idpf: convert control queue mutex to a spinlock

[ Upstream commit b2beb5b ]

With VIRTCHNL2_CAP_MACFILTER enabled, the following warning is generated
on module load:

[  324.701677] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:578
[  324.701684] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 1582, name: NetworkManager
[  324.701689] preempt_count: 201, expected: 0
[  324.701693] RCU nest depth: 0, expected: 0
[  324.701697] 2 locks held by NetworkManager/1582:
[  324.701702]  #0: ffffffff9f7be770 (rtnl_mutex){....}-{3:3}, at: rtnl_newlink+0x791/0x21e0
[  324.701730]  #1: ff1100216c380368 (_xmit_ETHER){....}-{2:2}, at: __dev_open+0x3f0/0x870
[  324.701749] Preemption disabled at:
[  324.701752] [<ffffffff9cd23b9d>] __dev_open+0x3dd/0x870
[  324.701765] CPU: 30 UID: 0 PID: 1582 Comm: NetworkManager Not tainted 6.15.0-rc5+ #2 PREEMPT(voluntary)
[  324.701771] Hardware name: Intel Corporation M50FCP2SBSTD/M50FCP2SBSTD, BIOS SE5C741.86B.01.01.0001.2211140926 11/14/2022
[  324.701774] Call Trace:
[  324.701777]  <TASK>
[  324.701779]  dump_stack_lvl+0x5d/0x80
[  324.701788]  ? __dev_open+0x3dd/0x870
[  324.701793]  __might_resched.cold+0x1ef/0x23d
<..>
[  324.701818]  __mutex_lock+0x113/0x1b80
<..>
[  324.701917]  idpf_ctlq_clean_sq+0xad/0x4b0 [idpf]
[  324.701935]  ? kasan_save_track+0x14/0x30
[  324.701941]  idpf_mb_clean+0x143/0x380 [idpf]
<..>
[  324.701991]  idpf_send_mb_msg+0x111/0x720 [idpf]
[  324.702009]  idpf_vc_xn_exec+0x4cc/0x990 [idpf]
[  324.702021]  ? rcu_is_watching+0x12/0xc0
[  324.702035]  idpf_add_del_mac_filters+0x3ed/0xb50 [idpf]
<..>
[  324.702122]  __hw_addr_sync_dev+0x1cf/0x300
[  324.702126]  ? find_held_lock+0x32/0x90
[  324.702134]  idpf_set_rx_mode+0x317/0x390 [idpf]
[  324.702152]  __dev_open+0x3f8/0x870
[  324.702159]  ? __pfx___dev_open+0x10/0x10
[  324.702174]  __dev_change_flags+0x443/0x650
<..>
[  324.702208]  netif_change_flags+0x80/0x160
[  324.702218]  do_setlink.isra.0+0x16a0/0x3960
<..>
[  324.702349]  rtnl_newlink+0x12fd/0x21e0

The sequence is as follows:
	rtnl_newlink()->
	__dev_change_flags()->
	__dev_open()->
	dev_set_rx_mode() - >  # disables BH and grabs "dev->addr_list_lock"
	idpf_set_rx_mode() ->  # proceed only if VIRTCHNL2_CAP_MACFILTER is ON
	__dev_uc_sync() ->
	idpf_add_mac_filter ->
	idpf_add_del_mac_filters ->
	idpf_send_mb_msg() ->
	idpf_mb_clean() ->
	idpf_ctlq_clean_sq()   # mutex_lock(cq_lock)

Fix by converting cq_lock to a spinlock. All operations under the new
lock are safe except freeing the DMA memory, which may use vunmap(). Fix
by requesting a contiguous physical memory for the DMA mapping.

Fixes: a251eee ("idpf: add SRIOV support and other ndo_ops")
Reviewed-by: Aleksandr Loktionov <aleksandr.loktionov@intel.com>
Signed-off-by: Ahmed Zaki <ahmed.zaki@intel.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Tested-by: Samuel Salin <Samuel.salin@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>

drivers/net/ethernet/intel/idpf/idpf_controlq.c

@@ -96,16 +96,15 @@ static void idpf_ctlq_init_rxq_bufs(struct idpf_ctlq_info *cq)
  */
 static void idpf_ctlq_shutdown(struct idpf_hw *hw, struct idpf_ctlq_info *cq)
 {
-	mutex_lock(&cq->cq_lock);
+	spin_lock(&cq->cq_lock);
 
 	/* free ring buffers and the ring itself */
 	idpf_ctlq_dealloc_ring_res(hw, cq);
 
 	/* Set ring_size to 0 to indicate uninitialized queue */
 	cq->ring_size = 0;
 
-	mutex_unlock(&cq->cq_lock);
-	mutex_destroy(&cq->cq_lock);
+	spin_unlock(&cq->cq_lock);
 }
 
 /**
@@ -173,7 +172,7 @@ int idpf_ctlq_add(struct idpf_hw *hw,
 
 	idpf_ctlq_init_regs(hw, cq, is_rxq);
 
-	mutex_init(&cq->cq_lock);
+	spin_lock_init(&cq->cq_lock);
 
 	list_add(&cq->cq_list, &hw->cq_list_head);
 
@@ -272,7 +271,7 @@ int idpf_ctlq_send(struct idpf_hw *hw, struct idpf_ctlq_info *cq,
 	int err = 0;
 	int i;
 
-	mutex_lock(&cq->cq_lock);
+	spin_lock(&cq->cq_lock);
 
 	/* Ensure there are enough descriptors to send all messages */
 	num_desc_avail = IDPF_CTLQ_DESC_UNUSED(cq);
@@ -332,7 +331,7 @@ int idpf_ctlq_send(struct idpf_hw *hw, struct idpf_ctlq_info *cq,
 	wr32(hw, cq->reg.tail, cq->next_to_use);
 
 err_unlock:
-	mutex_unlock(&cq->cq_lock);
+	spin_unlock(&cq->cq_lock);
 
 	return err;
 }
@@ -364,7 +363,7 @@ int idpf_ctlq_clean_sq(struct idpf_ctlq_info *cq, u16 *clean_count,
 	if (*clean_count > cq->ring_size)
 		return -EBADR;
 
-	mutex_lock(&cq->cq_lock);
+	spin_lock(&cq->cq_lock);
 
 	ntc = cq->next_to_clean;
 
@@ -397,7 +396,7 @@ int idpf_ctlq_clean_sq(struct idpf_ctlq_info *cq, u16 *clean_count,
 
 	cq->next_to_clean = ntc;
 
-	mutex_unlock(&cq->cq_lock);
+	spin_unlock(&cq->cq_lock);
 
 	/* Return number of descriptors actually cleaned */
 	*clean_count = i;
@@ -435,7 +434,7 @@ int idpf_ctlq_post_rx_buffs(struct idpf_hw *hw, struct idpf_ctlq_info *cq,
 	if (*buff_count > 0)
 		buffs_avail = true;
 
-	mutex_lock(&cq->cq_lock);
+	spin_lock(&cq->cq_lock);
 
 	if (tbp >= cq->ring_size)
 		tbp = 0;
@@ -524,7 +523,7 @@ int idpf_ctlq_post_rx_buffs(struct idpf_hw *hw, struct idpf_ctlq_info *cq,
 		wr32(hw, cq->reg.tail, cq->next_to_post);
 	}
 
-	mutex_unlock(&cq->cq_lock);
+	spin_unlock(&cq->cq_lock);
 
 	/* return the number of buffers that were not posted */
 	*buff_count = *buff_count - i;
@@ -552,7 +551,7 @@ int idpf_ctlq_recv(struct idpf_ctlq_info *cq, u16 *num_q_msg,
 	u16 i;
 
 	/* take the lock before we start messing with the ring */
-	mutex_lock(&cq->cq_lock);
+	spin_lock(&cq->cq_lock);
 
 	ntc = cq->next_to_clean;
 
@@ -614,7 +613,7 @@ int idpf_ctlq_recv(struct idpf_ctlq_info *cq, u16 *num_q_msg,
 
 	cq->next_to_clean = ntc;
 
-	mutex_unlock(&cq->cq_lock);
+	spin_unlock(&cq->cq_lock);
 
 	*num_q_msg = i;
 	if (*num_q_msg == 0)

drivers/net/ethernet/intel/idpf/idpf_controlq_api.h

@@ -99,7 +99,7 @@ struct idpf_ctlq_info {
 
 	enum idpf_ctlq_type cq_type;
 	int q_id;
-	struct mutex cq_lock;		/* control queue lock */
+	spinlock_t cq_lock;		/* control queue lock */
 	/* used for interrupt processing */
 	u16 next_to_use;
 	u16 next_to_clean;

drivers/net/ethernet/intel/idpf/idpf_lib.c

@@ -2315,8 +2315,12 @@ void *idpf_alloc_dma_mem(struct idpf_hw *hw, struct idpf_dma_mem *mem, u64 size)
 	struct idpf_adapter *adapter = hw->back;
 	size_t sz = ALIGN(size, 4096);
 
-	mem->va = dma_alloc_coherent(&adapter->pdev->dev, sz,
-				     &mem->pa, GFP_KERNEL);
+	/* The control queue resources are freed under a spinlock, contiguous
+	 * pages will avoid IOMMU remapping and the use vmap (and vunmap in
+	 * dma_free_*() path.
+	 */
+	mem->va = dma_alloc_attrs(&adapter->pdev->dev, sz, &mem->pa,
+				  GFP_KERNEL, DMA_ATTR_FORCE_CONTIGUOUS);
 	mem->size = sz;
 
 	return mem->va;
@@ -2331,8 +2335,8 @@ void idpf_free_dma_mem(struct idpf_hw *hw, struct idpf_dma_mem *mem)
 {
 	struct idpf_adapter *adapter = hw->back;
 
-	dma_free_coherent(&adapter->pdev->dev, mem->size,
-			  mem->va, mem->pa);
+	dma_free_attrs(&adapter->pdev->dev, mem->size,
+		       mem->va, mem->pa, DMA_ATTR_FORCE_CONTIGUOUS);
 	mem->size = 0;
 	mem->va = NULL;
 	mem->pa = 0;