net/qede/base: avoid possible race condition
authorRasesh Mody <rasesh.mody@cavium.com>
Tue, 19 Sep 2017 01:29:58 +0000 (18:29 -0700)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 6 Oct 2017 00:49:49 +0000 (02:49 +0200)
There's a possible race in multiple VF scenarios for base driver users
that use the optional APIs ecore_iov_pf_get_and_clear_pending_events,
ecore_iov_pf_add_pending_events. If the client doesn't synchronize the two
calls, it's possible for the PF to clear a VF pending message indication
without ever getting it [as 'get & clear' isn't atomic], leading to VF
timeout on the command.

The solution is to switch into a per-VF indication rather than having a
bitfield for the various VFs with pending events. As part of the solution,
the setting/clearing of the indications is done internally by base driver.
As a result, ecore_iov_pf_add_pending_events is no longer needed and
ecore_iov_pf_get_and_clear_pending_events loses the 'and_clear' from its
name as its now a proper getter.

A VF would be considered 'pending' [I.e., get_pending_events() should
have '1' for it in its bitfield] beginning with the PF's base driver
recognizing a message sent by that VF [in SP_DPC] and ending only when
that VF message is processed.

Signed-off-by: Rasesh Mody <rasesh.mody@cavium.com>
drivers/net/qede/base/ecore_iov_api.h
drivers/net/qede/base/ecore_sriov.c
drivers/net/qede/base/ecore_sriov.h

index 4fce6b6..4ec6217 100644 (file)
@@ -345,21 +345,13 @@ ecore_iov_get_public_vf_info(struct ecore_hwfn *p_hwfn,
                             u16 vfid, bool b_enabled_only);
 
 /**
- * @brief Set pending events bitmap for given @vfid
- *
- * @param p_hwfn
- * @param vfid
- */
-void ecore_iov_pf_add_pending_events(struct ecore_hwfn *p_hwfn, u8 vfid);
-
-/**
- * @brief Copy pending events bitmap in @events and clear
- *       original copy of events
+ * @brief fills a bitmask of all VFs which have pending unhandled
+ *        messages.
  *
  * @param p_hwfn
  */
-void ecore_iov_pf_get_and_clear_pending_events(struct ecore_hwfn *p_hwfn,
-                                              u64 *events);
+void ecore_iov_pf_get_pending_events(struct ecore_hwfn *p_hwfn,
+                                    u64 *events);
 
 /**
  * @brief Copy VF's message to PF's buffer
index b8e03d0..7ac533e 100644 (file)
@@ -3755,8 +3755,7 @@ cleanup:
                ack_vfs[vfid / 32] |= (1 << (vfid % 32));
                p_hwfn->pf_iov_info->pending_flr[rel_vf_id / 64] &=
                    ~(1ULL << (rel_vf_id % 64));
-               p_hwfn->pf_iov_info->pending_events[rel_vf_id / 64] &=
-                   ~(1ULL << (rel_vf_id % 64));
+               p_vf->vf_mbx.b_pending_msg = false;
        }
 
        return rc;
@@ -3886,12 +3885,22 @@ void ecore_iov_process_mbx_req(struct ecore_hwfn *p_hwfn,
        mbx = &p_vf->vf_mbx;
 
        /* ecore_iov_process_mbx_request */
-       DP_VERBOSE(p_hwfn,
-                  ECORE_MSG_IOV,
-                  "VF[%02x]: Processing mailbox message\n", p_vf->abs_vf_id);
+#ifndef CONFIG_ECORE_SW_CHANNEL
+       if (!mbx->b_pending_msg) {
+               DP_NOTICE(p_hwfn, true,
+                         "VF[%02x]: Trying to process mailbox message when none is pending\n",
+                         p_vf->abs_vf_id);
+               return;
+       }
+       mbx->b_pending_msg = false;
+#endif
 
        mbx->first_tlv = mbx->req_virt->first_tlv;
 
+       DP_VERBOSE(p_hwfn, ECORE_MSG_IOV,
+                  "VF[%02x]: Processing mailbox message [type %04x]\n",
+                  p_vf->abs_vf_id, mbx->first_tlv.tl.type);
+
        OSAL_IOV_VF_MSG_TYPE(p_hwfn,
                             p_vf->relative_vf_id,
                             mbx->first_tlv.tl.type);
@@ -4016,26 +4025,20 @@ void ecore_iov_process_mbx_req(struct ecore_hwfn *p_hwfn,
 #endif
 }
 
-void ecore_iov_pf_add_pending_events(struct ecore_hwfn *p_hwfn, u8 vfid)
+void ecore_iov_pf_get_pending_events(struct ecore_hwfn *p_hwfn,
+                                    u64 *events)
 {
-       u64 add_bit = 1ULL << (vfid % 64);
+       int i;
 
-       /* TODO - add locking mechanisms [no atomics in ecore, so we can't
-       * add the lock inside the ecore_pf_iov struct].
-       */
-       p_hwfn->pf_iov_info->pending_events[vfid / 64] |= add_bit;
-}
+       OSAL_MEM_ZERO(events, sizeof(u64) * ECORE_VF_ARRAY_LENGTH);
 
-void ecore_iov_pf_get_and_clear_pending_events(struct ecore_hwfn *p_hwfn,
-                                              u64 *events)
-{
-       u64 *p_pending_events = p_hwfn->pf_iov_info->pending_events;
+       ecore_for_each_vf(p_hwfn, i) {
+               struct ecore_vf_info *p_vf;
 
-       /* TODO - Take a lock */
-       OSAL_MEMCPY(events, p_pending_events,
-                   sizeof(u64) * ECORE_VF_ARRAY_LENGTH);
-       OSAL_MEMSET(p_pending_events, 0,
-                   sizeof(u64) * ECORE_VF_ARRAY_LENGTH);
+               p_vf = &p_hwfn->pf_iov_info->vfs_array[i];
+               if (p_vf->vf_mbx.b_pending_msg)
+                       events[i / 64] |= 1ULL << (i % 64);
+       }
 }
 
 static struct ecore_vf_info *
@@ -4069,6 +4072,8 @@ static enum _ecore_status_t ecore_sriov_vfpf_msg(struct ecore_hwfn *p_hwfn,
         */
        p_vf->vf_mbx.pending_req = (((u64)vf_msg->hi) << 32) | vf_msg->lo;
 
+       p_vf->vf_mbx.b_pending_msg = true;
+
        return OSAL_PF_VF_MSG(p_hwfn, p_vf->relative_vf_id);
 }
 
index d190126..8923730 100644 (file)
@@ -45,6 +45,9 @@ struct ecore_iov_vf_mbx {
        /* Address in VF where a pending message is located */
        dma_addr_t              pending_req;
 
+       /* Message from VF awaits handling */
+       bool                    b_pending_msg;
+
        u8 *offset;
 
 #ifdef CONFIG_ECORE_SW_CHANNEL
@@ -168,7 +171,6 @@ struct ecore_vf_info {
  */
 struct ecore_pf_iov {
        struct ecore_vf_info    vfs_array[E4_MAX_NUM_VFS];
-       u64                     pending_events[ECORE_VF_ARRAY_LENGTH];
        u64                     pending_flr[ECORE_VF_ARRAY_LENGTH];
 
 #ifndef REMOVE_DBG