cxgb4/ch_ktls: creating skbs causes panic

Creating SKB per tls record and freeing the original one causes
panic. There will be race if connection reset is requested. By
freeing original skb, refcnt will be decremented and that means,
there is no pending record to send, and so tls_dev_del will be
requested in control path while SKB of related connection is in
queue.
 Better approach is to use same SKB to send one record (partial
data) at a time. We still have to create a new SKB when partial
last part of a record is requested.
 This fix introduces new API cxgb4_write_partial_sgl() to send
partial part of skb. Present cxgb4_write_sgl can only provide
feasibility to start from an offset which limits to header only
and it can write sgls for the whole skb len. But this new API
will help in both. It can start from any offset and can end
writing in middle of the skb.

v4->v5:
- Removed extra changes.

Fixes: 429765a1 ("chcr: handle partial end part of a record")
Signed-off-by: Rohit Maheshwari <rohitm@chelsio.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

drivers/net/ethernet/chelsio/cxgb4/cxgb4.h

@@ -2124,6 +2124,9 @@ void cxgb4_inline_tx_skb(const struct sk_buff *skb, const struct sge_txq *q,
 void cxgb4_write_sgl(const struct sk_buff *skb, struct sge_txq *q,
 		     struct ulptx_sgl *sgl, u64 *end, unsigned int start,
 		     const dma_addr_t *addr);
+void cxgb4_write_partial_sgl(const struct sk_buff *skb, struct sge_txq *q,
+			     struct ulptx_sgl *sgl, u64 *end,
+			     const dma_addr_t *addr, u32 start, u32 send_len);
 void cxgb4_ring_tx_db(struct adapter *adap, struct sge_txq *q, int n);
 int t4_set_vlan_acl(struct adapter *adap, unsigned int mbox, unsigned int vf,
 		    u16 vlan);

drivers/net/ethernet/chelsio/cxgb4/sge.c

@@ -890,6 +890,114 @@ void cxgb4_write_sgl(const struct sk_buff *skb, struct sge_txq *q,
 }
 EXPORT_SYMBOL(cxgb4_write_sgl);
 
+/*	cxgb4_write_partial_sgl - populate SGL for partial packet
+ *	@skb: the packet
+ *	@q: the Tx queue we are writing into
+ *	@sgl: starting location for writing the SGL
+ *	@end: points right after the end of the SGL
+ *	@addr: the list of bus addresses for the SGL elements
+ *	@start: start offset in the SKB where partial data starts
+ *	@len: length of data from @start to send out
+ *
+ *	This API will handle sending out partial data of a skb if required.
+ *	Unlike cxgb4_write_sgl, @start can be any offset into the skb data,
+ *	and @len will decide how much data after @start offset to send out.
+ */
+void cxgb4_write_partial_sgl(const struct sk_buff *skb, struct sge_txq *q,
+			     struct ulptx_sgl *sgl, u64 *end,
+			     const dma_addr_t *addr, u32 start, u32 len)
+{
+	struct ulptx_sge_pair buf[MAX_SKB_FRAGS / 2 + 1] = {0}, *to;
+	u32 frag_size, skb_linear_data_len = skb_headlen(skb);
+	struct skb_shared_info *si = skb_shinfo(skb);
+	u8 i = 0, frag_idx = 0, nfrags = 0;
+	skb_frag_t *frag;
+
+	/* Fill the first SGL either from linear data or from partial
+	 * frag based on @start.
+	 */
+	if (unlikely(start < skb_linear_data_len)) {
+		frag_size = min(len, skb_linear_data_len - start);
+		sgl->len0 = htonl(frag_size);
+		sgl->addr0 = cpu_to_be64(addr[0] + start);
+		len -= frag_size;
+		nfrags++;
+	} else {
+		start -= skb_linear_data_len;
+		frag = &si->frags[frag_idx];
+		frag_size = skb_frag_size(frag);
+		/* find the first frag */
+		while (start >= frag_size) {
+			start -= frag_size;
+			frag_idx++;
+			frag = &si->frags[frag_idx];
+			frag_size = skb_frag_size(frag);
+		}
+
+		frag_size = min(len, skb_frag_size(frag) - start);
+		sgl->len0 = cpu_to_be32(frag_size);
+		sgl->addr0 = cpu_to_be64(addr[frag_idx + 1] + start);
+		len -= frag_size;
+		nfrags++;
+		frag_idx++;
+	}
+
+	/* If the entire partial data fit in one SGL, then send it out
+	 * now.
+	 */
+	if (!len)
+		goto done;
+
+	/* Most of the complexity below deals with the possibility we hit the
+	 * end of the queue in the middle of writing the SGL.  For this case
+	 * only we create the SGL in a temporary buffer and then copy it.
+	 */
+	to = (u8 *)end > (u8 *)q->stat ? buf : sgl->sge;
+
+	/* If the skb couldn't fit in first SGL completely, fill the
+	 * rest of the frags in subsequent SGLs. Note that each SGL
+	 * pair can store 2 frags.
+	 */
+	while (len) {
+		frag_size = min(len, skb_frag_size(&si->frags[frag_idx]));
+		to->len[i & 1] = cpu_to_be32(frag_size);
+		to->addr[i & 1] = cpu_to_be64(addr[frag_idx + 1]);
+		if (i && (i & 1))
+			to++;
+		nfrags++;
+		frag_idx++;
+		i++;
+		len -= frag_size;
+	}
+
+	/* If we ended in an odd boundary, then set the second SGL's
+	 * length in the pair to 0.
+	 */
+	if (i & 1)
+		to->len[1] = cpu_to_be32(0);
+
+	/* Copy from temporary buffer to Tx ring, in case we hit the
+	 * end of the queue in the middle of writing the SGL.
+	 */
+	if (unlikely((u8 *)end > (u8 *)q->stat)) {
+		u32 part0 = (u8 *)q->stat - (u8 *)sgl->sge, part1;
+
+		if (likely(part0))
+			memcpy(sgl->sge, buf, part0);
+		part1 = (u8 *)end - (u8 *)q->stat;
+		memcpy(q->desc, (u8 *)buf + part0, part1);
+		end = (void *)q->desc + part1;
+	}
+
+	/* 0-pad to multiple of 16 */
+	if ((uintptr_t)end & 8)
+		*end = 0;
+done:
+	sgl->cmd_nsge = htonl(ULPTX_CMD_V(ULP_TX_SC_DSGL) |
+			ULPTX_NSGE_V(nfrags));
+}
+EXPORT_SYMBOL(cxgb4_write_partial_sgl);
+
 /* This function copies 64 byte coalesced work request to
  * memory mapped BAR2 space. For coalesced WR SGE fetches
  * data from the FIFO instead of from Host.