iwlwifi: pcie: work around DMA hardware bug

There's a hardware bug in the flow handler (DMA engine), if the
address + len of some TB wraps around a 2^32 boundary, the carry
bit is then carried over into the next TB.

Work around this by copying the data to a new page when we find
this situation, and then copy it in a way that we cannot hit the
very end of the page.

To be able to free the new page again later we need to chain it
to the TSO page, use the last pointer there to make sure we can
never use the page fully for DMA, and thus cannot cause the same
overflow situation on this page.

This leaves a few potential places (where we didn't observe the
problem) unaddressed:
 * The second TB could reach or cross the end of a page (and thus
   2^32) due to the way we allocate the dev_cmd for the header
 * For host commands, a similar thing could happen since they're
   just kmalloc().
We'll address these in further commits.
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: Luca Coelho <luciano.coelho@intel.com>

drivers/net/wireless/intel/iwlwifi/pcie/tx-gen2.c

@@ -213,6 +213,16 @@ static void iwl_pcie_gen2_free_tfd(struct iwl_trans *trans, struct iwl_txq *txq)
 	}
 }
 
+/*
+ * We need this inline in case dma_addr_t is only 32-bits - since the
+ * hardware is always 64-bit, the issue can still occur in that case,
+ * so use u64 for 'phys' here to force the addition in 64-bit.
+ */
+static inline bool crosses_4g_boundary(u64 phys, u16 len)
+{
+	return upper_32_bits(phys) != upper_32_bits(phys + len);
+}
+
 static int iwl_pcie_gen2_set_tb(struct iwl_trans *trans,
 				struct iwl_tfh_tfd *tfd, dma_addr_t addr,
 				u16 len)
@@ -240,6 +250,107 @@ static int iwl_pcie_gen2_set_tb(struct iwl_trans *trans,
 	return idx;
 }
 
+static struct page *get_workaround_page(struct iwl_trans *trans,
+					struct sk_buff *skb)
+{
+	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
+	struct page **page_ptr;
+	struct page *ret;
+
+	page_ptr = (void *)((u8 *)skb->cb + trans_pcie->page_offs);
+
+	ret = alloc_page(GFP_ATOMIC);
+	if (!ret)
+		return NULL;
+
+	/* set the chaining pointer to the previous page if there */
+	*(void **)(page_address(ret) + PAGE_SIZE - sizeof(void *)) = *page_ptr;
+	*page_ptr = ret;
+
+	return ret;
+}
+
+/*
+ * Add a TB and if needed apply the FH HW bug workaround;
+ * meta != NULL indicates that it's a page mapping and we
+ * need to dma_unmap_page() and set the meta->tbs bit in
+ * this case.
+ */
+static int iwl_pcie_gen2_set_tb_with_wa(struct iwl_trans *trans,
+					struct sk_buff *skb,
+					struct iwl_tfh_tfd *tfd,
+					dma_addr_t phys, void *virt,
+					u16 len, struct iwl_cmd_meta *meta)
+{
+	dma_addr_t oldphys = phys;
+	struct page *page;
+	int ret;
+
+	if (unlikely(dma_mapping_error(trans->dev, phys)))
+		return -ENOMEM;
+
+	if (likely(!crosses_4g_boundary(phys, len))) {
+		ret = iwl_pcie_gen2_set_tb(trans, tfd, phys, len);
+
+		if (ret < 0)
+			goto unmap;
+
+		if (meta)
+			meta->tbs |= BIT(ret);
+
+		ret = 0;
+		goto trace;
+	}
+
+	/*
+	 * Work around a hardware bug. If (as expressed in the
+	 * condition above) the TB ends on a 32-bit boundary,
+	 * then the next TB may be accessed with the wrong
+	 * address.
+	 * To work around it, copy the data elsewhere and make
+	 * a new mapping for it so the device will not fail.
+	 */
+
+	if (WARN_ON(len > PAGE_SIZE - sizeof(void *))) {
+		ret = -ENOBUFS;
+		goto unmap;
+	}
+
+	page = get_workaround_page(trans, skb);
+	if (!page) {
+		ret = -ENOMEM;
+		goto unmap;
+	}
+
+	memcpy(page_address(page), virt, len);
+
+	phys = dma_map_single(trans->dev, page_address(page), len,
+			      DMA_TO_DEVICE);
+	if (unlikely(dma_mapping_error(trans->dev, phys)))
+		return -ENOMEM;
+	ret = iwl_pcie_gen2_set_tb(trans, tfd, phys, len);
+	if (ret < 0) {
+		/* unmap the new allocation as single */
+		oldphys = phys;
+		meta = NULL;
+		goto unmap;
+	}
+	IWL_WARN(trans,
+		 "TB bug workaround: copied %d bytes from 0x%llx to 0x%llx\n",
+		 len, (unsigned long long)oldphys, (unsigned long long)phys);
+
+	ret = 0;
+unmap:
+	if (meta)
+		dma_unmap_page(trans->dev, oldphys, len, DMA_TO_DEVICE);
+	else
+		dma_unmap_single(trans->dev, oldphys, len, DMA_TO_DEVICE);
+trace:
+	trace_iwlwifi_dev_tx_tb(trans->dev, skb, virt, phys, len);
+
+	return ret;
+}
+
 static int iwl_pcie_gen2_build_amsdu(struct iwl_trans *trans,
 				     struct sk_buff *skb,
 				     struct iwl_tfh_tfd *tfd, int start_len,
@@ -327,6 +438,11 @@ static int iwl_pcie_gen2_build_amsdu(struct iwl_trans *trans,
 			dev_kfree_skb(csum_skb);
 			goto out_err;
 		}
+		/*
+		 * No need for _with_wa, this is from the TSO page and
+		 * we leave some space at the end of it so can't hit
+		 * the buggy scenario.
+		 */
 		iwl_pcie_gen2_set_tb(trans, tfd, tb_phys, tb_len);
 		trace_iwlwifi_dev_tx_tb(trans->dev, skb, start_hdr,
 					tb_phys, tb_len);
@@ -338,16 +454,18 @@ static int iwl_pcie_gen2_build_amsdu(struct iwl_trans *trans,
 
 		/* put the payload */
 		while (data_left) {
+			int ret;
+
 			tb_len = min_t(unsigned int, tso.size, data_left);
 			tb_phys = dma_map_single(trans->dev, tso.data,
 						 tb_len, DMA_TO_DEVICE);
-			if (unlikely(dma_mapping_error(trans->dev, tb_phys))) {
+			ret = iwl_pcie_gen2_set_tb_with_wa(trans, skb, tfd,
+							   tb_phys, tso.data,
+							   tb_len, NULL);
+			if (ret) {
 				dev_kfree_skb(csum_skb);
 				goto out_err;
 			}
-			iwl_pcie_gen2_set_tb(trans, tfd, tb_phys, tb_len);
-			trace_iwlwifi_dev_tx_tb(trans->dev, skb, tso.data,
-						tb_phys, tb_len);
 
 			data_left -= tb_len;
 			tso_build_data(skb, &tso, tb_len);
@@ -381,6 +499,11 @@ iwl_tfh_tfd *iwl_pcie_gen2_build_tx_amsdu(struct iwl_trans *trans,
 
 	tb_phys = iwl_pcie_get_first_tb_dma(txq, idx);
 
+	/*
+	 * No need for _with_wa, the first TB allocation is aligned up
+	 * to a 64-byte boundary and thus can't be at the end or cross
+	 * a page boundary (much less a 2^32 boundary).
+	 */
 	iwl_pcie_gen2_set_tb(trans, tfd, tb_phys, IWL_FIRST_TB_SIZE);
 
 	/*
@@ -425,24 +548,19 @@ static int iwl_pcie_gen2_tx_add_frags(struct iwl_trans *trans,
 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 		dma_addr_t tb_phys;
-		int tb_idx;
+		unsigned int fragsz = skb_frag_size(frag);
+		int ret;
 
-		if (!skb_frag_size(frag))
+		if (!fragsz)
 			continue;
 
 		tb_phys = skb_frag_dma_map(trans->dev, frag, 0,
-					   skb_frag_size(frag), DMA_TO_DEVICE);
-
-		if (unlikely(dma_mapping_error(trans->dev, tb_phys)))
-			return -ENOMEM;
-		tb_idx = iwl_pcie_gen2_set_tb(trans, tfd, tb_phys,
-					      skb_frag_size(frag));
-		trace_iwlwifi_dev_tx_tb(trans->dev, skb, skb_frag_address(frag),
-					tb_phys, skb_frag_size(frag));
-		if (tb_idx < 0)
-			return tb_idx;
-
-		out_meta->tbs |= BIT(tb_idx);
+					   fragsz, DMA_TO_DEVICE);
+		ret = iwl_pcie_gen2_set_tb_with_wa(trans, skb, tfd, tb_phys,
+						   skb_frag_address(frag),
+						   fragsz, out_meta);
+		if (ret)
+			return ret;
 	}
 
 	return 0;
@@ -470,6 +588,11 @@ iwl_tfh_tfd *iwl_pcie_gen2_build_tx(struct iwl_trans *trans,
 	/* The first TB points to bi-directional DMA data */
 	memcpy(&txq->first_tb_bufs[idx], dev_cmd, IWL_FIRST_TB_SIZE);
 
+	/*
+	 * No need for _with_wa, the first TB allocation is aligned up
+	 * to a 64-byte boundary and thus can't be at the end or cross
+	 * a page boundary (much less a 2^32 boundary).
+	 */
 	iwl_pcie_gen2_set_tb(trans, tfd, tb_phys, IWL_FIRST_TB_SIZE);
 
 	/*
@@ -499,26 +622,30 @@ iwl_tfh_tfd *iwl_pcie_gen2_build_tx(struct iwl_trans *trans,
 	tb2_len = skb_headlen(skb) - hdr_len;
 
 	if (tb2_len > 0) {
+		int ret;
+
 		tb_phys = dma_map_single(trans->dev, skb->data + hdr_len,
 					 tb2_len, DMA_TO_DEVICE);
-		if (unlikely(dma_mapping_error(trans->dev, tb_phys)))
+		ret = iwl_pcie_gen2_set_tb_with_wa(trans, skb, tfd, tb_phys,
+						   skb->data + hdr_len, tb2_len,
+						   NULL);
+		if (ret)
 			goto out_err;
-		iwl_pcie_gen2_set_tb(trans, tfd, tb_phys, tb2_len);
-		trace_iwlwifi_dev_tx_tb(trans->dev, skb, skb->data + hdr_len,
-					tb_phys, tb2_len);
 	}
 
 	if (iwl_pcie_gen2_tx_add_frags(trans, skb, tfd, out_meta))
 		goto out_err;
 
 	skb_walk_frags(skb, frag) {
+		int ret;
+
 		tb_phys = dma_map_single(trans->dev, frag->data,
 					 skb_headlen(frag), DMA_TO_DEVICE);
-		if (unlikely(dma_mapping_error(trans->dev, tb_phys)))
+		ret = iwl_pcie_gen2_set_tb_with_wa(trans, skb, tfd, tb_phys,
+						   frag->data,
+						   skb_headlen(frag), NULL);
+		if (ret)
 			goto out_err;
-		iwl_pcie_gen2_set_tb(trans, tfd, tb_phys, skb_headlen(frag));
-		trace_iwlwifi_dev_tx_tb(trans->dev, skb, frag->data,
-					tb_phys, skb_headlen(frag));
 		if (iwl_pcie_gen2_tx_add_frags(trans, frag, tfd, out_meta))
 			goto out_err;
 	}

drivers/net/wireless/intel/iwlwifi/pcie/tx.c

@@ -624,12 +624,18 @@ void iwl_pcie_free_tso_page(struct iwl_trans_pcie *trans_pcie,
 			    struct sk_buff *skb)
 {
 	struct page **page_ptr;
+	struct page *next;
 
 	page_ptr = (void *)((u8 *)skb->cb + trans_pcie->page_offs);
+	next = *page_ptr;
+	*page_ptr = NULL;
 
-	if (*page_ptr) {
-		__free_page(*page_ptr);
-		*page_ptr = NULL;
+	while (next) {
+		struct page *tmp = next;
+
+		next = *(void **)(page_address(next) + PAGE_SIZE -
+				  sizeof(void *));
+		__free_page(tmp);
 	}
 }
 
@@ -2067,8 +2073,18 @@ struct iwl_tso_hdr_page *get_page_hdr(struct iwl_trans *trans, size_t len,
 	if (!p->page)
 		goto alloc;
 
-	/* enough room on this page */
-	if (p->pos + len < (u8 *)page_address(p->page) + PAGE_SIZE)
+	/*
+	 * Check if there's enough room on this page
+	 *
+	 * Note that we put a page chaining pointer *last* in the
+	 * page - we need it somewhere, and if it's there then we
+	 * avoid DMA mapping the last bits of the page which may
+	 * trigger the 32-bit boundary hardware bug.
+	 *
+	 * (see also get_workaround_page() in tx-gen2.c)
+	 */
+	if (p->pos + len < (u8 *)page_address(p->page) + PAGE_SIZE -
+			   sizeof(void *))
 		goto out;
 
 	/* We don't have enough room on this page, get a new one. */
@@ -2079,6 +2095,8 @@ struct iwl_tso_hdr_page *get_page_hdr(struct iwl_trans *trans, size_t len,
 	if (!p->page)
 		return NULL;
 	p->pos = page_address(p->page);
+	/* set the chaining pointer to NULL */
+	*(void **)(page_address(p->page) + PAGE_SIZE - sizeof(void *)) = NULL;
 out:
 	*page_ptr = p->page;
 	get_page(p->page);