(&parent->ref); > - > - return parent; > + if (parent && refcount_dec_and_test(&parent->refcount)) > + complete(&parent->unreg_completion); > } So far, this is "obtain a reference if the reference is not 0 (implying the object is not ready to use) and notify waiters when the last reference is dropped". This still looks idiomatic enough. > > -static void mdev_put_parent(struct mdev_parent *parent) > +static void mdev_device_remove_common(struct mdev_device *mdev) > { > - if (parent) > - kref_put(&parent->ref, mdev_release_parent); > + struct mdev_parent *parent; > + struct mdev_type *type; > + int ret; > + > + type = to_mdev_type(mdev->type_kobj); > + mdev_remove_sysfs_files(&mdev->dev, type); > + device_del(&mdev->dev); > + parent = mdev->parent; > + ret = parent->ops->remove(mdev); > + if (ret) > + dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); > + > + /* Balances with device_initialize() */ > + put_device(&mdev->dev); > } > > static int mdev_device_remove_cb(struct device *dev, void *data) > { > if (dev_is_mdev(dev)) > - mdev_device_remove(dev); > + mdev_device_remove_common(to_mdev_device(dev)); > > return 0; > } > @@ -147,7 +154,8 @@ int mdev_register_device(struct device *dev, const struct mdev_parent_ops *ops) > goto add_dev_err; > } > > - kref_init(&parent->ref); > + refcount_set(&parent->refcount, 1); Initializing to 1 when creating is also fine. > + init_completion(&parent->unreg_completion); > > parent->dev = dev; > parent->ops = ops; > @@ -206,14 +214,27 @@ void mdev_unregister_device(struct device *dev) > dev_info(dev, "MDEV: Unregistering\n"); > > list_del(&parent->next); > + mutex_unlock(&parent_list_lock); > + > + /* Release the initial reference so that new create cannot start */ > + mdev_put_parent(parent); The comment is confusing: We do drop one reference, but this does not imply we're going to 0 (which would be the one thing that would block creating new devices). > + > + /* > + * Wait for all the create and remove references to drop. > + */ > + wait_for_completion(&parent->unreg_completion); It only reaches 0 after this wait. > + > + /* > + * New references cannot be taken and all users are done > + * using the parent. So it is safe to unregister parent. > + */ > class_compat_remove_link(mdev_bus_compat_class, dev, NULL); > > device_for_each_child(dev, NULL, mdev_device_remove_cb); > > parent_remove_sysfs_files(parent); > - > - mutex_unlock(&parent_list_lock); > - mdev_put_parent(parent); > + kfree(parent); > + put_device(dev); > } > EXPORT_SYMBOL(mdev_unregister_device); > > @@ -237,10 +258,11 @@ int mdev_device_create(struct kobject *kobj, > struct mdev_parent *parent; > struct mdev_type *type = to_mdev_type(kobj); > > - parent = mdev_get_parent(type->parent); > - if (!parent) > + if (!mdev_try_get_parent(type->parent)) If other calls are still running, the refcount won't be 0, and this will succeed, even if we really want to get rid of the device. > return -EINVAL; > > + parent = type->parent; > + > mutex_lock(&mdev_list_lock); > > /* Check for duplicate */ > @@ -287,6 +309,7 @@ int mdev_device_create(struct kobject *kobj, > > mdev->active = true; > dev_dbg(&mdev->dev, "MDEV: created\n"); > + mdev_put_parent(parent); > > return 0; > > @@ -306,7 +329,6 @@ int mdev_device_remove(struct device *dev) > struct mdev_device *mdev, *tmp; > struct mdev_parent *parent; > struct mdev_type *type; > - int ret; > > mdev = to_mdev_device(dev); > > @@ -330,15 +352,17 @@ int mdev_device_remove(struct device *dev) > mutex_unlock(&mdev_list_lock); > > type = to_mdev_type(mdev->type_kobj); > - mdev_remove_sysfs_files(dev, type); > - device_del(&mdev->dev); > - parent = mdev->parent; > - ret = parent->ops->remove(mdev); > - if (ret) > - dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); > + if (!mdev_try_get_parent(type->parent)) { Same here: Is there really a guarantee that the refcount is 0 when the parent is going away? > + /* > + * Parent unregistration have started. > + * No need to remove here. > + */ > + mutex_unlock(&mdev_list_lock); Btw., you already unlocked above. > + return -ENODEV; > + } > > - /* Balances with device_initialize() */ > - put_device(&mdev->dev); > + parent = mdev->parent; > + mdev_device_remove_common(mdev); > mdev_put_parent(parent); > > return 0; > diff --git a/drivers/vfio/mdev/mdev_private.h b/drivers/vfio/mdev/mdev_private.h > index 924ed2274941..55ebab0af7b0 100644 > --- a/drivers/vfio/mdev/mdev_private.h > +++ b/drivers/vfio/mdev/mdev_private.h > @@ -19,7 +19,11 @@ void mdev_bus_unregister(void); > struct mdev_parent { > struct device *dev; > const struct mdev_parent_ops *ops; > - struct kref ref; > + /* Protects unregistration to wait until create/remove > + * are completed. > + */ > + refcount_t refcount; > + struct completion unreg_completion; > struct list_head next; > struct kset *mdev_types_kset; > struct list_head type_list; I think what's really needed is to split up the different needs and not overload the 'refcount' concept. - If we need to make sure that a reference to the parent is held so that the parent may not go away while still in use, we should continue to use the kref (in the idiomatic way it is used before this patch.) - We need to protect against creation of new devices if the parent is going away. Maybe set a going_away marker in the parent structure for that so that creation bails out immediately? What happens if the creation has already started when parent removal kicks in, though? Do we need some child list locking and an indication whether a child is in progress of being registered/unregistered? - We also need to protect against removal of devices while unregister is in progress (same mechanism as above?) The second issue you describe above should be fixed then if the children keep a reference of the parent.
Hi Cornelia, > -----Original Message----- > From: Cornelia Huck <cohuck@redhat.com> > Sent: Friday, May 17, 2019 6:22 AM > To: Parav Pandit <parav@mellanox.com> > Cc: kvm@vger.kernel.org; linux-kernel@vger.kernel.org; > kwankhede@nvidia.com; alex.williamson@redhat.com; cjia@nvidia.com > Subject: Re: [PATCHv3 3/3] vfio/mdev: Synchronize device create/remove > with parent removal > > On Thu, 16 May 2019 18:30:34 -0500 > Parav Pandit <parav@mellanox.com> wrote: > > > In following sequences, child devices created while removing mdev > > parent device can be left out, or it may lead to race of removing half > > initialized child mdev devices. > > > > issue-1: > > -------- > > cpu-0 cpu-1 > > ----- ----- > > mdev_unregister_device() > > device_for_each_child() > > mdev_device_remove_cb() > > mdev_device_remove() > > create_store() > > mdev_device_create() [...] > > device_add() > > parent_remove_sysfs_files() > > > > /* BUG: device added by cpu-0 > > * whose parent is getting removed > > * and it won't process this mdev. > > */ > > > > issue-2: > > -------- > > Below crash is observed when user initiated remove is in progress and > > mdev_unregister_driver() completes parent unregistration. > > > > cpu-0 cpu-1 > > ----- ----- > > remove_store() > > mdev_device_remove() > > active = false; > > mdev_unregister_device() > > parent device removed. > > [...] > > parents->ops->remove() > > /* > > * BUG: Accessing invalid parent. > > */ > > > > This is similar race like create() racing with mdev_unregister_device(). > > > > BUG: unable to handle kernel paging request at ffffffffc0585668 PGD > > e8f618067 P4D e8f618067 PUD e8f61a067 PMD 85adca067 PTE 0 > > Oops: 0000 [#1] SMP PTI > > CPU: 41 PID: 37403 Comm: bash Kdump: loaded Not tainted > > 5.1.0-rc6-vdevbus+ #6 Hardware name: Supermicro > > SYS-6028U-TR4+/X10DRU-i+, BIOS 2.0b 08/09/2016 > > RIP: 0010:mdev_device_remove+0xfa/0x140 [mdev] Call Trace: > > remove_store+0x71/0x90 [mdev] > > kernfs_fop_write+0x113/0x1a0 > > vfs_write+0xad/0x1b0 > > ksys_write+0x5a/0xe0 > > do_syscall_64+0x5a/0x210 > > entry_SYSCALL_64_after_hwframe+0x49/0xbe > > > > Therefore, mdev core is improved as below to overcome above issues. > > > > Wait for any ongoing mdev create() and remove() to finish before > > unregistering parent device using refcount and completion. > > This continues to allow multiple create and remove to progress in > > parallel for different mdev devices as most common case. > > At the same time guard parent removal while parent is being access by > > create() and remove callbacks. > > > > Code is simplified from kref to use refcount as unregister_device() > > has to wait anyway for all create/remove to finish. > > > > While removing mdev devices during parent unregistration, there isn't > > need to acquire refcount of parent device, hence code is restructured > > using mdev_device_remove_common() to avoid it. > > > > Fixes: 7b96953bc640 ("vfio: Mediated device Core driver") > > Signed-off-by: Parav Pandit <parav@mellanox.com> > > --- > > drivers/vfio/mdev/mdev_core.c | 86 ++++++++++++++++++++------------ > > drivers/vfio/mdev/mdev_private.h | 6 ++- > > 2 files changed, 60 insertions(+), 32 deletions(-) > > I'm still not quite happy with this patch. I think most of my dislike comes > from how you are using a member called 'refcount' vs. what I believe a > refcount actually is. See below. > Comments below. > > > > diff --git a/drivers/vfio/mdev/mdev_core.c > > b/drivers/vfio/mdev/mdev_core.c index 0bef0cae1d4b..ca33246c1dc3 > > 100644 > > --- a/drivers/vfio/mdev/mdev_core.c > > +++ b/drivers/vfio/mdev/mdev_core.c > > @@ -78,34 +78,41 @@ static struct mdev_parent > *__find_parent_device(struct device *dev) > > return NULL; > > } > > > > -static void mdev_release_parent(struct kref *kref) > > +static bool mdev_try_get_parent(struct mdev_parent *parent) > > { > > - struct mdev_parent *parent = container_of(kref, struct > mdev_parent, > > - ref); > > - struct device *dev = parent->dev; > > - > > - kfree(parent); > > - put_device(dev); > > + if (parent) > > + return refcount_inc_not_zero(&parent->refcount); > > + return false; > > } > > > > -static struct mdev_parent *mdev_get_parent(struct mdev_parent > > *parent) > > +static void mdev_put_parent(struct mdev_parent *parent) > > { > > - if (parent) > > - kref_get(&parent->ref); > > - > > - return parent; > > + if (parent && refcount_dec_and_test(&parent->refcount)) > > + complete(&parent->unreg_completion); > > } > > So far, this is "obtain a reference if the reference is not 0 (implying the object > is not ready to use) and notify waiters when the last reference is dropped". > This still looks idiomatic enough. > > > > > -static void mdev_put_parent(struct mdev_parent *parent) > > +static void mdev_device_remove_common(struct mdev_device *mdev) > > { > > - if (parent) > > - kref_put(&parent->ref, mdev_release_parent); > > + struct mdev_parent *parent; > > + struct mdev_type *type; > > + int ret; > > + > > + type = to_mdev_type(mdev->type_kobj); > > + mdev_remove_sysfs_files(&mdev->dev, type); > > + device_del(&mdev->dev); > > + parent = mdev->parent; > > + ret = parent->ops->remove(mdev); > > + if (ret) > > + dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); > > + > > + /* Balances with device_initialize() */ > > + put_device(&mdev->dev); > > } > > > > static int mdev_device_remove_cb(struct device *dev, void *data) { > > if (dev_is_mdev(dev)) > > - mdev_device_remove(dev); > > + mdev_device_remove_common(to_mdev_device(dev)); > > > > return 0; > > } > > @@ -147,7 +154,8 @@ int mdev_register_device(struct device *dev, const > struct mdev_parent_ops *ops) > > goto add_dev_err; > > } > > > > - kref_init(&parent->ref); > > + refcount_set(&parent->refcount, 1); > > Initializing to 1 when creating is also fine. > > > + init_completion(&parent->unreg_completion); > > > > parent->dev = dev; > > parent->ops = ops; > > @@ -206,14 +214,27 @@ void mdev_unregister_device(struct device *dev) > > dev_info(dev, "MDEV: Unregistering\n"); > > > > list_del(&parent->next); > > + mutex_unlock(&parent_list_lock); > > + > > + /* Release the initial reference so that new create cannot start */ > > + mdev_put_parent(parent); > > The comment is confusing: We do drop one reference, but this does not > imply we're going to 0 (which would be the one thing that would block > creating new devices). > Ok. How about below comment. /* Balance with initial reference init */ > > + > > + /* > > + * Wait for all the create and remove references to drop. > > + */ > > + wait_for_completion(&parent->unreg_completion); > > It only reaches 0 after this wait. > Yes. > > + > > + /* > > + * New references cannot be taken and all users are done > > + * using the parent. So it is safe to unregister parent. > > + */ > > class_compat_remove_link(mdev_bus_compat_class, dev, NULL); > > > > device_for_each_child(dev, NULL, mdev_device_remove_cb); > > > > parent_remove_sysfs_files(parent); > > - > > - mutex_unlock(&parent_list_lock); > > - mdev_put_parent(parent); > > + kfree(parent); > > + put_device(dev); > > } > > EXPORT_SYMBOL(mdev_unregister_device); > > > > @@ -237,10 +258,11 @@ int mdev_device_create(struct kobject *kobj, > > struct mdev_parent *parent; > > struct mdev_type *type = to_mdev_type(kobj); > > > > - parent = mdev_get_parent(type->parent); > > - if (!parent) > > + if (!mdev_try_get_parent(type->parent)) > > If other calls are still running, the refcount won't be 0, and this will succeed, > even if we really want to get rid of the device. > Sure, if other calls are running, refcount won't be 0. Process creating them will eventually complete, and refcount will drop to zero. And new processes won't be able to start any more. So there is no differentiation between 'already in creation stage' and 'about to start' processes. > > return -EINVAL; > > > > + parent = type->parent; > > + > > mutex_lock(&mdev_list_lock); > > > > /* Check for duplicate */ > > @@ -287,6 +309,7 @@ int mdev_device_create(struct kobject *kobj, > > > > mdev->active = true; > > dev_dbg(&mdev->dev, "MDEV: created\n"); > > + mdev_put_parent(parent); > > > > return 0; > > > > @@ -306,7 +329,6 @@ int mdev_device_remove(struct device *dev) > > struct mdev_device *mdev, *tmp; > > struct mdev_parent *parent; > > struct mdev_type *type; > > - int ret; > > > > mdev = to_mdev_device(dev); > > > > @@ -330,15 +352,17 @@ int mdev_device_remove(struct device *dev) > > mutex_unlock(&mdev_list_lock); > > > > type = to_mdev_type(mdev->type_kobj); > > - mdev_remove_sysfs_files(dev, type); > > - device_del(&mdev->dev); > > - parent = mdev->parent; > > - ret = parent->ops->remove(mdev); > > - if (ret) > > - dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); > > + if (!mdev_try_get_parent(type->parent)) { > > Same here: Is there really a guarantee that the refcount is 0 when the parent > is going away? A WARN_ON after wait_for_completion or in freeing the parent is good to catch bugs. > > > + /* > > + * Parent unregistration have started. > > + * No need to remove here. > > + */ > > + mutex_unlock(&mdev_list_lock); > > Btw., you already unlocked above. > You are right. This unlock is wrong. I will revise the patch. > > + return -ENODEV; > > + } > > > > - /* Balances with device_initialize() */ > > - put_device(&mdev->dev); > > + parent = mdev->parent; > > + mdev_device_remove_common(mdev); > > mdev_put_parent(parent); > > > > return 0; > > diff --git a/drivers/vfio/mdev/mdev_private.h > > b/drivers/vfio/mdev/mdev_private.h > > index 924ed2274941..55ebab0af7b0 100644 > > --- a/drivers/vfio/mdev/mdev_private.h > > +++ b/drivers/vfio/mdev/mdev_private.h > > @@ -19,7 +19,11 @@ void mdev_bus_unregister(void); struct > mdev_parent > > { > > struct device *dev; > > const struct mdev_parent_ops *ops; > > - struct kref ref; > > + /* Protects unregistration to wait until create/remove > > + * are completed. > > + */ > > + refcount_t refcount; > > + struct completion unreg_completion; > > struct list_head next; > > struct kset *mdev_types_kset; > > struct list_head type_list; > > I think what's really needed is to split up the different needs and not > overload the 'refcount' concept. > Refcount tells that how many active references are present for this parent device. Those active reference could be create/remove processes and mdev core itself. So when parent unregisters, mdev module publishes that it is going away through this refcount. Hence new users cannot start. > - If we need to make sure that a reference to the parent is held so > that the parent may not go away while still in use, we should > continue to use the kref (in the idiomatic way it is used before this > patch.) > - We need to protect against creation of new devices if the parent is > going away. Maybe set a going_away marker in the parent structure for > that so that creation bails out immediately? Such marker will help to not start new processes. So an additional marker can be added to improve mdev_try_get_parent(). But I couldn't justify why differentiating those two users on time scale is desired. One reason could be that user continuously tries to create mdev and parent never gets a chance, to unregister? I guess, parent will run out mdev devices before this can happen. Additionally a stop marker is needed (counter) to tell that all users are done accessing it. Both purposes are served using a refcount scheme. > What happens if the > creation has already started when parent removal kicks in, though? That particular creation will succeed but newer cannot start, because mdev_put_parent() is done. > Do we need some child list locking and an indication whether a child > is in progress of being registered/unregistered? > - We also need to protect against removal of devices while unregister > is in progress (same mechanism as above?) The second issue you > describe above should be fixed then if the children keep a reference > of the parent. Parent unregistration publishes that its going away first, so no new device removal from user can start. Already on going removal by users anyway complete first. Once all remove() users are done, parent is getting unregistered.
On Fri, 17 May 2019 14:18:26 +0000 Parav Pandit <parav@mellanox.com> wrote: > > > @@ -206,14 +214,27 @@ void mdev_unregister_device(struct device *dev) > > > dev_info(dev, "MDEV: Unregistering\n"); > > > > > > list_del(&parent->next); > > > + mutex_unlock(&parent_list_lock); > > > + > > > + /* Release the initial reference so that new create cannot start */ > > > + mdev_put_parent(parent); > > > > The comment is confusing: We do drop one reference, but this does not > > imply we're going to 0 (which would be the one thing that would block > > creating new devices). > > > Ok. How about below comment. > /* Balance with initial reference init */ Well, 'release the initial reference' is fine; it's just the second part that is confusing. One thing that continues to irk me (and I'm sorry if I sound like a broken record) is that you give up the initial reference and then continue to use parent. For the more usual semantics of a reference count, that would be a bug (as the structure would be freed if the reference count dropped to zero), even though it is not a bug here. > > > > + > > > + /* > > > + * Wait for all the create and remove references to drop. > > > + */ > > > + wait_for_completion(&parent->unreg_completion); > > > > It only reaches 0 after this wait. > > > Yes. > > > > + > > > + /* > > > + * New references cannot be taken and all users are done > > > + * using the parent. So it is safe to unregister parent. > > > + */ > > > class_compat_remove_link(mdev_bus_compat_class, dev, NULL); > > > > > > device_for_each_child(dev, NULL, mdev_device_remove_cb); > > > > > > parent_remove_sysfs_files(parent); > > > - > > > - mutex_unlock(&parent_list_lock); > > > - mdev_put_parent(parent); > > > + kfree(parent); > > > + put_device(dev); > > > } > > > EXPORT_SYMBOL(mdev_unregister_device); > > > > > > @@ -237,10 +258,11 @@ int mdev_device_create(struct kobject *kobj, > > > struct mdev_parent *parent; > > > struct mdev_type *type = to_mdev_type(kobj); > > > > > > - parent = mdev_get_parent(type->parent); > > > - if (!parent) > > > + if (!mdev_try_get_parent(type->parent)) > > > > If other calls are still running, the refcount won't be 0, and this will succeed, > > even if we really want to get rid of the device. > > > Sure, if other calls are running, refcount won't be 0. Process creating them will eventually complete, and refcount will drop to zero. > And new processes won't be able to start any more. > So there is no differentiation between 'already in creation stage' and 'about to start' processes. Does it really make sense to allow creation to start if the parent is going away? > > > > return -EINVAL; > > > > > > + parent = type->parent; > > > + > > > mutex_lock(&mdev_list_lock); > > > > > > /* Check for duplicate */ > > > @@ -287,6 +309,7 @@ int mdev_device_create(struct kobject *kobj, > > > > > > mdev->active = true; > > > dev_dbg(&mdev->dev, "MDEV: created\n"); > > > + mdev_put_parent(parent); > > > > > > return 0; > > > > > > @@ -306,7 +329,6 @@ int mdev_device_remove(struct device *dev) > > > struct mdev_device *mdev, *tmp; > > > struct mdev_parent *parent; > > > struct mdev_type *type; > > > - int ret; > > > > > > mdev = to_mdev_device(dev); > > > > > > @@ -330,15 +352,17 @@ int mdev_device_remove(struct device *dev) > > > mutex_unlock(&mdev_list_lock); > > > > > > type = to_mdev_type(mdev->type_kobj); > > > - mdev_remove_sysfs_files(dev, type); > > > - device_del(&mdev->dev); > > > - parent = mdev->parent; > > > - ret = parent->ops->remove(mdev); > > > - if (ret) > > > - dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); > > > + if (!mdev_try_get_parent(type->parent)) { > > > > Same here: Is there really a guarantee that the refcount is 0 when the parent > > is going away? > A WARN_ON after wait_for_completion or in freeing the parent is good to catch bugs. I'd rather prefer to avoid having to add WARN_ONs :) This looks like it is supposed to be an early exit. However, if some other thread does any create or remove operation at the same time, we'll still do the remove, and we still might have have a race window (and this is getting really hard to follow in the code). > > > > > > + /* > > > + * Parent unregistration have started. > > > + * No need to remove here. > > > + */ > > > + mutex_unlock(&mdev_list_lock); > > > > Btw., you already unlocked above. > > > You are right. This unlock is wrong. I will revise the patch. > > > > + return -ENODEV; > > > + } > > > > > > - /* Balances with device_initialize() */ > > > - put_device(&mdev->dev); > > > + parent = mdev->parent; > > > + mdev_device_remove_common(mdev); > > > mdev_put_parent(parent); > > > > > > return 0; > > > diff --git a/drivers/vfio/mdev/mdev_private.h > > > b/drivers/vfio/mdev/mdev_private.h > > > index 924ed2274941..55ebab0af7b0 100644 > > > --- a/drivers/vfio/mdev/mdev_private.h > > > +++ b/drivers/vfio/mdev/mdev_private.h > > > @@ -19,7 +19,11 @@ void mdev_bus_unregister(void); struct > > mdev_parent > > > { > > > struct device *dev; > > > const struct mdev_parent_ops *ops; > > > - struct kref ref; > > > + /* Protects unregistration to wait until create/remove > > > + * are completed. > > > + */ > > > + refcount_t refcount; > > > + struct completion unreg_completion; > > > struct list_head next; > > > struct kset *mdev_types_kset; > > > struct list_head type_list; > > > > I think what's really needed is to split up the different needs and not > > overload the 'refcount' concept. > > > Refcount tells that how many active references are present for this parent device. > Those active reference could be create/remove processes and mdev core itself. > > So when parent unregisters, mdev module publishes that it is going away through this refcount. > Hence new users cannot start. But it does not actually do that -- if there are other create/remove operations running, userspace can still trigger a new create/remove. If it triggers enough create/remove processes, it can keep the parent around (even though that really is a pathological case.) > > > - If we need to make sure that a reference to the parent is held so > > that the parent may not go away while still in use, we should > > continue to use the kref (in the idiomatic way it is used before this > > patch.) > > - We need to protect against creation of new devices if the parent is > > going away. Maybe set a going_away marker in the parent structure for > > that so that creation bails out immediately? > Such marker will help to not start new processes. > So an additional marker can be added to improve mdev_try_get_parent(). > But I couldn't justify why differentiating those two users on time scale is desired. > One reason could be that user continuously tries to create mdev and parent never gets a chance, to unregister? > I guess, parent will run out mdev devices before this can happen. They can also run remove tasks in parallel (see above). > > Additionally a stop marker is needed (counter) to tell that all users are done accessing it. > Both purposes are served using a refcount scheme. Why not stop new create/remove tasks on remove, and do the final cleanup asynchronously? I think a refcount is fine to track accesses, but not to block new tasks. > > > What happens if the > > creation has already started when parent removal kicks in, though? > That particular creation will succeed but newer cannot start, because mdev_put_parent() is done. > > > Do we need some child list locking and an indication whether a child > > is in progress of being registered/unregistered? > > - We also need to protect against removal of devices while unregister > > is in progress (same mechanism as above?) The second issue you > > describe above should be fixed then if the children keep a reference > > of the parent. > Parent unregistration publishes that its going away first, so no new device removal from user can start. I don't think that this actually works as intended (see above). > Already on going removal by users anyway complete first. > > Once all remove() users are done, parent is getting unregistered.
> -----Original Message----- > From: Cornelia Huck <cohuck@redhat.com> > Sent: Monday, May 20, 2019 6:29 AM > To: Parav Pandit <parav@mellanox.com> > Cc: kvm@vger.kernel.org; linux-kernel@vger.kernel.org; > kwankhede@nvidia.com; alex.williamson@redhat.com; cjia@nvidia.com > Subject: Re: [PATCHv3 3/3] vfio/mdev: Synchronize device create/remove > with parent removal > > On Fri, 17 May 2019 14:18:26 +0000 > Parav Pandit <parav@mellanox.com> wrote: > > > > > @@ -206,14 +214,27 @@ void mdev_unregister_device(struct device > *dev) > > > > dev_info(dev, "MDEV: Unregistering\n"); > > > > > > > > list_del(&parent->next); > > > > + mutex_unlock(&parent_list_lock); > > > > + > > > > + /* Release the initial reference so that new create cannot start */ > > > > + mdev_put_parent(parent); > > > > > > The comment is confusing: We do drop one reference, but this does > > > not imply we're going to 0 (which would be the one thing that would > > > block creating new devices). > > > > > Ok. How about below comment. > > /* Balance with initial reference init */ > > Well, 'release the initial reference' is fine; it's just the second part that is > confusing. > > One thing that continues to irk me (and I'm sorry if I sound like a broken > record) is that you give up the initial reference and then continue to use > parent. For the more usual semantics of a reference count, that would be a > bug (as the structure would be freed if the reference count dropped to zero), > even though it is not a bug here. > Well, refcount cannot drop to zero if user is using it. But I understand that mdev_device caches it the parent in it, and hence it uses it. However, mdev_device child devices are terminated first when parent goes away, ensuring that no more parent user is active. So as you mentioned, its not a bug here. > > > > > > + > > > > + /* > > > > + * Wait for all the create and remove references to drop. > > > > + */ > > > > + wait_for_completion(&parent->unreg_completion); > > > > > > It only reaches 0 after this wait. > > > > > Yes. > > > > > > + > > > > + /* > > > > + * New references cannot be taken and all users are done > > > > + * using the parent. So it is safe to unregister parent. > > > > + */ > > > > class_compat_remove_link(mdev_bus_compat_class, dev, NULL); > > > > > > > > device_for_each_child(dev, NULL, mdev_device_remove_cb); > > > > > > > > parent_remove_sysfs_files(parent); > > > > - > > > > - mutex_unlock(&parent_list_lock); > > > > - mdev_put_parent(parent); > > > > + kfree(parent); > > > > + put_device(dev); > > > > } > > > > EXPORT_SYMBOL(mdev_unregister_device); > > > > > > > > @@ -237,10 +258,11 @@ int mdev_device_create(struct kobject *kobj, > > > > struct mdev_parent *parent; > > > > struct mdev_type *type = to_mdev_type(kobj); > > > > > > > > - parent = mdev_get_parent(type->parent); > > > > - if (!parent) > > > > + if (!mdev_try_get_parent(type->parent)) > > > > > > If other calls are still running, the refcount won't be 0, and this > > > will succeed, even if we really want to get rid of the device. > > > > > Sure, if other calls are running, refcount won't be 0. Process creating them > will eventually complete, and refcount will drop to zero. > > And new processes won't be able to start any more. > > So there is no differentiation between 'already in creation stage' and > 'about to start' processes. > > Does it really make sense to allow creation to start if the parent is going > away? > Its really a small time window, on how we draw the line. But it has important note that if user continues to keep creating, removing, parent is blocked on removal. > > > > > > return -EINVAL; > > > > > > > > + parent = type->parent; > > > > + > > > > mutex_lock(&mdev_list_lock); > > > > > > > > /* Check for duplicate */ > > > > @@ -287,6 +309,7 @@ int mdev_device_create(struct kobject *kobj, > > > > > > > > mdev->active = true; > > > > dev_dbg(&mdev->dev, "MDEV: created\n"); > > > > + mdev_put_parent(parent); > > > > > > > > return 0; > > > > > > > > @@ -306,7 +329,6 @@ int mdev_device_remove(struct device *dev) > > > > struct mdev_device *mdev, *tmp; > > > > struct mdev_parent *parent; > > > > struct mdev_type *type; > > > > - int ret; > > > > > > > > mdev = to_mdev_device(dev); > > > > > > > > @@ -330,15 +352,17 @@ int mdev_device_remove(struct device *dev) > > > > mutex_unlock(&mdev_list_lock); > > > > > > > > type = to_mdev_type(mdev->type_kobj); > > > > - mdev_remove_sysfs_files(dev, type); > > > > - device_del(&mdev->dev); > > > > - parent = mdev->parent; > > > > - ret = parent->ops->remove(mdev); > > > > - if (ret) > > > > - dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); > > > > + if (!mdev_try_get_parent(type->parent)) { > > > > > > Same here: Is there really a guarantee that the refcount is 0 when > > > the parent is going away? > > A WARN_ON after wait_for_completion or in freeing the parent is good to > catch bugs. > > I'd rather prefer to avoid having to add WARN_ONs :) > > This looks like it is supposed to be an early exit. remove() is doing early exit if it doesn't get reference to its parent. mdev_device_remove_common(). > However, if some > other thread does any create or remove operation at the same time, > we'll still do the remove, and we still might have have a race window > (and this is getting really hard to follow in the code). > Which part? We have only 4 functions to follow, register_device(), unregister_device(), create() and remove(). If you meant, two removes racing with each other? If so, that is currently guarded using not_so_well_defined active flag. I will cleanup that later once this series is done. > > > > > > > > > + /* > > > > + * Parent unregistration have started. > > > > + * No need to remove here. > > > > + */ > > > > + mutex_unlock(&mdev_list_lock); > > > > > > Btw., you already unlocked above. > > > > > You are right. This unlock is wrong. I will revise the patch. > > > > > > + return -ENODEV; > > > > + } > > > > > > > > - /* Balances with device_initialize() */ > > > > - put_device(&mdev->dev); > > > > + parent = mdev->parent; > > > > + mdev_device_remove_common(mdev); > > > > mdev_put_parent(parent); > > > > > > > > return 0; > > > > diff --git a/drivers/vfio/mdev/mdev_private.h > > > > b/drivers/vfio/mdev/mdev_private.h > > > > index 924ed2274941..55ebab0af7b0 100644 > > > > --- a/drivers/vfio/mdev/mdev_private.h > > > > +++ b/drivers/vfio/mdev/mdev_private.h > > > > @@ -19,7 +19,11 @@ void mdev_bus_unregister(void); struct > > > mdev_parent > > > > { > > > > struct device *dev; > > > > const struct mdev_parent_ops *ops; > > > > - struct kref ref; > > > > + /* Protects unregistration to wait until create/remove > > > > + * are completed. > > > > + */ > > > > + refcount_t refcount; > > > > + struct completion unreg_completion; > > > > struct list_head next; > > > > struct kset *mdev_types_kset; > > > > struct list_head type_list; > > > > > > I think what's really needed is to split up the different needs and not > > > overload the 'refcount' concept. > > > > > Refcount tells that how many active references are present for this parent > device. > > Those active reference could be create/remove processes and mdev core > itself. > > > > So when parent unregisters, mdev module publishes that it is going away > through this refcount. > > Hence new users cannot start. > > But it does not actually do that -- if there are other create/remove > operations running, userspace can still trigger a new create/remove. If > it triggers enough create/remove processes, it can keep the parent > around (even though that really is a pathological case.) > Yes. I agree that is still possible. And an extra flag can guard it. I see it as try_get_parent() can be improved as incremental to implement and honor that flag. Do you want to roll that flag in same patch in v4? > > > > > - If we need to make sure that a reference to the parent is held so > > > that the parent may not go away while still in use, we should > > > continue to use the kref (in the idiomatic way it is used before this > > > patch.) > > > - We need to protect against creation of new devices if the parent is > > > going away. Maybe set a going_away marker in the parent structure for > > > that so that creation bails out immediately? > > Such marker will help to not start new processes. > > So an additional marker can be added to improve mdev_try_get_parent(). > > But I couldn't justify why differentiating those two users on time scale is > desired. > > One reason could be that user continuously tries to create mdev and > parent never gets a chance, to unregister? > > I guess, parent will run out mdev devices before this can happen. > > They can also run remove tasks in parallel (see above). > Yes, remove() is guarded using active flag. > > > > Additionally a stop marker is needed (counter) to tell that all users are > done accessing it. > > Both purposes are served using a refcount scheme. > > Why not stop new create/remove tasks on remove, and do the final > cleanup asynchronously? I think a refcount is fine to track accesses, > but not to block new tasks. > So a new flag will guard new create/remove tasks by enhancing try_get_parent(). I just didn't see it as critical fix, but it's doable. See above. Async is certainly not a good idea. mdev_release_parent() in current code doesn't nothing other than freeing memory and parent reference. It take away the parent from the list early on, which is also wrong, because it was added to the list at the end. Unregister() sequence should be mirror image. Parent device files has to be removed before unregister_device() finishes, because they were added in register_device(). Otherwise, parent device_del() might be done, but files are still created under it. If we want to keep the memory around of parent, until kref drops, than we need two refcounts. One ensure that create and remove are done using it, other one that ensures that child are done using it. I fail to justify adding complexity of two counters, because such two_counter_desire hints that somehow child devices may be still active even after remove() calls are finished. And that should not be the case. Unless I miss such case. > > > > > What happens if the > > > creation has already started when parent removal kicks in, though? > > That particular creation will succeed but newer cannot start, because > mdev_put_parent() is done. > > > > > Do we need some child list locking and an indication whether a child > > > is in progress of being registered/unregistered? > > > - We also need to protect against removal of devices while unregister > > > is in progress (same mechanism as above?) The second issue you > > > describe above should be fixed then if the children keep a reference > > > of the parent. > > Parent unregistration publishes that its going away first, so no new device > removal from user can start. > > I don't think that this actually works as intended (see above). > It does work in most cases. Only if user space is creating hundreds of processes for creating mdevs, before they actually run out of creating new one. But as we talked a flag will guard it. So if refcount is ok, I can enhance it for flag. > > Already on going removal by users anyway complete first. > > > > Once all remove() users are done, parent is getting unregistered.
On Mon, 20 May 2019 19:15:15 +0000 Parav Pandit <parav@mellanox.com> wrote: > > -----Original Message----- > > From: Cornelia Huck <cohuck@redhat.com> > > Sent: Monday, May 20, 2019 6:29 AM > > To: Parav Pandit <parav@mellanox.com> > > Cc: kvm@vger.kernel.org; linux-kernel@vger.kernel.org; > > kwankhede@nvidia.com; alex.williamson@redhat.com; cjia@nvidia.com > > Subject: Re: [PATCHv3 3/3] vfio/mdev: Synchronize device create/remove > > with parent removal > > > > On Fri, 17 May 2019 14:18:26 +0000 > > Parav Pandit <parav@mellanox.com> wrote: > > > > > > > @@ -206,14 +214,27 @@ void mdev_unregister_device(struct device > > *dev) > > > > > dev_info(dev, "MDEV: Unregistering\n"); > > > > > > > > > > list_del(&parent->next); > > > > > + mutex_unlock(&parent_list_lock); > > > > > + > > > > > + /* Release the initial reference so that new create cannot start */ > > > > > + mdev_put_parent(parent); > > > > > > > > The comment is confusing: We do drop one reference, but this does > > > > not imply we're going to 0 (which would be the one thing that would > > > > block creating new devices). > > > > > > > Ok. How about below comment. > > > /* Balance with initial reference init */ > > > > Well, 'release the initial reference' is fine; it's just the second part that is > > confusing. > > > > One thing that continues to irk me (and I'm sorry if I sound like a broken > > record) is that you give up the initial reference and then continue to use > > parent. For the more usual semantics of a reference count, that would be a > > bug (as the structure would be freed if the reference count dropped to zero), > > even though it is not a bug here. > > > Well, refcount cannot drop to zero if user is using it. > But I understand that mdev_device caches it the parent in it, and hence it uses it. > However, mdev_device child devices are terminated first when parent goes away, ensuring that no more parent user is active. > So as you mentioned, its not a bug here. > > > > > > > > > + > > > > > + /* > > > > > + * Wait for all the create and remove references to drop. > > > > > + */ > > > > > + wait_for_completion(&parent->unreg_completion); > > > > > > > > It only reaches 0 after this wait. > > > > > > > Yes. > > > > > > > > + > > > > > + /* > > > > > + * New references cannot be taken and all users are done > > > > > + * using the parent. So it is safe to unregister parent. > > > > > + */ > > > > > class_compat_remove_link(mdev_bus_compat_class, dev, NULL); > > > > > > > > > > device_for_each_child(dev, NULL, mdev_device_remove_cb); > > > > > > > > > > parent_remove_sysfs_files(parent); > > > > > - > > > > > - mutex_unlock(&parent_list_lock); > > > > > - mdev_put_parent(parent); > > > > > + kfree(parent); > > > > > + put_device(dev); > > > > > } > > > > > EXPORT_SYMBOL(mdev_unregister_device); > > > > > > > > > > @@ -237,10 +258,11 @@ int mdev_device_create(struct kobject *kobj, > > > > > struct mdev_parent *parent; > > > > > struct mdev_type *type = to_mdev_type(kobj); > > > > > > > > > > - parent = mdev_get_parent(type->parent); > > > > > - if (!parent) > > > > > + if (!mdev_try_get_parent(type->parent)) > > > > > > > > If other calls are still running, the refcount won't be 0, and this > > > > will succeed, even if we really want to get rid of the device. > > > > > > > Sure, if other calls are running, refcount won't be 0. Process creating them > > will eventually complete, and refcount will drop to zero. > > > And new processes won't be able to start any more. > > > So there is no differentiation between 'already in creation stage' and > > 'about to start' processes. > > > > Does it really make sense to allow creation to start if the parent is going > > away? > > > Its really a small time window, on how we draw the line. > But it has important note that if user continues to keep creating, removing, parent is blocked on removal. > > > > > > > > > return -EINVAL; > > > > > > > > > > + parent = type->parent; > > > > > + > > > > > mutex_lock(&mdev_list_lock); > > > > > > > > > > /* Check for duplicate */ > > > > > @@ -287,6 +309,7 @@ int mdev_device_create(struct kobject *kobj, > > > > > > > > > > mdev->active = true; > > > > > dev_dbg(&mdev->dev, "MDEV: created\n"); > > > > > + mdev_put_parent(parent); > > > > > > > > > > return 0; > > > > > > > > > > @@ -306,7 +329,6 @@ int mdev_device_remove(struct device *dev) > > > > > struct mdev_device *mdev, *tmp; > > > > > struct mdev_parent *parent; > > > > > struct mdev_type *type; > > > > > - int ret; > > > > > > > > > > mdev = to_mdev_device(dev); > > > > > > > > > > @@ -330,15 +352,17 @@ int mdev_device_remove(struct device *dev) > > > > > mutex_unlock(&mdev_list_lock); > > > > > > > > > > type = to_mdev_type(mdev->type_kobj); > > > > > - mdev_remove_sysfs_files(dev, type); > > > > > - device_del(&mdev->dev); > > > > > - parent = mdev->parent; > > > > > - ret = parent->ops->remove(mdev); > > > > > - if (ret) > > > > > - dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); > > > > > + if (!mdev_try_get_parent(type->parent)) { > > > > > > > > Same here: Is there really a guarantee that the refcount is 0 when > > > > the parent is going away? > > > A WARN_ON after wait_for_completion or in freeing the parent is good to > > catch bugs. > > > > I'd rather prefer to avoid having to add WARN_ONs :) > > > > This looks like it is supposed to be an early exit. > remove() is doing early exit if it doesn't get reference to its parent. > mdev_device_remove_common(). > > > However, if some > > other thread does any create or remove operation at the same time, > > we'll still do the remove, and we still might have have a race window > > (and this is getting really hard to follow in the code). > > > Which part? > We have only 4 functions to follow, register_device(), unregister_device(), create() and remove(). > > If you meant, two removes racing with each other? > If so, that is currently guarded using not_so_well_defined active flag. > I will cleanup that later once this series is done. > > > > > > > > > > > > > + /* > > > > > + * Parent unregistration have started. > > > > > + * No need to remove here. > > > > > + */ > > > > > + mutex_unlock(&mdev_list_lock); > > > > > > > > Btw., you already unlocked above. > > > > > > > You are right. This unlock is wrong. I will revise the patch. > > > > > > > > + return -ENODEV; > > > > > + } > > > > > > > > > > - /* Balances with device_initialize() */ > > > > > - put_device(&mdev->dev); > > > > > + parent = mdev->parent; > > > > > + mdev_device_remove_common(mdev); > > > > > mdev_put_parent(parent); > > > > > > > > > > return 0; > > > > > diff --git a/drivers/vfio/mdev/mdev_private.h > > > > > b/drivers/vfio/mdev/mdev_private.h > > > > > index 924ed2274941..55ebab0af7b0 100644 > > > > > --- a/drivers/vfio/mdev/mdev_private.h > > > > > +++ b/drivers/vfio/mdev/mdev_private.h > > > > > @@ -19,7 +19,11 @@ void mdev_bus_unregister(void); struct > > > > mdev_parent > > > > > { > > > > > struct device *dev; > > > > > const struct mdev_parent_ops *ops; > > > > > - struct kref ref; > > > > > + /* Protects unregistration to wait until create/remove > > > > > + * are completed. > > > > > + */ > > > > > + refcount_t refcount; > > > > > + struct completion unreg_completion; > > > > > struct list_head next; > > > > > struct kset *mdev_types_kset; > > > > > struct list_head type_list; > > > > > > > > I think what's really needed is to split up the different needs and not > > > > overload the 'refcount' concept. > > > > > > > Refcount tells that how many active references are present for this parent > > device. > > > Those active reference could be create/remove processes and mdev core > > itself. > > > > > > So when parent unregisters, mdev module publishes that it is going away > > through this refcount. > > > Hence new users cannot start. > > > > But it does not actually do that -- if there are other create/remove > > operations running, userspace can still trigger a new create/remove. If > > it triggers enough create/remove processes, it can keep the parent > > around (even though that really is a pathological case.) > > > Yes. I agree that is still possible. And an extra flag can guard it. > I see it as try_get_parent() can be improved as incremental to implement and honor that flag. > Do you want to roll that flag in same patch in v4? > > > > > > > > - If we need to make sure that a reference to the parent is held so > > > > that the parent may not go away while still in use, we should > > > > continue to use the kref (in the idiomatic way it is used before this > > > > patch.) > > > > - We need to protect against creation of new devices if the parent is > > > > going away. Maybe set a going_away marker in the parent structure for > > > > that so that creation bails out immediately? > > > Such marker will help to not start new processes. > > > So an additional marker can be added to improve mdev_try_get_parent(). > > > But I couldn't justify why differentiating those two users on time scale is > > desired. > > > One reason could be that user continuously tries to create mdev and > > parent never gets a chance, to unregister? > > > I guess, parent will run out mdev devices before this can happen. > > > > They can also run remove tasks in parallel (see above). > > > Yes, remove() is guarded using active flag. > > > > > > > Additionally a stop marker is needed (counter) to tell that all users are > > done accessing it. > > > Both purposes are served using a refcount scheme. > > > > Why not stop new create/remove tasks on remove, and do the final > > cleanup asynchronously? I think a refcount is fine to track accesses, > > but not to block new tasks. > > > So a new flag will guard new create/remove tasks by enhancing try_get_parent(). > I just didn't see it as critical fix, but it's doable. See above. > > Async is certainly not a good idea. > mdev_release_parent() in current code doesn't nothing other than freeing memory and parent reference. > It take away the parent from the list early on, which is also wrong, because it was added to the list at the end. > Unregister() sequence should be mirror image. > Parent device files has to be removed before unregister_device() finishes, because they were added in register_device(). > Otherwise, parent device_del() might be done, but files are still created under it. > > If we want to keep the memory around of parent, until kref drops, than we need two refcounts. > One ensure that create and remove are done using it, other one that ensures that child are done using it. > I fail to justify adding complexity of two counters, because such two_counter_desire hints that somehow child devices may be still active even after remove() calls are finished. > And that should not be the case. Unless I miss such case. > > > > > > > > What happens if the > > > > creation has already started when parent removal kicks in, though? > > > That particular creation will succeed but newer cannot start, because > > mdev_put_parent() is done. > > > > > > > Do we need some child list locking and an indication whether a child > > > > is in progress of being registered/unregistered? > > > > - We also need to protect against removal of devices while unregister > > > > is in progress (same mechanism as above?) The second issue you > > > > describe above should be fixed then if the children keep a reference > > > > of the parent. > > > Parent unregistration publishes that its going away first, so no new device > > removal from user can start. > > > > I don't think that this actually works as intended (see above). > > > It does work in most cases. Only if user space is creating hundreds of processes for creating mdevs, before they actually run out of creating new one. > But as we talked a flag will guard it. > > So if refcount is ok, I can enhance it for flag. I agree with Connie's dislike of the refcount, where it seems we're really just using it as a read-writer lock. So why not simply use a rwsem? The parent unregistration path would do a down_write() and all the ancillary paths would do a down_read_trylock() as they should never see read contention unless the parent is being removed. As a bonus, we don't need to invent our own fairness algorithm, nor do we need to remove the krefs as they're at least useful to validate we haven't missed anyone. Thanks, Alex
Hi Alex, I was on travel for last 3 days, hence the slow response. Started working now. Please see inline response below. > -----Original Message----- > From: Alex Williamson <alex.williamson@redhat.com> > Sent: Tuesday, May 21, 2019 3:42 AM > To: Parav Pandit <parav@mellanox.com> > Cc: Cornelia Huck <cohuck@redhat.com>; kvm@vger.kernel.org; linux- > kernel@vger.kernel.org; kwankhede@nvidia.com; cjia@nvidia.com > Subject: Re: [PATCHv3 3/3] vfio/mdev: Synchronize device create/remove with > parent removal > > On Mon, 20 May 2019 19:15:15 +0000 > Parav Pandit <parav@mellanox.com> wrote: > > > > -----Original Message----- > > > From: Cornelia Huck <cohuck@redhat.com> > > > Sent: Monday, May 20, 2019 6:29 AM > > > To: Parav Pandit <parav@mellanox.com> > > > Cc: kvm@vger.kernel.org; linux-kernel@vger.kernel.org; > > > kwankhede@nvidia.com; alex.williamson@redhat.com; cjia@nvidia.com > > > Subject: Re: [PATCHv3 3/3] vfio/mdev: Synchronize device > > > create/remove with parent removal > > > > > > On Fri, 17 May 2019 14:18:26 +0000 > > > Parav Pandit <parav@mellanox.com> wrote: > > > > > > > > > @@ -206,14 +214,27 @@ void mdev_unregister_device(struct > > > > > > device > > > *dev) > > > > > > dev_info(dev, "MDEV: Unregistering\n"); > > > > > > > > > > > > list_del(&parent->next); > > > > > > + mutex_unlock(&parent_list_lock); > > > > > > + > > > > > > + /* Release the initial reference so that new create cannot start > */ > > > > > > + mdev_put_parent(parent); > > > > > > > > > > The comment is confusing: We do drop one reference, but this > > > > > does not imply we're going to 0 (which would be the one thing > > > > > that would block creating new devices). > > > > > > > > > Ok. How about below comment. > > > > /* Balance with initial reference init */ > > > > > > Well, 'release the initial reference' is fine; it's just the second > > > part that is confusing. > > > > > > One thing that continues to irk me (and I'm sorry if I sound like a > > > broken > > > record) is that you give up the initial reference and then continue > > > to use parent. For the more usual semantics of a reference count, > > > that would be a bug (as the structure would be freed if the > > > reference count dropped to zero), even though it is not a bug here. > > > > > Well, refcount cannot drop to zero if user is using it. > > But I understand that mdev_device caches it the parent in it, and hence it > uses it. > > However, mdev_device child devices are terminated first when parent goes > away, ensuring that no more parent user is active. > > So as you mentioned, its not a bug here. > > > > > > > > > > > > + > > > > > > + /* > > > > > > + * Wait for all the create and remove references to drop. > > > > > > + */ > > > > > > + wait_for_completion(&parent->unreg_completion); > > > > > > > > > > It only reaches 0 after this wait. > > > > > > > > > Yes. > > > > > > > > > > + > > > > > > + /* > > > > > > + * New references cannot be taken and all users are done > > > > > > + * using the parent. So it is safe to unregister parent. > > > > > > + */ > > > > > > class_compat_remove_link(mdev_bus_compat_class, dev, > NULL); > > > > > > > > > > > > device_for_each_child(dev, NULL, mdev_device_remove_cb); > > > > > > > > > > > > parent_remove_sysfs_files(parent); > > > > > > - > > > > > > - mutex_unlock(&parent_list_lock); > > > > > > - mdev_put_parent(parent); > > > > > > + kfree(parent); > > > > > > + put_device(dev); > > > > > > } > > > > > > EXPORT_SYMBOL(mdev_unregister_device); > > > > > > > > > > > > @@ -237,10 +258,11 @@ int mdev_device_create(struct kobject > *kobj, > > > > > > struct mdev_parent *parent; > > > > > > struct mdev_type *type = to_mdev_type(kobj); > > > > > > > > > > > > - parent = mdev_get_parent(type->parent); > > > > > > - if (!parent) > > > > > > + if (!mdev_try_get_parent(type->parent)) > > > > > > > > > > If other calls are still running, the refcount won't be 0, and > > > > > this will succeed, even if we really want to get rid of the device. > > > > > > > > > Sure, if other calls are running, refcount won't be 0. Process > > > > creating them > > > will eventually complete, and refcount will drop to zero. > > > > And new processes won't be able to start any more. > > > > So there is no differentiation between 'already in creation stage' > > > > and > > > 'about to start' processes. > > > > > > Does it really make sense to allow creation to start if the parent > > > is going away? > > > > > Its really a small time window, on how we draw the line. > > But it has important note that if user continues to keep creating, removing, > parent is blocked on removal. > > > > > > > > > > > > return -EINVAL; > > > > > > > > > > > > + parent = type->parent; > > > > > > + > > > > > > mutex_lock(&mdev_list_lock); > > > > > > > > > > > > /* Check for duplicate */ > > > > > > @@ -287,6 +309,7 @@ int mdev_device_create(struct kobject > > > > > > *kobj, > > > > > > > > > > > > mdev->active = true; > > > > > > dev_dbg(&mdev->dev, "MDEV: created\n"); > > > > > > + mdev_put_parent(parent); > > > > > > > > > > > > return 0; > > > > > > > > > > > > @@ -306,7 +329,6 @@ int mdev_device_remove(struct device *dev) > > > > > > struct mdev_device *mdev, *tmp; > > > > > > struct mdev_parent *parent; > > > > > > struct mdev_type *type; > > > > > > - int ret; > > > > > > > > > > > > mdev = to_mdev_device(dev); > > > > > > > > > > > > @@ -330,15 +352,17 @@ int mdev_device_remove(struct device > *dev) > > > > > > mutex_unlock(&mdev_list_lock); > > > > > > > > > > > > type = to_mdev_type(mdev->type_kobj); > > > > > > - mdev_remove_sysfs_files(dev, type); > > > > > > - device_del(&mdev->dev); > > > > > > - parent = mdev->parent; > > > > > > - ret = parent->ops->remove(mdev); > > > > > > - if (ret) > > > > > > - dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); > > > > > > + if (!mdev_try_get_parent(type->parent)) { > > > > > > > > > > Same here: Is there really a guarantee that the refcount is 0 > > > > > when the parent is going away? > > > > A WARN_ON after wait_for_completion or in freeing the parent is > > > > good to > > > catch bugs. > > > > > > I'd rather prefer to avoid having to add WARN_ONs :) > > > > > > This looks like it is supposed to be an early exit. > > remove() is doing early exit if it doesn't get reference to its parent. > > mdev_device_remove_common(). > > > > > However, if some > > > other thread does any create or remove operation at the same time, > > > we'll still do the remove, and we still might have have a race > > > window (and this is getting really hard to follow in the code). > > > > > Which part? > > We have only 4 functions to follow, register_device(), unregister_device(), > create() and remove(). > > > > If you meant, two removes racing with each other? > > If so, that is currently guarded using not_so_well_defined active flag. > > I will cleanup that later once this series is done. > > > > > > > > > > > > > > > > > + /* > > > > > > + * Parent unregistration have started. > > > > > > + * No need to remove here. > > > > > > + */ > > > > > > + mutex_unlock(&mdev_list_lock); > > > > > > > > > > Btw., you already unlocked above. > > > > > > > > > You are right. This unlock is wrong. I will revise the patch. > > > > > > > > > > + return -ENODEV; > > > > > > + } > > > > > > > > > > > > - /* Balances with device_initialize() */ > > > > > > - put_device(&mdev->dev); > > > > > > + parent = mdev->parent; > > > > > > + mdev_device_remove_common(mdev); > > > > > > mdev_put_parent(parent); > > > > > > > > > > > > return 0; > > > > > > diff --git a/drivers/vfio/mdev/mdev_private.h > > > > > > b/drivers/vfio/mdev/mdev_private.h > > > > > > index 924ed2274941..55ebab0af7b0 100644 > > > > > > --- a/drivers/vfio/mdev/mdev_private.h > > > > > > +++ b/drivers/vfio/mdev/mdev_private.h > > > > > > @@ -19,7 +19,11 @@ void mdev_bus_unregister(void); struct > > > > > mdev_parent > > > > > > { > > > > > > struct device *dev; > > > > > > const struct mdev_parent_ops *ops; > > > > > > - struct kref ref; > > > > > > + /* Protects unregistration to wait until create/remove > > > > > > + * are completed. > > > > > > + */ > > > > > > + refcount_t refcount; > > > > > > + struct completion unreg_completion; > > > > > > struct list_head next; > > > > > > struct kset *mdev_types_kset; > > > > > > struct list_head type_list; > > > > > > > > > > I think what's really needed is to split up the different needs > > > > > and not overload the 'refcount' concept. > > > > > > > > > Refcount tells that how many active references are present for > > > > this parent > > > device. > > > > Those active reference could be create/remove processes and mdev > > > > core > > > itself. > > > > > > > > So when parent unregisters, mdev module publishes that it is going > > > > away > > > through this refcount. > > > > Hence new users cannot start. > > > > > > But it does not actually do that -- if there are other create/remove > > > operations running, userspace can still trigger a new create/remove. > > > If it triggers enough create/remove processes, it can keep the > > > parent around (even though that really is a pathological case.) > > > > > Yes. I agree that is still possible. And an extra flag can guard it. > > I see it as try_get_parent() can be improved as incremental to implement and > honor that flag. > > Do you want to roll that flag in same patch in v4? > > > > > > > > > > > - If we need to make sure that a reference to the parent is held so > > > > > that the parent may not go away while still in use, we should > > > > > continue to use the kref (in the idiomatic way it is used before this > > > > > patch.) > > > > > - We need to protect against creation of new devices if the parent is > > > > > going away. Maybe set a going_away marker in the parent structure > for > > > > > that so that creation bails out immediately? > > > > Such marker will help to not start new processes. > > > > So an additional marker can be added to improve mdev_try_get_parent(). > > > > But I couldn't justify why differentiating those two users on time > > > > scale is > > > desired. > > > > One reason could be that user continuously tries to create mdev > > > > and > > > parent never gets a chance, to unregister? > > > > I guess, parent will run out mdev devices before this can happen. > > > > > > They can also run remove tasks in parallel (see above). > > > > > Yes, remove() is guarded using active flag. > > > > > > > > > > Additionally a stop marker is needed (counter) to tell that all > > > > users are > > > done accessing it. > > > > Both purposes are served using a refcount scheme. > > > > > > Why not stop new create/remove tasks on remove, and do the final > > > cleanup asynchronously? I think a refcount is fine to track > > > accesses, but not to block new tasks. > > > > > So a new flag will guard new create/remove tasks by enhancing > try_get_parent(). > > I just didn't see it as critical fix, but it's doable. See above. > > > > Async is certainly not a good idea. > > mdev_release_parent() in current code doesn't nothing other than freeing > memory and parent reference. > > It take away the parent from the list early on, which is also wrong, because it > was added to the list at the end. > > Unregister() sequence should be mirror image. > > Parent device files has to be removed before unregister_device() finishes, > because they were added in register_device(). > > Otherwise, parent device_del() might be done, but files are still created > under it. > > > > If we want to keep the memory around of parent, until kref drops, than we > need two refcounts. > > One ensure that create and remove are done using it, other one that ensures > that child are done using it. > > I fail to justify adding complexity of two counters, because such > two_counter_desire hints that somehow child devices may be still active even > after remove() calls are finished. > > And that should not be the case. Unless I miss such case. > > > > > > > > > > > What happens if the > > > > > creation has already started when parent removal kicks in, though? > > > > That particular creation will succeed but newer cannot start, > > > > because > > > mdev_put_parent() is done. > > > > > > > > > Do we need some child list locking and an indication whether a child > > > > > is in progress of being registered/unregistered? > > > > > - We also need to protect against removal of devices while unregister > > > > > is in progress (same mechanism as above?) The second issue you > > > > > describe above should be fixed then if the children keep a reference > > > > > of the parent. > > > > Parent unregistration publishes that its going away first, so no > > > > new device > > > removal from user can start. > > > > > > I don't think that this actually works as intended (see above). > > > > > It does work in most cases. Only if user space is creating hundreds of > processes for creating mdevs, before they actually run out of creating new one. > > But as we talked a flag will guard it. > > > > So if refcount is ok, I can enhance it for flag. > > I agree with Connie's dislike of the refcount, where it seems we're really just > using it as a read-writer lock. So why not simply use a rwsem? The parent > unregistration path would do a down_write() and all the ancillary paths would > do a down_read_trylock() as they should never see read contention unless the > parent is being removed. Ok. sounds good. I will send v4 using rwsem without removing kref. > As a bonus, we don't need to invent our own fairness > algorithm, nor do we need to remove the krefs as they're at least useful to > validate we haven't missed anyone. Thanks, Well if we really care for kref, put on parent kref should be done in mdev_device_release(). I do not see how device can be still using the parent after device_del() is done. Anyways, kref cleanup is different patch in 5.3. > > Alex
diff --git a/drivers/vfio/mdev/mdev_core.c b/drivers/vfio/mdev/mdev_core.c index 0bef0cae1d4b..ca33246c1dc3 100644 --- a/drivers/vfio/mdev/mdev_core.c +++ b/drivers/vfio/mdev/mdev_core.c @@ -78,34 +78,41 @@ static struct mdev_parent *__find_parent_device(struct device *dev) return NULL; } -static void mdev_release_parent(struct kref *kref) +static bool mdev_try_get_parent(struct mdev_parent *parent) { - struct mdev_parent *parent = container_of(kref, struct mdev_parent, - ref); - struct device *dev = parent->dev; - - kfree(parent); - put_device(dev); + if (parent) + return refcount_inc_not_zero(&parent->refcount); + return false; } -static struct mdev_parent *mdev_get_parent(struct mdev_parent *parent) +static void mdev_put_parent(struct mdev_parent *parent) { - if (parent) - kref_get(&parent->ref); - - return parent; + if (parent && refcount_dec_and_test(&parent->refcount)) + complete(&parent->unreg_completion); } -static void mdev_put_parent(struct mdev_parent *parent) +static void mdev_device_remove_common(struct mdev_device *mdev) { - if (parent) - kref_put(&parent->ref, mdev_release_parent); + struct mdev_parent *parent; + struct mdev_type *type; + int ret; + + type = to_mdev_type(mdev->type_kobj); + mdev_remove_sysfs_files(&mdev->dev, type); + device_del(&mdev->dev); + parent = mdev->parent; + ret = parent->ops->remove(mdev); + if (ret) + dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); + + /* Balances with device_initialize() */ + put_device(&mdev->dev); } static int mdev_device_remove_cb(struct device *dev, void *data) { if (dev_is_mdev(dev)) - mdev_device_remove(dev); + mdev_device_remove_common(to_mdev_device(dev)); return 0; } @@ -147,7 +154,8 @@ int mdev_register_device(struct device *dev, const struct mdev_parent_ops *ops) goto add_dev_err; } - kref_init(&parent->ref); + refcount_set(&parent->refcount, 1); + init_completion(&parent->unreg_completion); parent->dev = dev; parent->ops = ops; @@ -206,14 +214,27 @@ void mdev_unregister_device(struct device *dev) dev_info(dev, "MDEV: Unregistering\n"); list_del(&parent->next); + mutex_unlock(&parent_list_lock); + + /* Release the initial reference so that new create cannot start */ + mdev_put_parent(parent); + + /* + * Wait for all the create and remove references to drop. + */ + wait_for_completion(&parent->unreg_completion); + + /* + * New references cannot be taken and all users are done + * using the parent. So it is safe to unregister parent. + */ class_compat_remove_link(mdev_bus_compat_class, dev, NULL); device_for_each_child(dev, NULL, mdev_device_remove_cb); parent_remove_sysfs_files(parent); - - mutex_unlock(&parent_list_lock); - mdev_put_parent(parent); + kfree(parent); + put_device(dev); } EXPORT_SYMBOL(mdev_unregister_device); @@ -237,10 +258,11 @@ int mdev_device_create(struct kobject *kobj, struct mdev_parent *parent; struct mdev_type *type = to_mdev_type(kobj); - parent = mdev_get_parent(type->parent); - if (!parent) + if (!mdev_try_get_parent(type->parent)) return -EINVAL; + parent = type->parent; + mutex_lock(&mdev_list_lock); /* Check for duplicate */ @@ -287,6 +309,7 @@ int mdev_device_create(struct kobject *kobj, mdev->active = true; dev_dbg(&mdev->dev, "MDEV: created\n"); + mdev_put_parent(parent); return 0; @@ -306,7 +329,6 @@ int mdev_device_remove(struct device *dev) struct mdev_device *mdev, *tmp; struct mdev_parent *parent; struct mdev_type *type; - int ret; mdev = to_mdev_device(dev); @@ -330,15 +352,17 @@ int mdev_device_remove(struct device *dev) mutex_unlock(&mdev_list_lock); type = to_mdev_type(mdev->type_kobj); - mdev_remove_sysfs_files(dev, type); - device_del(&mdev->dev); - parent = mdev->parent; - ret = parent->ops->remove(mdev); - if (ret) - dev_err(&mdev->dev, "Remove failed: err=%d\n", ret); + if (!mdev_try_get_parent(type->parent)) { + /* + * Parent unregistration have started. + * No need to remove here. + */ + mutex_unlock(&mdev_list_lock); + return -ENODEV; + } - /* Balances with device_initialize() */ - put_device(&mdev->dev); + parent = mdev->parent; + mdev_device_remove_common(mdev); mdev_put_parent(parent); return 0; diff --git a/drivers/vfio/mdev/mdev_private.h b/drivers/vfio/mdev/mdev_private.h index 924ed2274941..55ebab0af7b0 100644 --- a/drivers/vfio/mdev/mdev_private.h +++ b/drivers/vfio/mdev/mdev_private.h @@ -19,7 +19,11 @@ void mdev_bus_unregister(void); struct mdev_parent { struct device *dev; const struct mdev_parent_ops *ops; - struct kref ref; + /* Protects unregistration to wait until create/remove + * are completed. + */ + refcount_t refcount; + struct completion unreg_completion; struct list_head next; struct kset *mdev_types_kset; struct list_head type_list;
In following sequences, child devices created while removing mdev parent device can be left out, or it may lead to race of removing half initialized child mdev devices. issue-1: -------- cpu-0 cpu-1 ----- ----- mdev_unregister_device() device_for_each_child() mdev_device_remove_cb() mdev_device_remove() create_store() mdev_device_create() [...] device_add() parent_remove_sysfs_files() /* BUG: device added by cpu-0 * whose parent is getting removed * and it won't process this mdev. */ issue-2: -------- Below crash is observed when user initiated remove is in progress and mdev_unregister_driver() completes parent unregistration. cpu-0 cpu-1 ----- ----- remove_store() mdev_device_remove() active = false; mdev_unregister_device() parent device removed. [...] parents->ops->remove() /* * BUG: Accessing invalid parent. */ This is similar race like create() racing with mdev_unregister_device(). BUG: unable to handle kernel paging request at ffffffffc0585668 PGD e8f618067 P4D e8f618067 PUD e8f61a067 PMD 85adca067 PTE 0 Oops: 0000 [#1] SMP PTI CPU: 41 PID: 37403 Comm: bash Kdump: loaded Not tainted 5.1.0-rc6-vdevbus+ #6 Hardware name: Supermicro SYS-6028U-TR4+/X10DRU-i+, BIOS 2.0b 08/09/2016 RIP: 0010:mdev_device_remove+0xfa/0x140 [mdev] Call Trace: remove_store+0x71/0x90 [mdev] kernfs_fop_write+0x113/0x1a0 vfs_write+0xad/0x1b0 ksys_write+0x5a/0xe0 do_syscall_64+0x5a/0x210 entry_SYSCALL_64_after_hwframe+0x49/0xbe Therefore, mdev core is improved as below to overcome above issues. Wait for any ongoing mdev create() and remove() to finish before unregistering parent device using refcount and completion. This continues to allow multiple create and remove to progress in parallel for different mdev devices as most common case. At the same time guard parent removal while parent is being access by create() and remove callbacks. Code is simplified from kref to use refcount as unregister_device() has to wait anyway for all create/remove to finish. While removing mdev devices during parent unregistration, there isn't need to acquire refcount of parent device, hence code is restructured using mdev_device_remove_common() to avoid it. Fixes: 7b96953bc640 ("vfio: Mediated device Core driver") Signed-off-by: Parav Pandit <parav@mellanox.com> --- drivers/vfio/mdev/mdev_core.c | 86 ++++++++++++++++++++------------ drivers/vfio/mdev/mdev_private.h | 6 ++- 2 files changed, 60 insertions(+), 32 deletions(-)