nbd lifetimes fixes
Subsystems (2)
Hi Josef and Jens,
this series tries to deal with the fallout of the recent lock scope reduction as pointed out by Tetsuo and szybot and inspired by / reused from the catchall patch by Tetsuo. One big change is that I finally decided to kill off the ->destroy_complete scheme entirely because I'm pretty sure it is not needed with the various fixes and we can just return an error for the tiny race condition where it matters. Xiubo, can you double check this with your nbd-runner setup? nbd-runner itself seems pretty generic and not directly reproduce anything here.
Note that the syzbot reproduer still fails eventually, but in devtmpfsd in a way that does not look related to the loop code at all.
PATCH 1/6 nbd: add missing locking to the nbd_dev_add error path
From: Tetsuo Handa <[email protected]>
idr_remove needs external synchronization.
--- drivers/block/nbd.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/block/nbd.c +2 -0
@@ -1774,7 +1774,9 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)return nbd; out_free_idr: +mutex_lock (&nbd_index_mutex);idr_remove (&nbd_index_idr, index); +mutex_unlock (&nbd_index_mutex); out_free_tags:blk_mq_free_tag_set (&nbd->tag_set); out_free_nbd:
Signature · 1 line
2.30.2
PATCH 2/6 nbd: reset NBD to NULL when restarting in nbd_genl_connect
In reply to Christoph HellwigWhen nbd_genl_connect restarts to wait for a disconnecting device, nbd needs to be reset to NULL. Do that by facoring out a helper to find an unused device.
--- drivers/block/nbd.c | 32 ++++++++++++++++++-------------- 1 file changed, 18 insertions(+), 14 deletions(-)
drivers/block/nbd.c +18 -14
@@ -1785,6 +1785,20 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)return ERR_PTR (err); } +static struct nbd_device *nbd_find_unused (void ) +{ +struct nbd_device *nbd; +int id; + +lockdep_assert_held (&nbd_index_mutex); + +idr_for_each_entry (&nbd_index_idr, nbd, id) +if (!refcount_read (&nbd->config_refs)) +return nbd; + +return NULL ; +} +/* Netlink interface. */ static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX +1 ] = { [NBD_ATTR_INDEX ] = { .type =NLA_U32 },
@@ -1832,7 +1846,7 @@ static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)static int nbd_genl_connect (struct sk_buff *skb,struct genl_info *info) {DECLARE_COMPLETION_ONSTACK (destroy_complete); -struct nbd_device *nbd =NULL ; +struct nbd_device *nbd;struct nbd_config *config;int index =-1 ;int ret;
@@ -1853,20 +1867,10 @@ static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info) } again:mutex_lock (&nbd_index_mutex); -if (index ==-1 ) { -struct nbd_device *tmp; -int id; - -idr_for_each_entry (&nbd_index_idr, tmp, id) { -if (!refcount_read (&tmp->config_refs)) { - nbd = tmp; -break ; - } - } - }else { +if (index ==-1 ) + nbd =nbd_find_unused (); +else nbd =idr_find (&nbd_index_idr, index); - } -if (nbd) {if (test_bit (NBD_DESTROY_ON_DISCONNECT , &nbd->flags) &&test_bit (NBD_DISCONNECT_REQUESTED , &nbd->flags)) {
Signature · 1 line
2.30.2
PATCH 3/6 nbd: prevent IDR lookups from finding partially initialized devices
In reply to Christoph HellwigFrom: Tetsuo Handa <[email protected]>
Previously nbd_index_mutex was held during whole add/remove/lookup operations in order to guarantee that partially initialized devices are not reachable via idr_find() or idr_for_each(). But now that partially initialized devices become reachable as soon as idr_alloc() succeeds, we need to skip partially initialized devices. Since it seems that all functions use refcount_inc_not_zero(&nbd->refs) in order to skip destroying devices, update nbd->refs from zero to non-zero as the last step of device initialization in order to also skip partially initialized devices.
--- drivers/block/nbd.c | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-)
drivers/block/nbd.c +10 -1
@@ -1751,7 +1751,11 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)mutex_init (&nbd->config_lock);refcount_set (&nbd->config_refs,0 ); -refcount_set (&nbd->refs, refs); +/* +* Start out with a zero references to keep other threads from using +* this device until it is fully initialized. +*/ +refcount_set (&nbd->refs,0 );INIT_LIST_HEAD (&nbd->list); disk->major =NBD_MAJOR ;
@@ -1770,6 +1774,11 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs) disk->private_data = nbd;sprintf (disk->disk_name,"nbd%d" , index);add_disk (disk); + +/* +* Now publish the device. +*/ +refcount_set (&nbd->refs, refs); nbd_total_devices++;return nbd;
Signature · 1 line
2.30.2
PATCH 4/6 nbd: set nbd->index before releasing nbd_index_mutex
In reply to Christoph HellwigFrom: Tetsuo Handa <[email protected]>
Set nbd->index before releasing nbd_index_mutex, as populate_nbd_status() might access nbd->index as soon as nbd_index_mutex is released.
--- drivers/block/nbd.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/block/nbd.c +1 -1
@@ -1724,10 +1724,10 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)if (err >=0 ) index = err; } + nbd->index = index;mutex_unlock (&nbd_index_mutex);if (err <0 ) goto out_free_tags; - nbd->index = index; disk =blk_mq_alloc_disk (&nbd->tag_set,NULL );if (IS_ERR (disk)) {
Signature · 1 line
2.30.2
PATCH 5/6 nbd: only return usable devices from nbd_find_unused
In reply to Christoph HellwigDevice marked as NBD_DESTROY_ON_DISCONNECT can and should be skipped given that they won't survive the disconnect. So skip them and try to grab a reference directly and just continue if the the devices is being torn down or created and thus has a zero refcount.
--- drivers/block/nbd.c | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-)
drivers/block/nbd.c +9 -7
@@ -1794,16 +1794,20 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)return ERR_PTR (err); } -static struct nbd_device *nbd_find_unused (void ) +static struct nbd_device *nbd_find_get_unused (void ) {struct nbd_device *nbd;int id;lockdep_assert_held (&nbd_index_mutex); -idr_for_each_entry (&nbd_index_idr, nbd, id) -if (!refcount_read (&nbd->config_refs)) +idr_for_each_entry (&nbd_index_idr, nbd, id) { +if (refcount_read (&nbd->config_refs) || +test_bit (NBD_DESTROY_ON_DISCONNECT , &nbd->flags)) +continue ; +if (refcount_inc_not_zero (&nbd->refs))return nbd; + }return NULL ; }
@@ -1877,10 +1881,10 @@ static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info) again:mutex_lock (&nbd_index_mutex);if (index ==-1 ) - nbd =nbd_find_unused (); + nbd =nbd_find_get_unused ();else nbd =idr_find (&nbd_index_idr, index); -if (nbd) { +if (nbd && index !=-1 ) {if (test_bit (NBD_DESTROY_ON_DISCONNECT , &nbd->flags) &&test_bit (NBD_DISCONNECT_REQUESTED , &nbd->flags)) { nbd->destroy_complete = &destroy_complete;
@@ -1893,8 +1897,6 @@ static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)if (!refcount_inc_not_zero (&nbd->refs)) {mutex_unlock (&nbd_index_mutex); -if (index ==-1 ) - goto again;pr_err ("nbd: device at index %d is going down\n" , index);return -EINVAL ;
Signature · 1 line
2.30.2
PATCH 6/6 nbd: remove nbd->destroy_complete
In reply to Christoph HellwigThe nbd->destroy_complete pointer is not really needed. For creating a device without a specific index we now simplify skip devices marked NBD_DESTROY_ON_DISCONNECT as there is not much point to reuse them. For device creation with a specific index there is no real need to treat the case of a requested but not finished disconnect different than any other device that is being shutdown, i.e. we can just return an error, as a slightly different race window would anyway.
--- drivers/block/nbd.c | 52 ++++++++++++--------------------------------- 1 file changed, 14 insertions(+), 38 deletions(-)
drivers/block/nbd.c +14 -38
@@ -120,7 +120,6 @@ struct nbd_device {struct task_struct *task_recv;struct task_struct *task_setup; -struct completion *destroy_complete;unsigned long flags;char *backend;
@@ -235,19 +234,6 @@ static const struct device_attribute backend_attr = { .show = backend_show, }; -/* -* Place this in the last just before the nbd is freed to -* make sure that the disk and the related kobject are also -* totally removed to avoid duplicate creation of the same -* one. -*/ -static void nbd_notify_destroy_completion (struct nbd_device *nbd) -{ -if (test_bit (NBD_DESTROY_ON_DISCONNECT , &nbd->flags) && - nbd->destroy_complete) -complete (nbd->destroy_complete); -} -static void nbd_dev_remove (struct nbd_device *nbd) {struct gendisk *disk = nbd->disk;
@@ -262,7 +248,6 @@ static void nbd_dev_remove(struct nbd_device *nbd)*/ mutex_lock (&nbd_index_mutex);idr_remove (&nbd_index_idr, nbd->index); -nbd_notify_destroy_completion (nbd);mutex_unlock (&nbd_index_mutex);kfree (nbd);
@@ -1706,7 +1691,6 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)BLK_MQ_F_BLOCKING ; nbd->tag_set.driver_data = nbd;INIT_WORK (&nbd->remove_work, nbd_dev_remove_work); - nbd->destroy_complete =NULL ; nbd->backend =NULL ; err =blk_mq_alloc_tag_set (&nbd->tag_set);
@@ -1858,7 +1842,6 @@ static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)static int nbd_genl_connect (struct sk_buff *skb,struct genl_info *info) { -DECLARE_COMPLETION_ONSTACK (destroy_complete);struct nbd_device *nbd;struct nbd_config *config;int index =-1 ;
@@ -1880,31 +1863,24 @@ static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info) } again:mutex_lock (&nbd_index_mutex); -if (index ==-1 ) +if (index ==-1 ) { nbd =nbd_find_get_unused (); -else + }else { nbd =idr_find (&nbd_index_idr, index); -if (nbd && index !=-1 ) { -if (test_bit (NBD_DESTROY_ON_DISCONNECT , &nbd->flags) && -test_bit (NBD_DISCONNECT_REQUESTED , &nbd->flags)) { - nbd->destroy_complete = &destroy_complete; -mutex_unlock (&nbd_index_mutex); - -/* wait until the nbd device is completely destroyed */ -wait_for_completion (&destroy_complete); - goto again; - } - -if (!refcount_inc_not_zero (&nbd->refs)) { -mutex_unlock (&nbd_index_mutex); -pr_err ("nbd: device at index %d is going down\n" , - index); -return -EINVAL ; +if (nbd) { +if ((test_bit (NBD_DESTROY_ON_DISCONNECT , &nbd->flags) && +test_bit (NBD_DISCONNECT_REQUESTED , &nbd->flags)) || + !refcount_inc_not_zero (&nbd->refs)) { +mutex_unlock (&nbd_index_mutex); +pr_err ("nbd: device at index %d is going down\n" , + index); +return -EINVAL ; + } } -mutex_unlock (&nbd_index_mutex); - }else { -mutex_unlock (&nbd_index_mutex); + } +mutex_unlock (&nbd_index_mutex); +if (!nbd) { nbd =nbd_dev_add (index,2 );if (IS_ERR (nbd)) {pr_err ("nbd: failed to add new device\n" );
Signature · 1 line
2.30.2
Quoted 12 lines · Christoph Hellwig
Hi Josef and Jens,
this series tries to deal with the fallout of the recent lock scope reduction as pointed out by Tetsuo and szybot and inspired by / reused from the catchall patch by Tetsuo. One big change is that I finally decided to kill off the ->destroy_complete scheme entirely because I'm pretty sure it is not needed with the various fixes and we can just return an error for the tiny race condition where it matters. Xiubo, can you double check this with your nbd-runner setup? nbd-runner itself seems pretty generic and not directly reproduce anything here.
Quoted 4 lines
Note that the syzbot reproduer still fails eventually, but in devtmpfsd in a way that does not look related to the loop code at all.
Looks good to me, you can add
Thanks,
Josef
Quoted 12 lines · Christoph Hellwig
Hi Josef and Jens,
this series tries to deal with the fallout of the recent lock scope reduction as pointed out by Tetsuo and szybot and inspired by / reused from the catchall patch by Tetsuo. One big change is that I finally decided to kill off the ->destroy_complete scheme entirely because I'm pretty sure it is not needed with the various fixes and we can just return an error for the tiny race condition where it matters. Xiubo, can you double check this with your nbd-runner setup? nbd-runner itself seems pretty generic and not directly reproduce anything here.
Quoted 3 lines
Note that the syzbot reproduer still fails eventually, but in devtmpfsd in a way that does not look related to the loop code at all.
Applied, thanks.
Signature · 1 line
Jens Axboe
Quoted 12 lines · Christoph Hellwig
Hi Josef and Jens,
this series tries to deal with the fallout of the recent lock scope reduction as pointed out by Tetsuo and szybot and inspired by / reused from the catchall patch by Tetsuo. One big change is that I finally decided to kill off the ->destroy_complete scheme entirely because I'm pretty sure it is not needed with the various fixes and we can just return an error for the tiny race condition where it matters. Xiubo, can you double check this with your nbd-runner setup? nbd-runner itself seems pretty generic and not directly reproduce anything here.
Quoted 4 lines
Note that the syzbot reproduer still fails eventually, but in devtmpfsd in a way that does not look related to the loop code at all.
Sorry, I think I have missed this thread.
Test this and works well for me by using the nbd-runner.