]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
qed: make 'ethtool -d' 10 times faster
authorMichal Schmidt <mschmidt@redhat.com>
Mon, 30 Sep 2024 20:13:04 +0000 (22:13 +0200)
committerJakub Kicinski <kuba@kernel.org>
Fri, 4 Oct 2024 16:25:14 +0000 (09:25 -0700)
As a side effect of commit 5401c3e09928 ("qed: allow sleep in
qed_mcp_trace_dump()"), 'ethtool -d' became much slower.
Almost all the time is spent collecting the "mcp_trace".
It is caused by sleeping too long in _qed_mcp_cmd_and_union.
When called with sleeping not allowed, the function delays for 10 µs
between firmware polls. But if sleeping is allowed, it sleeps for 10 ms
instead.

The sleeps in _qed_mcp_cmd_and_union are unnecessarily long.
Replace msleep with usleep_range, which allows to achieve a similar
polling interval like in the no-sleeping mode (10 - 20 µs).

The only caller, qed_mcp_cmd_and_union, can stop doing the
multiplication/division of the usecs/max_retries. The polling interval
and the number of retries do not need to be parameters at all.

On my test system, 'ethtool -d' now takes 4 seconds instead of 44.

Signed-off-by: Michal Schmidt <mschmidt@redhat.com>
Link: https://patch.msgid.link/20240930201307.330692-2-mschmidt@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
drivers/net/ethernet/qlogic/qed/qed_mcp.c

index 16e6bd4661433f9b6d4c216eed055605df4bfe3c..00f0abc1c2d26df5ba988624bfc7f2b624da1176 100644 (file)
@@ -459,12 +459,11 @@ static void qed_mcp_print_cpu_info(struct qed_hwfn *p_hwfn,
 static int
 _qed_mcp_cmd_and_union(struct qed_hwfn *p_hwfn,
                       struct qed_ptt *p_ptt,
-                      struct qed_mcp_mb_params *p_mb_params,
-                      u32 max_retries, u32 usecs)
+                      struct qed_mcp_mb_params *p_mb_params)
 {
-       u32 cnt = 0, msecs = DIV_ROUND_UP(usecs, 1000);
        struct qed_mcp_cmd_elem *p_cmd_elem;
        u16 seq_num;
+       u32 cnt = 0;
        int rc = 0;
 
        /* Wait until the mailbox is non-occupied */
@@ -488,12 +487,13 @@ _qed_mcp_cmd_and_union(struct qed_hwfn *p_hwfn,
                spin_unlock_bh(&p_hwfn->mcp_info->cmd_lock);
 
                if (QED_MB_FLAGS_IS_SET(p_mb_params, CAN_SLEEP))
-                       msleep(msecs);
+                       usleep_range(QED_MCP_RESP_ITER_US,
+                                    QED_MCP_RESP_ITER_US * 2);
                else
-                       udelay(usecs);
-       } while (++cnt < max_retries);
+                       udelay(QED_MCP_RESP_ITER_US);
+       } while (++cnt < QED_DRV_MB_MAX_RETRIES);
 
-       if (cnt >= max_retries) {
+       if (cnt >= QED_DRV_MB_MAX_RETRIES) {
                DP_NOTICE(p_hwfn,
                          "The MFW mailbox is occupied by an uncompleted command. Failed to send command 0x%08x [param 0x%08x].\n",
                          p_mb_params->cmd, p_mb_params->param);
@@ -520,9 +520,10 @@ _qed_mcp_cmd_and_union(struct qed_hwfn *p_hwfn,
                 */
 
                if (QED_MB_FLAGS_IS_SET(p_mb_params, CAN_SLEEP))
-                       msleep(msecs);
+                       usleep_range(QED_MCP_RESP_ITER_US,
+                                    QED_MCP_RESP_ITER_US * 2);
                else
-                       udelay(usecs);
+                       udelay(QED_MCP_RESP_ITER_US);
 
                spin_lock_bh(&p_hwfn->mcp_info->cmd_lock);
 
@@ -536,9 +537,9 @@ _qed_mcp_cmd_and_union(struct qed_hwfn *p_hwfn,
                        goto err;
 
                spin_unlock_bh(&p_hwfn->mcp_info->cmd_lock);
-       } while (++cnt < max_retries);
+       } while (++cnt < QED_DRV_MB_MAX_RETRIES);
 
-       if (cnt >= max_retries) {
+       if (cnt >= QED_DRV_MB_MAX_RETRIES) {
                DP_NOTICE(p_hwfn,
                          "The MFW failed to respond to command 0x%08x [param 0x%08x].\n",
                          p_mb_params->cmd, p_mb_params->param);
@@ -564,7 +565,8 @@ _qed_mcp_cmd_and_union(struct qed_hwfn *p_hwfn,
                   "MFW mailbox: response 0x%08x param 0x%08x [after %d.%03d ms]\n",
                   p_mb_params->mcp_resp,
                   p_mb_params->mcp_param,
-                  (cnt * usecs) / 1000, (cnt * usecs) % 1000);
+                  (cnt * QED_MCP_RESP_ITER_US) / 1000,
+                  (cnt * QED_MCP_RESP_ITER_US) % 1000);
 
        /* Clear the sequence number from the MFW response */
        p_mb_params->mcp_resp &= FW_MSG_CODE_MASK;
@@ -581,8 +583,6 @@ static int qed_mcp_cmd_and_union(struct qed_hwfn *p_hwfn,
                                 struct qed_mcp_mb_params *p_mb_params)
 {
        size_t union_data_size = sizeof(union drv_union_data);
-       u32 max_retries = QED_DRV_MB_MAX_RETRIES;
-       u32 usecs = QED_MCP_RESP_ITER_US;
 
        /* MCP not initialized */
        if (!qed_mcp_is_init(p_hwfn)) {
@@ -606,13 +606,7 @@ static int qed_mcp_cmd_and_union(struct qed_hwfn *p_hwfn,
                return -EINVAL;
        }
 
-       if (QED_MB_FLAGS_IS_SET(p_mb_params, CAN_SLEEP)) {
-               max_retries = DIV_ROUND_UP(max_retries, 1000);
-               usecs *= 1000;
-       }
-
-       return _qed_mcp_cmd_and_union(p_hwfn, p_ptt, p_mb_params, max_retries,
-                                     usecs);
+       return _qed_mcp_cmd_and_union(p_hwfn, p_ptt, p_mb_params);
 }
 
 static int _qed_mcp_cmd(struct qed_hwfn *p_hwfn,