net: qed: fix NVMe login fails over VFs

25ms sleep cycles in waiting for PF response are excessive and may lead
to different timeout failures.

Start to wait with short udelays, and in most cases polling will end
here. If the time was not sufficient, switch to msleeps.
usleep_range() may go far beyond 100us depending on platform and tick
configuration, hence atomic udelays for consistency.

Also add explicit DMA barriers since 'done' always comes from a shared
request-response DMA pool, and note that in the comment nearby.

Fixes: 1408cc1f ("qed: Introduce VFs")
Signed-off-by: Alexander Lobakin <alobakin@marvell.com>
Signed-off-by: Igor Russkikh <irusskikh@marvell.com>
Signed-off-by: Michal Kalderon <michal.kalderon@marvell.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/qlogic/qed/qed_vf.c

@@ -81,12 +81,17 @@ static void qed_vf_pf_req_end(struct qed_hwfn *p_hwfn, int req_status)
 	mutex_unlock(&(p_hwfn->vf_iov_info->mutex));
 }
 
+#define QED_VF_CHANNEL_USLEEP_ITERATIONS	90
+#define QED_VF_CHANNEL_USLEEP_DELAY		100
+#define QED_VF_CHANNEL_MSLEEP_ITERATIONS	10
+#define QED_VF_CHANNEL_MSLEEP_DELAY		25
+
 static int qed_send_msg2pf(struct qed_hwfn *p_hwfn, u8 *done, u32 resp_size)
 {
 	union vfpf_tlvs *p_req = p_hwfn->vf_iov_info->vf2pf_request;
 	struct ustorm_trigger_vf_zone trigger;
 	struct ustorm_vf_zone *zone_data;
-	int rc = 0, time = 100;
+	int iter, rc = 0;
 
 	zone_data = (struct ustorm_vf_zone *)PXP_VF_BAR0_START_USDM_ZONE_B;
 
@@ -126,11 +131,19 @@ static int qed_send_msg2pf(struct qed_hwfn *p_hwfn, u8 *done, u32 resp_size)
 	REG_WR(p_hwfn, (uintptr_t)&zone_data->trigger, *((u32 *)&trigger));
 
 	/* When PF would be done with the response, it would write back to the
-	 * `done' address. Poll until then.
+	 * `done' address from a coherent DMA zone. Poll until then.
 	 */
-	while ((!*done) && time) {
-		msleep(25);
-		time--;
+
+	iter = QED_VF_CHANNEL_USLEEP_ITERATIONS;
+	while (!*done && iter--) {
+		udelay(QED_VF_CHANNEL_USLEEP_DELAY);
+		dma_rmb();
+	}
+
+	iter = QED_VF_CHANNEL_MSLEEP_ITERATIONS;
+	while (!*done && iter--) {
+		msleep(QED_VF_CHANNEL_MSLEEP_DELAY);
+		dma_rmb();
 	}
 
 	if (!*done) {