qed: fix handling of concurrent ramrods.
authorTomer Tayar <Tomer.Tayar@qlogic.com>
Mon, 7 Dec 2015 11:25:57 +0000 (06:25 -0500)
committerDavid S. Miller <davem@davemloft.net>
Mon, 7 Dec 2015 19:14:03 +0000 (14:14 -0500)
Concurrent non-blocking slowpath ramrods can be completed
out-of-order on the completion chain. Recycling completed elements,
while previously sent elements are still completion pending,
can lead to overriding of active elements on the chain. Furthermore,
sending pending slowpath ramrods currently lacks the update of the
chain element physical pointer.

This patch:
* Ensures that ramrods are sent to the FW with
  consecutive echo values.
* Handles out-of-order completions by freeing only first
  successive completed entries.
* Updates the chain element physical pointer when copying
  a pending element into a free element for sending.

Signed-off-by: Tomer Tayar <Tomer.Tayar@qlogic.com>
Signed-off-by: Manish Chopra <manish.chopra@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/qlogic/qed/qed_sp.h
drivers/net/ethernet/qlogic/qed/qed_spq.c
include/linux/qed/common_hsi.h

index 31a1f1eb4f56a5d00f0ce3ae75a29af56a96ef12..287fadfab52d754e0a913d4ed044ca7e50cf1a09 100644 (file)
@@ -124,8 +124,12 @@ struct qed_spq {
        dma_addr_t              p_phys;
        struct qed_spq_entry    *p_virt;
 
-       /* Used as index for completions (returns on EQ by FW) */
-       u16                     echo_idx;
+#define SPQ_RING_SIZE \
+       (CORE_SPQE_PAGE_SIZE_BYTES / sizeof(struct slow_path_element))
+
+       /* Bitmap for handling out-of-order completions */
+       DECLARE_BITMAP(p_comp_bitmap, SPQ_RING_SIZE);
+       u8                      comp_bitmap_idx;
 
        /* Statistics */
        u32                     unlimited_pending_count;
index 7c0b8459666ed270ec754a87c9fa1c072d16dd49..3dd548ab8df14aabf9fa6a551cf54aa97f704d12 100644 (file)
@@ -112,8 +112,6 @@ static int
 qed_spq_fill_entry(struct qed_hwfn *p_hwfn,
                   struct qed_spq_entry *p_ent)
 {
-       p_ent->elem.hdr.echo = 0;
-       p_hwfn->p_spq->echo_idx++;
        p_ent->flags = 0;
 
        switch (p_ent->comp_mode) {
@@ -195,10 +193,12 @@ static int qed_spq_hw_post(struct qed_hwfn *p_hwfn,
                           struct qed_spq *p_spq,
                           struct qed_spq_entry *p_ent)
 {
-       struct qed_chain                *p_chain = &p_hwfn->p_spq->chain;
+       struct qed_chain *p_chain = &p_hwfn->p_spq->chain;
+       u16 echo = qed_chain_get_prod_idx(p_chain);
        struct slow_path_element        *elem;
        struct core_db_data             db;
 
+       p_ent->elem.hdr.echo    = cpu_to_le16(echo);
        elem = qed_chain_produce(p_chain);
        if (!elem) {
                DP_NOTICE(p_hwfn, "Failed to produce from SPQ chain\n");
@@ -437,7 +437,9 @@ void qed_spq_setup(struct qed_hwfn *p_hwfn)
        p_spq->comp_count               = 0;
        p_spq->comp_sent_count          = 0;
        p_spq->unlimited_pending_count  = 0;
-       p_spq->echo_idx                 = 0;
+
+       bitmap_zero(p_spq->p_comp_bitmap, SPQ_RING_SIZE);
+       p_spq->comp_bitmap_idx = 0;
 
        /* SPQ cid, cannot fail */
        qed_cxt_acquire_cid(p_hwfn, PROTOCOLID_CORE, &p_spq->cid);
@@ -582,26 +584,32 @@ qed_spq_add_entry(struct qed_hwfn *p_hwfn,
        struct qed_spq *p_spq = p_hwfn->p_spq;
 
        if (p_ent->queue == &p_spq->unlimited_pending) {
-               struct qed_spq_entry *p_en2;
 
                if (list_empty(&p_spq->free_pool)) {
                        list_add_tail(&p_ent->list, &p_spq->unlimited_pending);
                        p_spq->unlimited_pending_count++;
 
                        return 0;
-               }
+               } else {
+                       struct qed_spq_entry *p_en2;
 
-               p_en2 = list_first_entry(&p_spq->free_pool,
-                                        struct qed_spq_entry,
-                                        list);
-               list_del(&p_en2->list);
+                       p_en2 = list_first_entry(&p_spq->free_pool,
+                                                struct qed_spq_entry,
+                                                list);
+                       list_del(&p_en2->list);
+
+                       /* Copy the ring element physical pointer to the new
+                        * entry, since we are about to override the entire ring
+                        * entry and don't want to lose the pointer.
+                        */
+                       p_ent->elem.data_ptr = p_en2->elem.data_ptr;
 
-               /* Strcut assignment */
-               *p_en2 = *p_ent;
+                       *p_en2 = *p_ent;
 
-               kfree(p_ent);
+                       kfree(p_ent);
 
-               p_ent = p_en2;
+                       p_ent = p_en2;
+               }
        }
 
        /* entry is to be placed in 'pending' queue */
@@ -777,13 +785,38 @@ int qed_spq_completion(struct qed_hwfn *p_hwfn,
        list_for_each_entry_safe(p_ent, tmp, &p_spq->completion_pending,
                                 list) {
                if (p_ent->elem.hdr.echo == echo) {
+                       u16 pos = le16_to_cpu(echo) % SPQ_RING_SIZE;
+
                        list_del(&p_ent->list);
 
-                       qed_chain_return_produced(&p_spq->chain);
+                       /* Avoid overriding of SPQ entries when getting
+                        * out-of-order completions, by marking the completions
+                        * in a bitmap and increasing the chain consumer only
+                        * for the first successive completed entries.
+                        */
+                       bitmap_set(p_spq->p_comp_bitmap, pos, SPQ_RING_SIZE);
+
+                       while (test_bit(p_spq->comp_bitmap_idx,
+                                       p_spq->p_comp_bitmap)) {
+                               bitmap_clear(p_spq->p_comp_bitmap,
+                                            p_spq->comp_bitmap_idx,
+                                            SPQ_RING_SIZE);
+                               p_spq->comp_bitmap_idx++;
+                               qed_chain_return_produced(&p_spq->chain);
+                       }
+
                        p_spq->comp_count++;
                        found = p_ent;
                        break;
                }
+
+               /* This is relatively uncommon - depends on scenarios
+                * which have mutliple per-PF sent ramrods.
+                */
+               DP_VERBOSE(p_hwfn, QED_MSG_SPQ,
+                          "Got completion for echo %04x - doesn't match echo %04x in completion pending list\n",
+                          le16_to_cpu(echo),
+                          le16_to_cpu(p_ent->elem.hdr.echo));
        }
 
        /* Release lock before callback, as callback may post
index 6a4347639c0329235a735d723a564acecf59f022..1d1ba2c5ee7a26f6a35b1835c016a6bd7755fcd8 100644 (file)
@@ -9,6 +9,8 @@
 #ifndef __COMMON_HSI__
 #define __COMMON_HSI__
 
+#define CORE_SPQE_PAGE_SIZE_BYTES                       4096
+
 #define FW_MAJOR_VERSION       8
 #define FW_MINOR_VERSION       4
 #define FW_REVISION_VERSION    2