return (s16)(fp->bp->tx_ring_size) - used;
}
-static void bnx2x_tx_int(struct bnx2x_fastpath *fp, int work)
+static void bnx2x_tx_int(struct bnx2x_fastpath *fp)
{
struct bnx2x *bp = fp->bp;
struct netdev_queue *txq;
bd_cons = bnx2x_free_tx_pkt(bp, fp, pkt_cons);
sw_cons++;
done++;
-
- if (done == work)
- break;
}
fp->tx_pkt_cons = sw_cons;
struct bnx2x_fastpath *fp = &bp->fp[0];
int rc;
- bnx2x_tx_int(fp, 1000);
+ bnx2x_tx_int(fp);
rc = bnx2x_rx_int(fp, 1000);
}
cnt = 1000;
while (bnx2x_has_tx_work_unload(fp)) {
- bnx2x_tx_int(fp, 1000);
+ bnx2x_tx_int(fp);
if (!cnt) {
BNX2X_ERR("timeout waiting for queue[%d]\n",
i);
bnx2x_update_fpsb_idx(fp);
if (bnx2x_has_tx_work(fp))
- bnx2x_tx_int(fp, budget);
+ bnx2x_tx_int(fp);
- if (bnx2x_has_rx_work(fp))
+ if (bnx2x_has_rx_work(fp)) {
work_done = bnx2x_rx_int(fp, budget);
- rmb(); /* BNX2X_HAS_WORK() reads the status block */
+ /* must not complete if we consumed full budget */
+ if (work_done >= budget)
+ goto poll_again;
+ }
- /* must not complete if we consumed full budget */
- if ((work_done < budget) && !BNX2X_HAS_WORK(fp)) {
+ /* BNX2X_HAS_WORK() reads the status block, thus we need to
+ * ensure that status block indices have been actually read
+ * (bnx2x_update_fpsb_idx) prior to this check (BNX2X_HAS_WORK)
+ * so that we won't write the "newer" value of the status block to IGU
+ * (if there was a DMA right after BNX2X_HAS_WORK and
+ * if there is no rmb, the memory reading (bnx2x_update_fpsb_idx)
+ * may be postponed to right before bnx2x_ack_sb). In this case
+ * there will never be another interrupt until there is another update
+ * of the status block, while there is still unhandled work.
+ */
+ rmb();
+ if (!BNX2X_HAS_WORK(fp)) {
#ifdef BNX2X_STOP_ON_ERROR
poll_panic:
#endif
le16_to_cpu(fp->fp_c_idx), IGU_INT_ENABLE, 1);
}
+poll_again:
return work_done;
}