RDMA: Use refcount_t for reference counting
cover RDMA: Use refcount_t for reference counting
There are some refcnt in type of atomic_t in RDMA subsystem, almost all of them is wrote before the refcount_t is acheived in kernel. refcount_t is better than integer for reference counting, it will WARN on overflow/underflow and avoid use-after-free risks.
Weihang Li (6):
RDMA/core: Use refcount_t instead of atomic_t for reference counting RDMA/hns: Use refcount_t instead of atomic_t for reference counting RDMA/hns: Use refcount_t APIs for HEM
RDMA/cxgb4: Use refcount_t instead of atomic_t for reference counting RDMA/i40iw: Use refcount_t instead of atomic_t for reference counting RDMA/ipoib: Use refcount_t instead of atomic_t for reference counting
drivers/infiniband/core/iwcm.c | 9 +++-- drivers/infiniband/core/iwcm.h | 2 +- drivers/infiniband/core/iwpm_util.c | 12 ++++--- drivers/infiniband/core/iwpm_util.h | 2 +- drivers/infiniband/core/mad_priv.h | 2 +- drivers/infiniband/core/multicast.c | 30 ++++++++-------- drivers/infiniband/core/uverbs.h | 2 +- drivers/infiniband/core/uverbs_main.c | 12 +++---- drivers/infiniband/hw/cxgb4/cq.c | 6 ++-- drivers/infiniband/hw/cxgb4/ev.c | 8 ++--- drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 2 +- drivers/infiniband/hw/hns/hns_roce_cq.c | 8 ++--- drivers/infiniband/hw/hns/hns_roce_device.h | 6 ++-- drivers/infiniband/hw/hns/hns_roce_hem.c | 33 +++++++++--------- drivers/infiniband/hw/hns/hns_roce_hem.h | 4 +-- drivers/infiniband/hw/hns/hns_roce_qp.c | 12 +++---- drivers/infiniband/hw/hns/hns_roce_srq.c | 8 ++--- drivers/infiniband/hw/i40iw/i40iw.h | 2 +- drivers/infiniband/hw/i40iw/i40iw_cm.c | 54 ++++++++++++++--------------- drivers/infiniband/hw/i40iw/i40iw_cm.h | 4 +-- drivers/infiniband/hw/i40iw/i40iw_main.c | 2 +- drivers/infiniband/hw/i40iw/i40iw_puda.h | 2 +- drivers/infiniband/hw/i40iw/i40iw_utils.c | 10 +++--- drivers/infiniband/ulp/ipoib/ipoib.h | 4 +-- drivers/infiniband/ulp/ipoib/ipoib_main.c | 8 ++--- 25 files changed, 123 insertions(+), 121 deletions(-)
Signature · 1 line
2.7.4
PATCH 1/6 RDMA/core: Use refcount_t instead of atomic_t for reference counting
The refcount_t API will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks. Increase refcount_t from 0 to 1 is regarded as there is a risk about use-after-free. So it should be set to 1 directly during initialization.
--- drivers/infiniband/core/iwcm.c | 9 ++++----- drivers/infiniband/core/iwcm.h | 2 +- drivers/infiniband/core/iwpm_util.c | 12 ++++++++---- drivers/infiniband/core/iwpm_util.h | 2 +- drivers/infiniband/core/mad_priv.h | 2 +- drivers/infiniband/core/multicast.c | 30 +++++++++++++++--------------- drivers/infiniband/core/uverbs.h | 2 +- drivers/infiniband/core/uverbs_main.c | 12 ++++++------ 8 files changed, 37 insertions(+), 34 deletions(-)
drivers/infiniband/core/iwcm.c +4 -5
@@ -211,8 +211,7 @@ static void free_cm_id(struct iwcm_id_private *cm_id_priv)*/ static int iwcm_deref_id (struct iwcm_id_private *cm_id_priv) { -BUG_ON (atomic_read (&cm_id_priv->refcount)==0 ); -if (atomic_dec_and_test (&cm_id_priv->refcount)) { +if (refcount_dec_and_test (&cm_id_priv->refcount)) {BUG_ON (!list_empty (&cm_id_priv->work_list));free_cm_id (cm_id_priv);return 1 ;
@@ -225,7 +224,7 @@ static void add_ref(struct iw_cm_id *cm_id) {struct iwcm_id_private *cm_id_priv; cm_id_priv = container_of(cm_id, struct iwcm_id_private, id); -atomic_inc (&cm_id_priv->refcount); +refcount_inc (&cm_id_priv->refcount); }static void rem_ref (struct iw_cm_id *cm_id)
@@ -257,7 +256,7 @@ struct iw_cm_id *iw_create_cm_id(struct ib_device *device, cm_id_priv->id.add_ref = add_ref; cm_id_priv->id.rem_ref = rem_ref;spin_lock_init (&cm_id_priv->lock); -atomic_set (&cm_id_priv->refcount,1 ); +refcount_set (&cm_id_priv->refcount,1 );init_waitqueue_head (&cm_id_priv->connect_wait);init_completion (&cm_id_priv->destroy_comp);INIT_LIST_HEAD (&cm_id_priv->work_list);
@@ -1094,7 +1093,7 @@ static int cm_event_handler(struct iw_cm_id *cm_id, } } -atomic_inc (&cm_id_priv->refcount); +refcount_inc (&cm_id_priv->refcount);if (list_empty (&cm_id_priv->work_list)) {list_add_tail (&work->list, &cm_id_priv->work_list);queue_work (iwcm_wq, &work->work);
drivers/infiniband/core/iwcm.h +1 -1
@@ -52,7 +52,7 @@ struct iwcm_id_private {wait_queue_head_t connect_wait;struct list_head work_list;spinlock_t lock; -atomic_t refcount; +refcount_t refcount;struct list_head work_free_list; };
drivers/infiniband/core/iwpm_util.c +8 -4
@@ -61,7 +61,7 @@ int iwpm_init(u8 nl_client) {int ret =0 ;mutex_lock (&iwpm_admin_lock); -if (atomic_read (&iwpm_admin.refcount) ==0 ) { +if (!refcount_read (&iwpm_admin.refcount)) { iwpm_hash_bucket =kcalloc (IWPM_MAPINFO_HASH_SIZE ,sizeof (struct hlist_head ),GFP_KERNEL );
@@ -77,8 +77,12 @@ int iwpm_init(u8 nl_client) ret = -ENOMEM ; goto init_exit; } + +refcount_set (&iwpm_admin.refcount,1 ); + }else { +refcount_inc (&iwpm_admin.refcount); } -atomic_inc (&iwpm_admin.refcount); + init_exit:mutex_unlock (&iwpm_admin_lock);if (!ret) {
@@ -105,12 +109,12 @@ int iwpm_exit(u8 nl_client)if (!iwpm_valid_client (nl_client))return -EINVAL ;mutex_lock (&iwpm_admin_lock); -if (atomic_read (&iwpm_admin.refcount) ==0 ) { +if (!refcount_read (&iwpm_admin.refcount)) {mutex_unlock (&iwpm_admin_lock);pr_err ("%s Incorrect usage - negative refcount\n" , __func__);return -EINVAL ; } -if (atomic_dec_and_test (&iwpm_admin.refcount)) { +if (refcount_dec_and_test (&iwpm_admin.refcount)) {free_hash_bucket ();free_reminfo_bucket ();pr_debug ("%s: Resources are destroyed\n" , __func__);
drivers/infiniband/core/iwpm_util.h +1 -1
@@ -90,7 +90,7 @@ struct iwpm_remote_info { };struct iwpm_admin_data { -atomic_t refcount; +refcount_t refcount;atomic_t nlmsg_seq;int client_list[RDMA_NL_NUM_CLIENTS ];u32 reg_list[RDMA_NL_NUM_CLIENTS ];
drivers/infiniband/core/mad_priv.h +1 -1
@@ -115,7 +115,7 @@ struct ib_mad_snoop_private {struct ib_mad_qp_info *qp_info;int snoop_index;int mad_snoop_flags; -atomic_t refcount; +refcount_t refcount;struct completion comp; };
drivers/infiniband/core/multicast.c +15 -15
@@ -61,7 +61,7 @@ struct mcast_port {struct mcast_device *dev;spinlock_t lock;struct rb_root table; -atomic_t refcount; +refcount_t refcount;struct completion comp;u32 port_num; };
@@ -103,7 +103,7 @@ struct mcast_group {struct list_head active_list;struct mcast_member *last_join;int members[NUM_JOIN_MEMBERSHIP_TYPES ]; -atomic_t refcount; +refcount_t refcount;enum mcast_group_state state;struct ib_sa_query *query;u16 pkey_index;
@@ -117,7 +117,7 @@ struct mcast_member {struct mcast_group *group;struct list_head list;enum mcast_state state; -atomic_t refcount; +refcount_t refcount;struct completion comp; };
@@ -178,7 +178,7 @@ static struct mcast_group *mcast_insert(struct mcast_port *port,static void deref_port (struct mcast_port *port) { -if (atomic_dec_and_test (&port->refcount)) +if (refcount_dec_and_test (&port->refcount))complete (&port->comp); }
@@ -188,7 +188,7 @@ static void release_group(struct mcast_group *group)unsigned long flags;spin_lock_irqsave (&port->lock, flags); -if (atomic_dec_and_test (&group->refcount)) { +if (refcount_dec_and_test (&group->refcount)) {rb_erase (&group->node, &port->table);spin_unlock_irqrestore (&port->lock, flags);kfree (group);
@@ -199,7 +199,7 @@ static void release_group(struct mcast_group *group)static void deref_member (struct mcast_member *member) { -if (atomic_dec_and_test (&member->refcount)) +if (refcount_dec_and_test (&member->refcount))complete (&member->comp); }
@@ -212,7 +212,7 @@ static void queue_join(struct mcast_member *member)list_add_tail (&member->list, &group->pending_list);if (group->state ==MCAST_IDLE ) { group->state =MCAST_BUSY ; -atomic_inc (&group->refcount); +refcount_inc (&group->refcount);queue_work (mcast_wq, &group->work); }spin_unlock_irqrestore (&group->lock, flags);
@@ -401,7 +401,7 @@ static void process_group_error(struct mcast_group *group)while (!list_empty (&group->active_list)) { member =list_entry (group->active_list.next, struct mcast_member, list); -atomic_inc (&member->refcount); +refcount_inc (&member->refcount);list_del_init (&member->list);adjust_membership (group, member->multicast.rec.join_state,-1 ); member->state =MCAST_ERROR ;
@@ -445,7 +445,7 @@ static void mcast_work_handler(struct work_struct *work) struct mcast_member, list); multicast = &member->multicast; join_state = multicast->rec.join_state; -atomic_inc (&member->refcount); +refcount_inc (&member->refcount);if (join_state == (group->rec.join_state & join_state)) { status = cmp_rec(&group->rec, &multicast->rec,
@@ -497,7 +497,7 @@ static void process_join_error(struct mcast_group *group, int status) member =list_entry (group->pending_list.next, struct mcast_member, list);if (group->last_join == member) { -atomic_inc (&member->refcount); +refcount_inc (&member->refcount);list_del_init (&member->list);spin_unlock_irq (&group->lock); ret = member->multicast.callback (status, &member->multicast);
@@ -589,9 +589,9 @@ static struct mcast_group *acquire_group(struct mcast_port *port,kfree (group); group = cur_group; }else -atomic_inc (&port->refcount); +refcount_inc (&port->refcount); found: -atomic_inc (&group->refcount); +refcount_inc (&group->refcount);spin_unlock_irqrestore (&port->lock, flags);return group; }
@@ -632,7 +632,7 @@ ib_sa_join_multicast(struct ib_sa_client *client, member->multicast.callback = callback; member->multicast.context = context;init_completion (&member->comp); -atomic_set (&member->refcount,1 ); +refcount_set (&member->refcount,1 ); member->state =MCAST_JOINING ; member->group = acquire_group(&dev->port[port_num - dev->start_port],
@@ -780,7 +780,7 @@ static void mcast_groups_event(struct mcast_port *port, group = rb_entry(node, struct mcast_group, node);spin_lock (&group->lock);if (group->state ==MCAST_IDLE ) { -atomic_inc (&group->refcount); +refcount_inc (&group->refcount);queue_work (mcast_wq, &group->work); }if (group->state !=MCAST_GROUP_ERROR )
@@ -840,7 +840,7 @@ static int mcast_add_one(struct ib_device *device)spin_lock_init (&port->lock); port->table =RB_ROOT ;init_completion (&port->comp); -atomic_set (&port->refcount,1 ); +refcount_set (&port->refcount,1 ); ++count; }
drivers/infiniband/core/uverbs.h +1 -1
@@ -97,7 +97,7 @@ ib_uverbs_init_udata_buf_or_null(struct ib_udata *udata,*/ struct ib_uverbs_device { -atomic_t refcount; +refcount_t refcount;u32 num_comp_vectors;struct completion comp;struct device dev;
drivers/infiniband/core/uverbs_main.c +6 -6
@@ -197,7 +197,7 @@ void ib_uverbs_release_file(struct kref *ref)module_put (ib_dev->ops.owner);srcu_read_unlock (&file->device->disassociate_srcu, srcu_key); -if (atomic_dec_and_test (&file->device->refcount)) +if (refcount_dec_and_test (&file->device->refcount))ib_uverbs_comp_dev (file->device);if (file->default_async_file)
@@ -891,7 +891,7 @@ static int ib_uverbs_open(struct inode *inode, struct file *filp)int srcu_key; dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev); -if (!atomic_inc_not_zero (&dev->refcount)) +if (!refcount_inc_not_zero (&dev->refcount))return -ENXIO ;get_device (&dev->dev);
@@ -955,7 +955,7 @@ static int ib_uverbs_open(struct inode *inode, struct file *filp) err:mutex_unlock (&dev->lists_mutex);srcu_read_unlock (&dev->disassociate_srcu, srcu_key); -if (atomic_dec_and_test (&dev->refcount)) +if (refcount_dec_and_test (&dev->refcount))ib_uverbs_comp_dev (dev);put_device (&dev->dev);
@@ -1124,7 +1124,7 @@ static int ib_uverbs_add_one(struct ib_device *device) uverbs_dev->dev.release = ib_uverbs_release_dev; uverbs_dev->groups[0 ] = &dev_attr_group; uverbs_dev->dev.groups = uverbs_dev->groups; -atomic_set (&uverbs_dev->refcount,1 ); +refcount_set (&uverbs_dev->refcount,1 );init_completion (&uverbs_dev->comp); uverbs_dev->xrcd_tree =RB_ROOT ;mutex_init (&uverbs_dev->xrcd_tree_mutex);
@@ -1166,7 +1166,7 @@ static int ib_uverbs_add_one(struct ib_device *device) err_uapi:ida_free (&uverbs_ida, devnum); err: -if (atomic_dec_and_test (&uverbs_dev->refcount)) +if (refcount_dec_and_test (&uverbs_dev->refcount))ib_uverbs_comp_dev (uverbs_dev);wait_for_completion (&uverbs_dev->comp);put_device (&uverbs_dev->dev);
@@ -1229,7 +1229,7 @@ static void ib_uverbs_remove_one(struct ib_device *device, void *client_data) wait_clients =0 ; } -if (atomic_dec_and_test (&uverbs_dev->refcount)) +if (refcount_dec_and_test (&uverbs_dev->refcount))ib_uverbs_comp_dev (uverbs_dev);if (wait_clients)wait_for_completion (&uverbs_dev->comp);
Signature · 1 line
2.7.4
Jason Gunthorpe
maintainerRe: [PATCH for-next 1/6] RDMA/core: Use refcount_t instead of atomic_t for reference counting
Quoted 4 lines · Weihang Li
The
refcount_tAPI will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks. Increaserefcount_tfrom 0 to 1 is regarded as there is a risk about use-after-free. So it should be set to 1 directly during initialization.
What does this comment about 0 to 1 mean?
This all seems like a good idea but I wish you had done one patch per variable changed
Jason
Re: [PATCH for-next 1/6] RDMA/core: Use refcount_t instead of atomic_t for reference counting
Quoted 8 lines · Jason Gunthorpe
Quoted 4 lines · Weihang Li
The
refcount_tAPI will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks. Increaserefcount_tfrom 0 to 1 is regarded as there is a risk about use-after-free. So it should be set to 1 directly during initialization.What does this comment about 0 to 1 mean?
Hi Jason,
I first thought refcount_inc() and atomic_inc() are exactly the same, but I got a warning about refcount_t on iwpm_init() after the replacement:
[ 16.882939] refcount_t: addition on 0; use-after-free. [ 16.888065] WARNING: CPU: 2 PID: 1 at lib/refcount.c:25 refcount_warn_saturate+0xa0/0x144 ... [ 17.014698] Call trace: [ 17.017135] refcount_warn_saturate+0xa0/0x144 [ 17.021559] iwpm_init+0x104/0x12c [ 17.024948] iw_cm_init+0x24/0xd0 [ 17.028248] do_one_initcall+0x54/0x2d0 [ 17.032068] kernel_init_freeable+0x224/0x294 [ 17.036407] kernel_init+0x20/0x12c [ 17.039880] ret_from_fork+0x10/0x18
Then I noticed that the comment of refcount_inc() says:
* Will WARN if the refcount is 0, as this represents a possible use-after-free
* condition.
so I made changes:
@@ -77,8 +77,12 @@ int iwpm_init(u8 nl_client)
ret = -ENOMEM;
goto init_exit;
}
+
+ refcount_set(&iwpm_admin.refcount, 1);
+ } else {
+ refcount_inc(&iwpm_admin.refcount);
}- refcount_inc(&iwpm_admin.refcount);
+
I wrote the comments because I thought someone might be confused by the above changes :)
Quoted 5 lines
This all seems like a good idea but I wish you had done one patch per variable changed
Jason
Sure, thanks.
Weihang
Re: [PATCH for-next 1/6] RDMA/core: Use refcount_t instead of atomic_t for reference counting
Quoted 47 lines · liweihang
Quoted 8 lines · Jason Gunthorpe
Quoted 4 lines · Weihang Li
The
refcount_tAPI will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks. Increaserefcount_tfrom 0 to 1 is regarded as there is a risk about use-after-free. So it should be set to 1 directly during initialization.What does this comment about 0 to 1 mean?
Hi Jason,
I first thought
refcount_inc()andatomic_inc()are exactly the same, but I got a warning aboutrefcount_toniwpm_init()after the replacement:[ 16.882939] refcount_t: addition on 0; use-after-free. [ 16.888065] WARNING: CPU: 2 PID: 1 at lib/refcount.c:25 refcount_warn_saturate+0xa0/0x144 ... [ 17.014698] Call trace: [ 17.017135] refcount_warn_saturate+0xa0/0x144 [ 17.021559] iwpm_init+0x104/0x12c [ 17.024948] iw_cm_init+0x24/0xd0 [ 17.028248] do_one_initcall+0x54/0x2d0 [ 17.032068] kernel_init_freeable+0x224/0x294 [ 17.036407] kernel_init+0x20/0x12c [ 17.039880] ret_from_fork+0x10/0x18Then I noticed that the comment of
refcount_inc()says:* Will WARN if the refcount is 0, as this represents a possible use-after-free
* condition.so I made changes:
drivers/infiniband/core/iwpm_util.c +5
-1@@ -77,8 +77,12 @@ int iwpm_init(u8 nl_client) ret = -ENOMEM ; goto init_exit; } + +refcount_set (&iwpm_admin.refcount,1 ); + }else { +refcount_inc (&iwpm_admin.refcount); } -refcount_inc (&iwpm_admin.refcount); +
Quoted 2 lines
I wrote the comments because I thought someone might be confused by the above changes :)
Stuff like this needs to be split into a single patch for iwpm_admin
Jason
Re: [PATCH for-next 1/6] RDMA/core: Use refcount_t instead of atomic_t for reference counting
Quoted 48 lines · Jason Gunthorpe
Quoted 47 lines · liweihang
Quoted 8 lines · Jason Gunthorpe
Quoted 4 lines · Weihang Li
The
refcount_tAPI will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks. Increaserefcount_tfrom 0 to 1 is regarded as there is a risk about use-after-free. So it should be set to 1 directly during initialization.What does this comment about 0 to 1 mean?
Hi Jason,
I first thought
refcount_inc()andatomic_inc()are exactly the same, but I got a warning aboutrefcount_toniwpm_init()after the replacement:[ 16.882939] refcount_t: addition on 0; use-after-free. [ 16.888065] WARNING: CPU: 2 PID: 1 at lib/refcount.c:25 refcount_warn_saturate+0xa0/0x144 ... [ 17.014698] Call trace: [ 17.017135] refcount_warn_saturate+0xa0/0x144 [ 17.021559] iwpm_init+0x104/0x12c [ 17.024948] iw_cm_init+0x24/0xd0 [ 17.028248] do_one_initcall+0x54/0x2d0 [ 17.032068] kernel_init_freeable+0x224/0x294 [ 17.036407] kernel_init+0x20/0x12c [ 17.039880] ret_from_fork+0x10/0x18Then I noticed that the comment of
refcount_inc()says:* Will WARN if the refcount is 0, as this represents a possible use-after-free
* condition.so I made changes:
drivers/infiniband/core/iwpm_util.c +5
-1@@ -77,8 +77,12 @@ int iwpm_init(u8 nl_client) ret = -ENOMEM ; goto init_exit; } + +refcount_set (&iwpm_admin.refcount,1 ); + }else { +refcount_inc (&iwpm_admin.refcount); } -refcount_inc (&iwpm_admin.refcount); +
Quoted 7 lines
Quoted 2 lines
I wrote the comments because I thought someone might be confused by the above changes :)
Stuff like this needs to be split into a single patch for
iwpm_adminJason
Sure, I will split all of the changes into separate patches.
Weihang
RE: [PATCH for-next 1/6] RDMA/core: Use refcount_t instead of atomic_t for reference counting
In reply to Weihang LiQuoted 19 lines
Subject: [PATCH for-next 1/6] RDMA/core: Use
refcount_tinstead ofatomic_tfor reference countingThe
refcount_tAPI will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks. Increaserefcount_tfrom 0 to 1 is regarded as there is a risk about use-after-free. So it should be set to 1 directly during initialization.Signed-off-by Weihang Li <[email protected]>--- drivers/infiniband/core/iwcm.c | 9 ++++----- drivers/infiniband/core/iwcm.h | 2 +- drivers/infiniband/core/iwpm_util.c | 12 ++++++++---- drivers/infiniband/core/iwpm_util.h | 2 +- drivers/infiniband/core/mad_priv.h | 2 +- drivers/infiniband/core/multicast.c | 30 +++++++++++++++--------------- drivers/infiniband/core/uverbs.h | 2 +- drivers/infiniband/core/uverbs_main.c | 12 ++++++------ 8 files changed, 37 insertions(+), 34 deletions(-)
[...]
drivers/infiniband/core/multicast.c static struct mcast_group *acquire_group(struct · line 595
Quoted 10 lines
drivers/infiniband/core/multicast.c +2
-2@@ -589,9 +589,9 @@ static struct mcast_group *acquire_group(structmcast_port *port,kfree (group ); group = cur_group; }else -atomic_inc (&port->refcount); +refcount_inc (&port->refcount); found: -atomic_inc (&group->refcount); +refcount_inc (&group->refcount);
Seems like there is refcount_inc with refcount = 0 when the group is first created?
Quoted 3 lines
drivers/infiniband/core/multicast.c +0
-0spin_unlock_irqrestore (&port->lock, flags);return group;}
Re: [PATCH for-next 1/6] RDMA/core: Use refcount_t instead of atomic_t for reference counting
Quoted 22 lines · Shiraz
Quoted 19 lines
Subject: [PATCH for-next 1/6] RDMA/core: Use
refcount_tinstead ofatomic_tfor reference countingThe
refcount_tAPI will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks. Increaserefcount_tfrom 0 to 1 is regarded as there is a risk about use-after-free. So it should be set to 1 directly during initialization.Signed-off-by Weihang Li <[email protected]>--- drivers/infiniband/core/iwcm.c | 9 ++++----- drivers/infiniband/core/iwcm.h | 2 +- drivers/infiniband/core/iwpm_util.c | 12 ++++++++---- drivers/infiniband/core/iwpm_util.h | 2 +- drivers/infiniband/core/mad_priv.h | 2 +- drivers/infiniband/core/multicast.c | 30 +++++++++++++++--------------- drivers/infiniband/core/uverbs.h | 2 +- drivers/infiniband/core/uverbs_main.c | 12 ++++++------ 8 files changed, 37 insertions(+), 34 deletions(-)[...]
Quoted 12 lines
drivers/infiniband/core/multicast.c static struct mcast_group *acquire_group(struct · line 595
Quoted 10 lines
drivers/infiniband/core/multicast.c +2
-2@@ -589,9 +589,9 @@ static struct mcast_group *acquire_group(structmcast_port *port,kfree (group ); group = cur_group; }else -atomic_inc (&port->refcount); +refcount_inc (&port->refcount); found: -atomic_inc (&group->refcount); +refcount_inc (&group->refcount);Seems like there is
refcount_incwith refcount = 0 when the group is first created?
Yes, one of "refcount_inc(&group->refcount)" led to the issue that Leon had reported. I will fix it, thank you.
Weihang
Quoted 6 lines
Quoted 4 lines
drivers/infiniband/core/multicast.c +0
-0spin_unlock_irqrestore (&port->lock, flags);return group;}
PATCH 2/6 RDMA/hns: Use refcount_t instead of atomic_t for reference counting
In reply to Weihang LiThe refcount_t API will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks, so the type of refcnt in QP/CQ/SRQ is changed from atomic_t to refcount_t.
--- drivers/infiniband/hw/hns/hns_roce_cq.c | 8 ++++---- drivers/infiniband/hw/hns/hns_roce_device.h | 6 +++--- drivers/infiniband/hw/hns/hns_roce_qp.c | 12 ++++++------ drivers/infiniband/hw/hns/hns_roce_srq.c | 8 ++++---- 4 files changed, 17 insertions(+), 17 deletions(-)
drivers/infiniband/hw/hns/hns_roce_cq.c +4 -4
@@ -154,7 +154,7 @@ static int alloc_cqc(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq) hr_cq->cons_index =0 ; hr_cq->arm_sn =1 ; -atomic_set (&hr_cq->refcount,1 ); +refcount_set (&hr_cq->refcount,1 );init_completion (&hr_cq->free);return 0 ;
@@ -188,7 +188,7 @@ static void free_cqc(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq)synchronize_irq (hr_dev->eq_table.eq[hr_cq->vector].irq);/* wait for all interrupt processed */ -if (atomic_dec_and_test (&hr_cq->refcount)) +if (refcount_dec_and_test (&hr_cq->refcount))complete (&hr_cq->free);wait_for_completion (&hr_cq->free);
@@ -481,7 +481,7 @@ void hns_roce_cq_event(struct hns_roce_dev *hr_dev, u32 cqn, int event_type)return ; } -atomic_inc (&hr_cq->refcount); +refcount_inc (&hr_cq->refcount); ibcq = &hr_cq->ib_cq;if (ibcq->event_handler) {
@@ -491,7 +491,7 @@ void hns_roce_cq_event(struct hns_roce_dev *hr_dev, u32 cqn, int event_type) ibcq->event_handler (&event, ibcq->cq_context); } -if (atomic_dec_and_test (&hr_cq->refcount)) +if (refcount_dec_and_test (&hr_cq->refcount))complete (&hr_cq->free); }
drivers/infiniband/hw/hns/hns_roce_device.h +3 -3
@@ -446,7 +446,7 @@ struct hns_roce_cq {int cqe_size;unsigned long cqn;u32 vector; -atomic_t refcount; +refcount_t refcount;struct completion free;struct list_head sq_list;/* all qps on this send cq */ struct list_head rq_list;/* all qps on this recv cq */
@@ -473,7 +473,7 @@ struct hns_roce_srq {u32 xrcdn;void __iomem *db_reg; -atomic_t refcount; +refcount_t refcount;struct completion free;struct hns_roce_mtr buf_mtr;
@@ -642,7 +642,7 @@ struct hns_roce_qp {u32 xrcdn; -atomic_t refcount; +refcount_t refcount;struct completion free;struct hns_roce_sge sge;
drivers/infiniband/hw/hns/hns_roce_qp.c +6 -6
@@ -65,7 +65,7 @@ static void flush_work_handle(struct work_struct *work)* make sure we signal QP destroy leg that flush QP was completed * so that it can safely proceed ahead now and destroy QP */ -if (atomic_dec_and_test (&hr_qp->refcount)) +if (refcount_dec_and_test (&hr_qp->refcount))complete (&hr_qp->free); }
@@ -75,7 +75,7 @@ void init_flush_work(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp) flush_work->hr_dev = hr_dev;INIT_WORK (&flush_work->work, flush_work_handle); -atomic_inc (&hr_qp->refcount); +refcount_inc (&hr_qp->refcount);queue_work (hr_dev->irq_workq, &flush_work->work); }
@@ -87,7 +87,7 @@ void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)xa_lock (&hr_dev->qp_table_xa); qp =__hns_roce_qp_lookup (hr_dev, qpn);if (qp) -atomic_inc (&qp->refcount); +refcount_inc (&qp->refcount);xa_unlock (&hr_dev->qp_table_xa);if (!qp) {
@@ -108,7 +108,7 @@ void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type) qp->event (qp, (enum hns_roce_event )event_type); -if (atomic_dec_and_test (&qp->refcount)) +if (refcount_dec_and_test (&qp->refcount))complete (&qp->free); }
@@ -1076,7 +1076,7 @@ static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev, hr_qp->ibqp.qp_num = hr_qp->qpn; hr_qp->event = hns_roce_ib_qp_event; -atomic_set (&hr_qp->refcount,1 ); +refcount_set (&hr_qp->refcount,1 );init_completion (&hr_qp->free);return 0 ;
@@ -1099,7 +1099,7 @@ static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,void hns_roce_qp_destroy (struct hns_roce_dev *hr_dev,struct hns_roce_qp *hr_qp,struct ib_udata *udata) { -if (atomic_dec_and_test (&hr_qp->refcount)) +if (refcount_dec_and_test (&hr_qp->refcount))complete (&hr_qp->free);wait_for_completion (&hr_qp->free);
drivers/infiniband/hw/hns/hns_roce_srq.c +4 -4
@@ -17,7 +17,7 @@ void hns_roce_srq_event(struct hns_roce_dev *hr_dev, u32 srqn, int event_type)xa_lock (&srq_table->xa); srq =xa_load (&srq_table->xa, srqn & (hr_dev->caps.num_srqs -1 ));if (srq) -atomic_inc (&srq->refcount); +refcount_inc (&srq->refcount);xa_unlock (&srq_table->xa);if (!srq) {
@@ -27,7 +27,7 @@ void hns_roce_srq_event(struct hns_roce_dev *hr_dev, u32 srqn, int event_type) srq->event (srq, event_type); -if (atomic_dec_and_test (&srq->refcount)) +if (refcount_dec_and_test (&srq->refcount))complete (&srq->free); }
@@ -149,7 +149,7 @@ static void free_srqc(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)xa_erase (&srq_table->xa, srq->srqn); -if (atomic_dec_and_test (&srq->refcount)) +if (refcount_dec_and_test (&srq->refcount))complete (&srq->free);wait_for_completion (&srq->free);
@@ -417,7 +417,7 @@ int hns_roce_create_srq(struct ib_srq *ib_srq, srq->db_reg = hr_dev->reg_base +SRQ_DB_REG ; srq->event = hns_roce_ib_srq_event; -atomic_set (&srq->refcount,1 ); +refcount_set (&srq->refcount,1 );init_completion (&srq->free);return 0 ;
Signature · 1 line
2.7.4
PATCH 3/6 RDMA/hns: Use refcount_t APIs for HEM
In reply to Weihang Lirefcount_t is better than integer for reference counting, it will WARN on overflow/underflow and avoid use-after-free risks.
--- drivers/infiniband/hw/hns/hns_roce_hem.c | 33 ++++++++++++++++---------------- drivers/infiniband/hw/hns/hns_roce_hem.h | 4 ++-- 2 files changed, 18 insertions(+), 19 deletions(-)
drivers/infiniband/hw/hns/hns_roce_hem.c +16 -17
@@ -271,7 +271,7 @@ static struct hns_roce_hem *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,if (!hem)return NULL ; - hem->refcount =0 ; +refcount_set (&hem->refcount,0 );INIT_LIST_HEAD (&hem->chunk_list); order =get_order (hem_alloc_size);
@@ -618,7 +618,7 @@ static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev,mutex_lock (&table->mutex);if (table->hem[index.buf]) { - ++table->hem[index.buf]->refcount; +refcount_inc (&table->hem[index.buf]->refcount); goto out; }
@@ -637,7 +637,7 @@ static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev, } } - ++table->hem[index.buf]->refcount; +refcount_set (&table->hem[index.buf]->refcount,1 ); goto out; err_alloc:
@@ -663,7 +663,7 @@ int hns_roce_table_get(struct hns_roce_dev *hr_dev,mutex_lock (&table->mutex);if (table->hem[i]) { - ++table->hem[i]->refcount; +refcount_inc (&table->hem[i]->refcount); goto out; }
@@ -686,7 +686,7 @@ int hns_roce_table_get(struct hns_roce_dev *hr_dev, goto out; } - ++table->hem[i]->refcount; +refcount_set (&table->hem[i]->refcount,1 ); out:mutex_unlock (&table->mutex);return ret;
@@ -753,11 +753,11 @@ static void hns_roce_table_mhop_put(struct hns_roce_dev *hr_dev,return ; } -mutex_lock (&table->mutex); -if (check_refcount && (--table->hem[index.buf]->refcount >0 )) { -mutex_unlock (&table->mutex); +if (!check_refcount) +mutex_lock (&table->mutex); +else if (!refcount_dec_and_mutex_lock (&table->hem[index.buf]->refcount, + &table->mutex))return ; - }clear_mhop_hem (hr_dev, table, obj, &mhop, &index);free_mhop_hem (hr_dev, table, &mhop, &index);
@@ -779,16 +779,15 @@ void hns_roce_table_put(struct hns_roce_dev *hr_dev, i = (obj & (table->num_obj -1 )) / (table->table_chunk_size / table->obj_size); -mutex_lock (&table->mutex); +if (!refcount_dec_and_mutex_lock (&table->hem[i]->refcount, + &table->mutex)) +return ; -if (--table->hem[i]->refcount ==0 ) { -/* Clear HEM base address */ -if (hr_dev->hw->clear_hem (hr_dev, table, obj,0 )) -dev_warn (dev,"Clear HEM base address failed.\n" ); +if (hr_dev->hw->clear_hem (hr_dev, table, obj,0 )) +dev_warn (dev,"failed to clear HEM base address.\n" ); -hns_roce_free_hem (hr_dev, table->hem[i]); - table->hem[i] =NULL ; - } +hns_roce_free_hem (hr_dev, table->hem[i]); + table->hem[i] =NULL ;mutex_unlock (&table->mutex); }
drivers/infiniband/hw/hns/hns_roce_hem.h +2 -2
@@ -88,8 +88,8 @@ struct hns_roce_hem_chunk { };struct hns_roce_hem { -struct list_head chunk_list; -int refcount; +struct list_head chunk_list; +refcount_t refcount; };struct hns_roce_hem_iter {
Signature · 1 line
2.7.4
PATCH 4/6 RDMA/cxgb4: Use refcount_t instead of atomic_t for reference counting
In reply to Weihang LiThe refcount_t API will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks.
--- drivers/infiniband/hw/cxgb4/cq.c | 6 +++--- drivers/infiniband/hw/cxgb4/ev.c | 8 ++++---- drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 2 +- 3 files changed, 8 insertions(+), 8 deletions(-)
drivers/infiniband/hw/cxgb4/cq.c +3 -3
@@ -976,8 +976,8 @@ int c4iw_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata) chp =to_c4iw_cq (ib_cq);xa_erase_irq (&chp->rhp->cqs, chp->cq.cqid); -atomic_dec (&chp->refcnt); -wait_event (chp->wait, !atomic_read (&chp->refcnt)); +refcount_dec (&chp->refcnt); +wait_event (chp->wait, !refcount_read (&chp->refcnt)); ucontext = rdma_udata_to_drv_context(udata, struct c4iw_ucontext, ibucontext);
@@ -1080,7 +1080,7 @@ int c4iw_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, chp->ibcq.cqe = entries -2 ;spin_lock_init (&chp->lock);spin_lock_init (&chp->comp_handler_lock); -atomic_set (&chp->refcnt,1 ); +refcount_set (&chp->refcnt,1 );init_waitqueue_head (&chp->wait); ret =xa_insert_irq (&rhp->cqs, chp->cq.cqid, chp,GFP_KERNEL );if (ret)
drivers/infiniband/hw/cxgb4/ev.c +4 -4
@@ -151,7 +151,7 @@ void c4iw_ev_dispatch(struct c4iw_dev *dev, struct t4_cqe *err_cqe) }c4iw_qp_add_ref (&qhp->ibqp); -atomic_inc (&chp->refcnt); +refcount_inc (&chp->refcnt);xa_unlock_irq (&dev->qps);/* Bad incoming write */
@@ -213,7 +213,7 @@ void c4iw_ev_dispatch(struct c4iw_dev *dev, struct t4_cqe *err_cqe)break ; } done: -if (atomic_dec_and_test (&chp->refcnt)) +if (refcount_dec_and_test (&chp->refcnt))wake_up (&chp->wait);c4iw_qp_rem_ref (&qhp->ibqp); out:
@@ -228,13 +228,13 @@ int c4iw_ev_handler(struct c4iw_dev *dev, u32 qid)xa_lock_irqsave (&dev->cqs, flag); chp =xa_load (&dev->cqs, qid);if (chp) { -atomic_inc (&chp->refcnt); +refcount_inc (&chp->refcnt);xa_unlock_irqrestore (&dev->cqs, flag);t4_clear_cq_armed (&chp->cq);spin_lock_irqsave (&chp->comp_handler_lock, flag); (*chp->ibcq.comp_handler)(&chp->ibcq, chp->ibcq.cq_context);spin_unlock_irqrestore (&chp->comp_handler_lock, flag); -if (atomic_dec_and_test (&chp->refcnt)) +if (refcount_dec_and_test (&chp->refcnt))wake_up (&chp->wait); }else {pr_debug ("unknown cqid 0x%x\n" , qid);
drivers/infiniband/hw/cxgb4/iw_cxgb4.h +1 -1
@@ -427,7 +427,7 @@ struct c4iw_cq {struct t4_cq cq;spinlock_t lock;spinlock_t comp_handler_lock; -atomic_t refcnt; +refcount_t refcnt;wait_queue_head_t wait;struct c4iw_wr_wait *wr_waitp; };
Signature · 1 line
2.7.4
PATCH 5/6 RDMA/i40iw: Use refcount_t instead of atomic_t for reference counting
In reply to Weihang LiThe refcount_t API will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks.
--- drivers/infiniband/hw/i40iw/i40iw.h | 2 +- drivers/infiniband/hw/i40iw/i40iw_cm.c | 54 +++++++++++++++---------------- drivers/infiniband/hw/i40iw/i40iw_cm.h | 4 +-- drivers/infiniband/hw/i40iw/i40iw_main.c | 2 +- drivers/infiniband/hw/i40iw/i40iw_puda.h | 2 +- drivers/infiniband/hw/i40iw/i40iw_utils.c | 10 +++--- 6 files changed, 37 insertions(+), 37 deletions(-)
drivers/infiniband/hw/i40iw/i40iw.h +1 -1
@@ -137,7 +137,7 @@ struct i40iw_cqp_request {struct cqp_commands_info info;wait_queue_head_t waitq;struct list_head list; -atomic_t refcount; +refcount_t refcount;void (*callback_fcn)(struct i40iw_cqp_request *,u32 );void *param;struct i40iw_cqp_compl_info compl_info;
drivers/infiniband/hw/i40iw/i40iw_cm.c +27 -27
213 lines not on this page · raw message
drivers/infiniband/hw/i40iw/i40iw_cm.h +2 -2
@@ -291,7 +291,7 @@ struct i40iw_cm_listener {u32 loc_addr[4 ];u16 loc_port;struct iw_cm_id *cm_id; -atomic_t ref_count; +refcount_t ref_count;struct i40iw_device *iwdev;atomic_t pend_accepts_cnt;int backlog;
@@ -319,7 +319,7 @@ struct i40iw_cm_node {enum i40iw_cm_node_state state;u8 loc_mac[ETH_ALEN ];u8 rem_mac[ETH_ALEN ]; -atomic_t ref_count; +refcount_t ref_count;struct i40iw_qp *iwqp;struct i40iw_device *iwdev;struct i40iw_sc_dev *dev;
drivers/infiniband/hw/i40iw/i40iw_main.c +1 -1
@@ -1125,7 +1125,7 @@ static enum i40iw_status_code i40iw_alloc_local_mac_ipaddr_entry(struct i40iw_de }/* increment refcount, because we need the cqp request ret value */ -atomic_inc (&cqp_request->refcount); +refcount_inc (&cqp_request->refcount); cqp_info = &cqp_request->info; cqp_info->cqp_cmd =OP_ALLOC_LOCAL_MAC_IPADDR_ENTRY ;
drivers/infiniband/hw/i40iw/i40iw_puda.h +1 -1
@@ -90,7 +90,7 @@ struct i40iw_puda_buf {u8 tcphlen;/* tcp length in bytes */ u8 maclen;/* mac length in bytes */ u32 totallen;/* machlen+iphlen+tcphlen+datalen */ -atomic_t refcount; +refcount_t refcount;u8 hdrlen;bool ipv4;u32 seqnum;
drivers/infiniband/hw/i40iw/i40iw_utils.c +5 -5
@@ -384,10 +384,10 @@ struct i40iw_cqp_request *i40iw_get_cqp_request(struct i40iw_cqp *cqp, bool wait }if (wait) { -atomic_set (&cqp_request->refcount,2 ); +refcount_set (&cqp_request->refcount,2 ); cqp_request->waiting = true; }else { -atomic_set (&cqp_request->refcount,1 ); +refcount_set (&cqp_request->refcount,1 ); }return cqp_request; }
@@ -424,7 +424,7 @@ void i40iw_free_cqp_request(struct i40iw_cqp *cqp, struct i40iw_cqp_request *cqpvoid i40iw_put_cqp_request (struct i40iw_cqp *cqp,struct i40iw_cqp_request *cqp_request) { -if (atomic_dec_and_test (&cqp_request->refcount)) +if (refcount_dec_and_test (&cqp_request->refcount))i40iw_free_cqp_request (cqp, cqp_request); }
@@ -445,7 +445,7 @@ static void i40iw_free_pending_cqp_request(struct i40iw_cqp *cqp, }i40iw_put_cqp_request (cqp, cqp_request);wait_event_timeout (iwdev->close_wq, - !atomic_read (&cqp_request->refcount), + !refcount_read (&cqp_request->refcount),1000 ); }
@@ -1005,7 +1005,7 @@ static void i40iw_cqp_manage_hmc_fcn_callback(struct i40iw_cqp_request *cqp_requif (hmcfcninfo && hmcfcninfo->callback_fcn) {i40iw_debug (&iwdev->sc_dev,I40IW_DEBUG_HMC ,"%s1\n" , __func__); -atomic_inc (&cqp_request->refcount); +refcount_inc (&cqp_request->refcount); work = &iwdev->virtchnl_w[hmcfcninfo->iw_vf_idx]; work->cqp_request = cqp_request;INIT_WORK (&work->work, i40iw_cqp_manage_hmc_fcn_worker);
Signature · 1 line
2.7.4
PATCH 6/6 RDMA/ipoib: Use refcount_t instead of atomic_t for reference counting
In reply to Weihang LiThe refcount_t API will WARN on underflow and overflow of a reference counter, and avoid use-after-free risks.
--- drivers/infiniband/ulp/ipoib/ipoib.h | 4 ++-- drivers/infiniband/ulp/ipoib/ipoib_main.c | 8 ++++---- 2 files changed, 6 insertions(+), 6 deletions(-)
drivers/infiniband/ulp/ipoib/ipoib.h +2 -2
@@ -454,7 +454,7 @@ struct ipoib_neigh {struct list_head list;struct ipoib_neigh __rcu *hnext;struct rcu_head rcu; -atomic_t refcnt; +refcount_t refcnt;unsigned long alive; };
@@ -464,7 +464,7 @@ struct ipoib_neigh {void ipoib_neigh_dtor (struct ipoib_neigh *neigh);static inline void ipoib_neigh_put (struct ipoib_neigh *neigh) { -if (atomic_dec_and_test (&neigh->refcnt)) +if (refcount_dec_and_test (&neigh->refcnt))ipoib_neigh_dtor (neigh); }struct ipoib_neigh *ipoib_neigh_get (struct net_device *dev,u8 *daddr);
drivers/infiniband/ulp/ipoib/ipoib_main.c +4 -4
@@ -1287,7 +1287,7 @@ struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr) neigh = rcu_dereference_bh(neigh->hnext)) {if (memcmp (daddr, neigh->daddr,INFINIBAND_ALEN ) ==0 ) {/* found, take one ref on behalf of the caller */ -if (!atomic_inc_not_zero (&neigh->refcnt)) { +if (!refcount_inc_not_zero (&neigh->refcnt)) {/* deleted */ neigh =NULL ; goto out_unlock;
@@ -1382,7 +1382,7 @@ static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,INIT_LIST_HEAD (&neigh->list);ipoib_cm_set (neigh,NULL );/* one ref on behalf of the caller */ -atomic_set (&neigh->refcnt,1 ); +refcount_set (&neigh->refcnt,1 );return neigh; }
@@ -1414,7 +1414,7 @@ struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr, lockdep_is_held(&priv->lock))) {if (memcmp (daddr, neigh->daddr,INFINIBAND_ALEN ) ==0 ) {/* found, take one ref on behalf of the caller */ -if (!atomic_inc_not_zero (&neigh->refcnt)) { +if (!refcount_inc_not_zero (&neigh->refcnt)) {/* deleted */ neigh =NULL ;break ;
@@ -1429,7 +1429,7 @@ struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr, goto out_unlock;/* one ref on behalf of the hash table */ -atomic_inc (&neigh->refcnt); +refcount_inc (&neigh->refcnt); neigh->alive = jiffies;/* put in hash */ rcu_assign_pointer(neigh->hnext,
Signature · 1 line
2.7.4
Quoted 5 lines · Weihang Li
There are some refcnt in type of
atomic_tin RDMA subsystem, almost all of them is wrote before therefcount_tis acheived in kernel.refcount_tis better than integer for reference counting, it will WARN on overflow/underflow and avoid use-after-free risks.
Quoted 7 lines
Weihang Li (6):
RDMA/core: Userefcount_tinstead ofatomic_tfor reference counting RDMA/hns: Userefcount_tinstead ofatomic_tfor reference counting RDMA/hns: Userefcount_tAPIs for HEM
RDMA/cxgb4: Userefcount_tinstead ofatomic_tfor reference counting RDMA/i40iw: Userefcount_tinstead ofatomic_tfor reference counting RDMA/ipoib: Userefcount_tinstead ofatomic_tfor reference counting
This series causes to the following splat when restarting mlx4 or mlx5 drivers.
[ 227.529930] ------------[ cut here ]------------ [ 227.530966] refcount_t: addition on 0; use-after-free. [ 227.531890] WARNING: CPU: 1 PID: 579 at lib/refcount.c:25 refcount_warn_saturate+0xb4/0x130 [ 227.533427] Modules linked in: bonding nf_tables ipip tunnel4 geneve ip6_gre ip6_tunnel tunnel6 ip_gre gre mlx5_ib mlx5_core ptp pps_core rdma_ucm ib_uverbs ib_ipoib ib_umad openvswitch nsh xt_conntrack xt_MASQUERADE nf_conntrack_netlink nfnetlink xt_addrtype iptable_nat nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 br_netfilter rpcrdma ib_iser libiscsi scsi_transport_iscsi rdma_cm iw_cm ib_cm ib_core overlay fuse [last unloaded: ib_uverbs] [ 227.539872] CPU: 1 PID: 579 Comm: kworker/u20:7 Not tainted 5.13.0-rc1_2021_05_14_02_56_31_ #1 [ 227.541458] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014 [ 227.543424] Workqueue: ipoib_wq ipoib_mcast_join_task [ib_ipoib] [ 227.544497] RIP: 0010:refcount_warn_saturate+0xb4/0x130 [ 227.545515] Code: 5b c3 0f b6 1d 69 19 40 01 80 fb 01 0f 87 60 fd 52 00 83 e3 01 75 91 48 c7 c7 88 9a 57 82 c6 05 4d 19 40 01 01 e8 35 50 4e 00 <0f> 0b 5b c3 0f b6 1d 3f 19 40 01 80 fb 01 0f 87 70 fd 52 00 83 e3 [ 227.548723] RSP: 0018:ffff888133d23cb0 EFLAGS: 00010082 [ 227.549739] RAX: 0000000000000000 RBX: 0000000000000000 RCX: ffff8882f58a78b8 [ 227.551077] RDX: 00000000ffffffd8 RSI: 0000000000000027 RDI: ffff8882f58a78b0 [ 227.552306] RBP: ffff888133d23d30 R08: ffffffff83b5fd40 R09: 0000000000000001 [ 227.553589] R10: 0000000000000001 R11: 746e756f63666572 R12: ffff88813e00c700 [ 227.554885] R13: ffff8881182a3720 R14: ffff88810d06f938 R15: ffff8881182a3600 [ 227.556104] FS: 0000000000000000(0000) GS:ffff8882f5880000(0000) knlGS:0000000000000000 [ 227.557594] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 227.558719] CR2: 000055fb4a2ee768 CR3: 000000010303e006 CR4: 0000000000370ea0 [ 227.559940] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 227.561223] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 227.562566] Call Trace: [ 227.563088] ib_sa_join_multicast+0x4ec/0x580 [ib_core] [ 227.564059] ipoib_mcast_join+0x16d/0x220 [ib_ipoib] [ 227.564961] ? ipoib_mcast_join+0x220/0x220 [ib_ipoib] [ 227.565991] ipoib_mcast_join_task+0x16e/0x320 [ib_ipoib] [ 227.567014] ? process_one_work+0x1d7/0x5b0 [ 227.567809] process_one_work+0x25a/0x5b0 [ 227.568587] worker_thread+0x4f/0x3e0 [ 227.569349] ? process_one_work+0x5b0/0x5b0 [ 227.570210] kthread+0x129/0x140 [ 227.570874] ? __kthread_bind_mask+0x60/0x60 [ 227.571677] ret_from_fork+0x1f/0x30 [ 227.572416] irq event stamp: 150582 [ 227.573190] hardirqs last enabled at (150581): [<ffffffff81c4b337>] _raw_spin_unlock_irqrestore+0x47/0x50 [ 227.575046] hardirqs last disabled at (150582): [<ffffffff81c4b130>] _raw_spin_lock_irqsave+0x50/0x60 [ 227.576793] softirqs last enabled at (150576): [<ffffffffa00bcbef>] ipoib_mcast_join_task+0xef/0x320 [ib_ipoib] [ 227.578657] softirqs last disabled at (150574): [<ffffffffa00bcb9f>] ipoib_mcast_join_task+0x9f/0x320 [ib_ipoib] [ 227.580416] ---[ end trace 52a055538498ac03 ]---
Thanks
Quoted 55 lines · Leon Romanovsky
Quoted 5 lines · Weihang Li
There are some refcnt in type of
atomic_tin RDMA subsystem, almost all of them is wrote before therefcount_tis acheived in kernel.refcount_tis better than integer for reference counting, it will WARN on overflow/underflow and avoid use-after-free risks.Quoted 7 lines
Weihang Li (6):
RDMA/core: Userefcount_tinstead ofatomic_tfor reference counting RDMA/hns: Userefcount_tinstead ofatomic_tfor reference counting RDMA/hns: Userefcount_tAPIs for HEM
RDMA/cxgb4: Userefcount_tinstead ofatomic_tfor reference counting RDMA/i40iw: Userefcount_tinstead ofatomic_tfor reference counting RDMA/ipoib: Userefcount_tinstead ofatomic_tfor reference countingThis series causes to the following splat when restarting mlx4 or mlx5 drivers.
[ 227.529930] ------------[ cut here ]------------ [ 227.530966] refcount_t: addition on 0; use-after-free. [ 227.531890] WARNING: CPU: 1 PID: 579 at lib/refcount.c:25 refcount_warn_saturate+0xb4/0x130 [ 227.533427] Modules linked in: bonding nf_tables ipip tunnel4 geneve ip6_gre ip6_tunnel tunnel6 ip_gre gre mlx5_ib mlx5_core ptp pps_core rdma_ucm ib_uverbs ib_ipoib ib_umad openvswitch nsh xt_conntrack xt_MASQUERADE nf_conntrack_netlink nfnetlink xt_addrtype iptable_nat nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 br_netfilter rpcrdma ib_iser libiscsi scsi_transport_iscsi rdma_cm iw_cm ib_cm ib_core overlay fuse [last unloaded: ib_uverbs] [ 227.539872] CPU: 1 PID: 579 Comm: kworker/u20:7 Not tainted 5.13.0-rc1_2021_05_14_02_56_31_ #1 [ 227.541458] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014 [ 227.543424] Workqueue: ipoib_wq ipoib_mcast_join_task [ib_ipoib] [ 227.544497] RIP: 0010:refcount_warn_saturate+0xb4/0x130 [ 227.545515] Code: 5b c3 0f b6 1d 69 19 40 01 80 fb 01 0f 87 60 fd 52 00 83 e3 01 75 91 48 c7 c7 88 9a 57 82 c6 05 4d 19 40 01 01 e8 35 50 4e 00 <0f> 0b 5b c3 0f b6 1d 3f 19 40 01 80 fb 01 0f 87 70 fd 52 00 83 e3 [ 227.548723] RSP: 0018:ffff888133d23cb0 EFLAGS: 00010082 [ 227.549739] RAX: 0000000000000000 RBX: 0000000000000000 RCX: ffff8882f58a78b8 [ 227.551077] RDX: 00000000ffffffd8 RSI: 0000000000000027 RDI: ffff8882f58a78b0 [ 227.552306] RBP: ffff888133d23d30 R08: ffffffff83b5fd40 R09: 0000000000000001 [ 227.553589] R10: 0000000000000001 R11: 746e756f63666572 R12: ffff88813e00c700 [ 227.554885] R13: ffff8881182a3720 R14: ffff88810d06f938 R15: ffff8881182a3600 [ 227.556104] FS: 0000000000000000(0000) GS:ffff8882f5880000(0000) knlGS:0000000000000000 [ 227.557594] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 227.558719] CR2: 000055fb4a2ee768 CR3: 000000010303e006 CR4: 0000000000370ea0 [ 227.559940] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 227.561223] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 227.562566] Call Trace: [ 227.563088] ib_sa_join_multicast+0x4ec/0x580 [ib_core] [ 227.564059] ipoib_mcast_join+0x16d/0x220 [ib_ipoib] [ 227.564961] ? ipoib_mcast_join+0x220/0x220 [ib_ipoib] [ 227.565991] ipoib_mcast_join_task+0x16e/0x320 [ib_ipoib] [ 227.567014] ? process_one_work+0x1d7/0x5b0 [ 227.567809] process_one_work+0x25a/0x5b0 [ 227.568587] worker_thread+0x4f/0x3e0 [ 227.569349] ? process_one_work+0x5b0/0x5b0 [ 227.570210] kthread+0x129/0x140 [ 227.570874] ? __kthread_bind_mask+0x60/0x60 [ 227.571677] ret_from_fork+0x1f/0x30 [ 227.572416] irq event stamp: 150582 [ 227.573190] hardirqs last enabled at (150581): [<ffffffff81c4b337>] _raw_spin_unlock_irqrestore+0x47/0x50 [ 227.575046] hardirqs last disabled at (150582): [<ffffffff81c4b130>] _raw_spin_lock_irqsave+0x50/0x60 [ 227.576793] softirqs last enabled at (150576): [<ffffffffa00bcbef>] ipoib_mcast_join_task+0xef/0x320 [ib_ipoib] [ 227.578657] softirqs last disabled at (150574): [<ffffffffa00bcb9f>] ipoib_mcast_join_task+0x9f/0x320 [ib_ipoib] [ 227.580416] ---[ end trace 52a055538498ac03 ]---
Quoted 2 lines
Thanks
Hi Leon,
Thanks for the test, it seems that the problem is caused by the refcount of mcast_group. I will reply to this email once I have made progress.
Weihang