Skip to contentSkip to conversation

Convert skb to use scatterlist

RFC In review Stephen Hemminger 6 participants 16 messages Jul 5, 2007 – Jul 11, 2007

Subsystems (9)
Networking Network drivers netfilter InfiniBand +5 more subsystems
lore
Share & tools
lore

Stephen Hemminger

author
#0 lore raw

cover Convert skb to use scatterlist

This topic came up at first Netconf. This patch changes skbuff to use scatterlist. Why? Devices can than use the pci_dma_sg routines to map the fraglist in one operation. This allows for better error handling (less unwinding), and some IOMMU's (PPC?) can be smarter.

Signature · 1 line
Stephen Hemminger <[email protected]>
author
#1 lore raw

RFC 1/2 skbuff: use scatterlist

Replace the skb frag list with the common scatterlist definition. This allows device drivers to use dma_scatter/gather operations which may be faster on some platforms. As a side benefit, it is easier to handle dma mapping error unwind.

This idea came up long ago, just never got implemented.

Signed-off-by Stephen Hemminger <[email protected]>
---
 drivers/atm/he.c                        |    2 -
 drivers/infiniband/hw/amso1100/c2.c     |    2 -
 drivers/net/3c59x.c                     |    8 ++--
 drivers/net/8139cp.c                    |    5 +-
 drivers/net/acenic.c                    |    8 ++--
 drivers/net/atl1/atl1_main.c            |    8 ++--
 drivers/net/bnx2.c                      |    8 ++--
 drivers/net/cassini.c                   |   19 ++++-----
 drivers/net/chelsio/sge.c               |    8 ++--
 drivers/net/cxgb3/adapter.h             |    2 -
 drivers/net/cxgb3/sge.c                 |   25 ++++++------
 drivers/net/e1000/e1000_main.c          |    8 ++--
 drivers/net/ehea/ehea_main.c            |    7 +--
 drivers/net/forcedeth.c                 |   22 ++++++-----
 drivers/net/ibm_emac/ibm_emac_core.c    |    2 -
 drivers/net/ibmveth.c                   |    2 -
 drivers/net/ixgb/ixgb_main.c            |    6 +--
 drivers/net/mv643xx_eth.c               |    2 -
 drivers/net/myri10ge/myri10ge.c         |   26 ++++++-------
 drivers/net/netxen/netxen_nic_main.c    |    6 +--
 drivers/net/ns83820.c                   |    9 ++--
 drivers/net/qla3xxx.c                   |    6 +--
 drivers/net/r8169.c                     |    4 +-
 drivers/net/s2io.c                      |   18 +++++----
 drivers/net/sk98lin/skge.c              |    8 ++--
 drivers/net/skge.c                      |    8 ++--
 drivers/net/sky2.c                      |   16 ++++----
 drivers/net/starfire.c                  |    9 +++-
 drivers/net/sungem.c                    |    4 +-
 drivers/net/sunhme.c                    |    4 +-
 drivers/net/tg3.c                       |   14 +++----
 drivers/net/tsi108_eth.c                |    2 -
 drivers/net/typhoon.c                   |    2 -
 drivers/net/via-velocity.c              |    2 -
 include/linux/skbuff.h                  |   21 ++++------
 net/appletalk/ddp.c                     |    4 +-
 net/core/datagram.c                     |    9 ++--
 net/core/pktgen.c                       |   42 ++++++++++-----------
 net/core/skbuff.c                       |   64 ++++++++++++++++----------------
 net/core/sock.c                         |    8 ++--
 net/core/user_dma.c                     |    2 -
 net/ipv4/ip_fragment.c                  |    4 +-
 net/ipv4/ip_output.c                    |    9 +++-
 net/ipv4/tcp.c                          |    9 ++--
 net/ipv4/tcp_output.c                   |    8 ++--
 net/ipv6/ip6_output.c                   |    7 ++-
 net/ipv6/netfilter/nf_conntrack_reasm.c |    2 -
 net/ipv6/reassembly.c                   |    2 -
 net/xfrm/xfrm_algo.c                    |    4 +-
 49 files changed, 239 insertions(+), 238 deletions(-)
include/linux/skbuff.h +8 -13
@@ -21,6 +21,7 @@
 
 #include <asm/atomic.h>
 #include <asm/types.h>
+#include <asm/scatterlist.h>
 #include <linux/spinlock.h>
 #include <linux/net.h>
 #include <linux/textsearch.h>

@@ -122,13 +123,7 @@ struct sk_buff;
 /* To allow 64K frame to be packed as single skb without frag_list */
 #define MAX_SKB_FRAGS (65536/PAGE_SIZE + 2)
 
-typedef struct skb_frag_struct skb_frag_t;
-
-struct skb_frag_struct {
-	struct page *page;
-	__u16 page_offset;
-	__u16 size;
-};
+typedef struct scatterlist skb_frag_t;
 
 /* This data is invariant across clones and lives at
  * the end of the header data, ie. at skb->end.

@@ -813,7 +808,7 @@ static inline int skb_pagelen(const stru
 	int i, len = 0;
 
 	for (i = (int)skb_shinfo(skb)->nr_frags - 1; i >= 0; i--)
-		len += skb_shinfo(skb)->frags[i].size;
+		len += skb_shinfo(skb)->frags[i].length;
 	return len + skb_headlen(skb);
 }
 

@@ -822,9 +817,9 @@ static inline void skb_fill_page_desc(st
 {
 	skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 
-	frag->page		  = page;
-	frag->page_offset	  = off;
-	frag->size		  = size;
+	frag->page	  = page;
+	frag->offset	  = off;
+	frag->length	  = size;
 	skb_shinfo(skb)->nr_frags = i + 1;
 }
 

@@ -1390,10 +1385,10 @@ static inline int skb_can_coalesce(struc
 				   struct page *page, int off)
 {
 	if (i) {
-		struct skb_frag_struct *frag = &skb_shinfo(skb)->frags[i - 1];
+		skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 
 		return page == frag->page &&
-		       off == frag->page_offset + frag->size;
+		       off == frag->offset + frag->length;
 	}
 	return 0;
 }
drivers/atm/he.c +1 -1
@@ -2803,7 +2803,7 @@ he_send(struct atm_vcc *vcc, struct sk_b
 		}
 
 		tpd->iovec[slot].addr = pci_map_single(he_dev->pci_dev,
-			(void *) page_address(frag->page) + frag->page_offset,
+			(void *) page_address(frag->page) + frag->offset,
 				frag->size, PCI_DMA_TODEVICE);
 		tpd->iovec[slot].len = frag->size;
 		++slot;
drivers/infiniband/hw/amso1100/c2.c +1 -1
@@ -801,7 +801,7 @@ static int c2_xmit_frame(struct sk_buff 
 			maplen = frag->size;
 			mapaddr =
 			    pci_map_page(c2dev->pcidev, frag->page,
-					 frag->page_offset, maplen,
+					 frag->offset, maplen,
 					 PCI_DMA_TODEVICE);
 
 			elem = elem->next;
drivers/net/3c59x.c +4 -4
@@ -2102,13 +2102,13 @@ boomerang_start_xmit(struct sk_buff *skb
 
 			vp->tx_ring[entry].frag[i+1].addr =
 					cpu_to_le32(pci_map_single(VORTEX_PCI(vp),
-											   (void*)page_address(frag->page) + frag->page_offset,
-											   frag->size, PCI_DMA_TODEVICE));
+											   (void*)page_address(frag->page) + frag->offset,
+											   frag->length, PCI_DMA_TODEVICE));
 
 			if (i == skb_shinfo(skb)->nr_frags-1)
-					vp->tx_ring[entry].frag[i+1].length = cpu_to_le32(frag->size|LAST_FRAG);
+					vp->tx_ring[entry].frag[i+1].length = cpu_to_le32(frag->length|LAST_FRAG);
 			else
-					vp->tx_ring[entry].frag[i+1].length = cpu_to_le32(frag->size);
+					vp->tx_ring[entry].frag[i+1].length = cpu_to_le32(frag->length);
 		}
 	}
 #else
drivers/net/8139cp.c +2 -3
@@ -831,14 +831,13 @@ static int cp_start_xmit (struct sk_buff
 
 		for (frag = 0; frag < skb_shinfo(skb)->nr_frags; frag++) {
 			skb_frag_t *this_frag = &skb_shinfo(skb)->frags[frag];
-			u32 len;
+			u32 len  = this_frag->length;
 			u32 ctrl;
 			dma_addr_t mapping;
 
-			len = this_frag->size;
 			mapping = pci_map_single(cp->pdev,
 						 ((void *) page_address(this_frag->page) +
-						  this_frag->page_offset),
+						  this_frag->offset),
 						 len, PCI_DMA_TODEVICE);
 			eor = (entry == (CP_TX_RING_SIZE - 1)) ? RingEnd : 0;
drivers/net/acenic.c +4 -4
@@ -2528,15 +2528,15 @@ restart:
 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 			struct tx_ring_info *info;
 
-			len += frag->size;
+			len += frag->length;
 			info = ap->skb->tx_skbuff + idx;
 			desc = ap->tx_ring + idx;
 
 			mapping = pci_map_page(ap->pdev, frag->page,
-					       frag->page_offset, frag->size,
+					       frag->offset, frag->length,
 					       PCI_DMA_TODEVICE);
 
-			flagsize = (frag->size << 16);
+			flagsize = (frag->length << 16);
 			if (skb->ip_summed == CHECKSUM_PARTIAL)
 				flagsize |= BD_FLG_TCP_UDP_SUM;
 			idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);

@@ -2555,7 +2555,7 @@ restart:
 				info->skb = NULL;
 			}
 			pci_unmap_addr_set(info, mapping, mapping);
-			pci_unmap_len_set(info, maplen, frag->size);
+			pci_unmap_len_set(info, maplen, frag->length);
 			ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
 		}
 	}
drivers/net/atl1/atl1_main.c +4 -4
@@ -1384,11 +1384,11 @@ static void atl1_tx_map(struct atl1_adap
 	}
 
 	for (f = 0; f < nr_frags; f++) {
-		struct skb_frag_struct *frag;
+		skb_frag_t *frag;
 		u16 lenf, i, m;
 
 		frag = &skb_shinfo(skb)->frags[f];
-		lenf = frag->size;
+		lenf = frag->length;
 
 		m = (lenf + MAX_TX_BUF_LEN - 1) / MAX_TX_BUF_LEN;
 		for (i = 0; i < m; i++) {

@@ -1401,7 +1401,7 @@ static void atl1_tx_map(struct atl1_adap
 			lenf -= buffer_info->length;
 			buffer_info->dma =
 			    pci_map_page(adapter->pdev, frag->page,
-					 frag->page_offset + i * MAX_TX_BUF_LEN,
+					 frag->offset + i * MAX_TX_BUF_LEN,
 					 buffer_info->length, PCI_DMA_TODEVICE);
 
 			if (++tpd_next_to_use == tpd_ring->count)

@@ -1516,7 +1516,7 @@ static int atl1_xmit_frame(struct sk_buf
 	/* nr_frags will be nonzero if we're doing scatter/gather (SG) */
 	nr_frags = skb_shinfo(skb)->nr_frags;
 	for (f = 0; f < nr_frags; f++) {
-		frag_size = skb_shinfo(skb)->frags[f].size;
+		frag_size = skb_shinfo(skb)->frags[f].length;
 		if (frag_size)
 			count +=
 			    (frag_size + MAX_TX_BUF_LEN - 1) / MAX_TX_BUF_LEN;
drivers/net/bnx2.c +4 -4
@@ -2038,7 +2038,7 @@ bnx2_tx_int(struct bnx2 *bp)
 				pci_unmap_addr(
 					&bp->tx_buf_ring[TX_RING_IDX(sw_cons)],
 				       	mapping),
-				skb_shinfo(skb)->frags[i].size,
+				skb_shinfo(skb)->frags[i].length,
 				PCI_DMA_TODEVICE);
 		}
 

@@ -4001,7 +4001,7 @@ bnx2_free_tx_skbs(struct bnx2 *bp)
 			tx_buf = &bp->tx_buf_ring[i + j + 1];
 			pci_unmap_page(bp->pdev,
 				pci_unmap_addr(tx_buf, mapping),
-				skb_shinfo(skb)->frags[j].size,
+				skb_shinfo(skb)->frags[j].length,
 				PCI_DMA_TODEVICE);
 		}
 		dev_kfree_skb(skb);

@@ -4922,8 +4922,8 @@ bnx2_start_xmit(struct sk_buff *skb, str
 		ring_prod = TX_RING_IDX(prod);
 		txbd = &bp->tx_desc_ring[ring_prod];
 
-		len = frag->size;
-		mapping = pci_map_page(bp->pdev, frag->page, frag->page_offset,
+		len = frag->length;
+		mapping = pci_map_page(bp->pdev, frag->page, frag->offset,
 			len, PCI_DMA_TODEVICE);
 		pci_unmap_addr_set(&bp->tx_buf_ring[ring_prod],
 				mapping, mapping);
drivers/net/cassini.c +9 -10
@@ -2067,8 +2067,8 @@ static int cas_rx_process_pkt(struct cas
 		get_page(page->buffer);
 		cas_buffer_inc(page);
 		frag->page = page->buffer;
-		frag->page_offset = off;
-		frag->size = hlen - swivel;
+		frag->offset = off;
+		frag->length = hlen - swivel;
 
 		/* any more data? */
 		if ((words[0] & RX_COMP1_SPLIT_PKT) && ((dlen -= hlen) > 0)) {

@@ -2092,8 +2092,8 @@ static int cas_rx_process_pkt(struct cas
 			get_page(page->buffer);
 			cas_buffer_inc(page);
 			frag->page = page->buffer;
-			frag->page_offset = 0;
-			frag->size = hlen;
+			frag->offset = 0;
+			frag->length = hlen;
 			RX_USED_ADD(page, hlen + cp->crc_size);
 		}
 

@@ -2860,12 +2860,11 @@ static inline int cas_xmit_tx_ringN(stru
 	for (frag = 0; frag < nr_frags; frag++) {
 		skb_frag_t *fragp = &skb_shinfo(skb)->frags[frag];
 
-		len = fragp->size;
-		mapping = pci_map_page(cp->pdev, fragp->page,
-				       fragp->page_offset, len,
-				       PCI_DMA_TODEVICE);
+		len = fragp->length;
+		mapping = pci_map_page(cp->pdev, fragp->page, fragp->offset,
+				       len, PCI_DMA_TODEVICE);
 
-		tabort = cas_calc_tabort(cp, fragp->page_offset, len);
+		tabort = cas_calc_tabort(cp, fragp->offset, len);
 		if (unlikely(tabort)) {
 			void *addr;
 

@@ -2876,7 +2875,7 @@ static inline int cas_xmit_tx_ringN(stru
 
 			addr = cas_page_map(fragp->page);
 			memcpy(tx_tiny_buf(cp, ring, entry),
-			       addr + fragp->page_offset + len - tabort,
+			       addr + fragp->offset + len - tabort,
 			       tabort);
 			cas_page_unmap(addr);
 			mapping = tx_tiny_map(cp, ring, entry, tentry);
drivers/net/chelsio/sge.c +4 -4
@@ -1130,7 +1130,7 @@ static inline unsigned int compute_large
 		}
 		for (i = 0; nfrags--; i++) {
 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
-			len = frag->size;
+			len = frag->length;
 			while (len > SGE_TX_DESC_MAX_PLEN) {
 				count++;
 				len -= SGE_TX_DESC_MAX_PLEN;

@@ -1272,10 +1272,10 @@ static inline void write_tx_descs(struct
 		}
 
 		mapping = pci_map_page(adapter->pdev, frag->page,
-				       frag->page_offset, frag->size,
+				       frag->offset, frag->length,
 				       PCI_DMA_TODEVICE);
 		desc_mapping = mapping;
-		desc_len = frag->size;
+		desc_len = frag->length;
 
 		pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
 						 &desc_mapping, &desc_len,

@@ -1285,7 +1285,7 @@ static inline void write_tx_descs(struct
 				      nfrags == 0);
 		ce->skb = NULL;
 		pci_unmap_addr_set(ce, dma_addr, mapping);
-		pci_unmap_len_set(ce, dma_len, frag->size);
+		pci_unmap_len_set(ce, dma_len, frag->length);
 	}
 	ce->skb = skb;
 	wmb();
drivers/net/cxgb3/sge.c +12 -13
@@ -246,7 +246,7 @@ static inline void unmap_skb(struct sk_b
 
 	while (frag_idx < nfrags && curflit < WR_FLITS) {
 		pci_unmap_page(pdev, be64_to_cpu(sgp->addr[j]),
-			       skb_shinfo(skb)->frags[frag_idx].size,
+			       skb_shinfo(skb)->frags[frag_idx].length,
 			       PCI_DMA_TODEVICE);
 		j ^= 1;
 		if (j == 0) {

@@ -433,8 +433,8 @@ static void refill_fl(struct adapter *ad
 					q->alloc_failed++;
 					break;
 				} else {
-					p->frag.size = RX_PAGE_SIZE;
-					p->frag.page_offset = 0;
+					p->frag.length = RX_PAGE_SIZE;
+					p->frag.offset = 0;
 					p->va = page_address(p->frag.page);
 				}
 			}

@@ -442,10 +442,10 @@ static void refill_fl(struct adapter *ad
 			memcpy(&sd->t, p, sizeof(*p));
 			va = p->va;
 
-			p->frag.page_offset += RX_PAGE_SIZE;
-			BUG_ON(p->frag.page_offset > PAGE_SIZE);
+			p->frag.offset += RX_PAGE_SIZE;
+			BUG_ON(p->frag.offset > PAGE_SIZE);
 			p->va += RX_PAGE_SIZE;
-			if (p->frag.page_offset == PAGE_SIZE)
+			if (p->frag.offset == PAGE_SIZE)
 				p->frag.page = NULL;
 			else
 				get_page(p->frag.page);

@@ -716,9 +716,9 @@ static inline unsigned int make_sgl(cons
 	for (i = 0; i < nfrags; i++) {
 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 
-		mapping = pci_map_page(pdev, frag->page, frag->page_offset,
-				       frag->size, PCI_DMA_TODEVICE);
-		sgp->len[j] = cpu_to_be32(frag->size);
+		mapping = pci_map_page(pdev, frag->page, frag->offset,
+				       frag->length, PCI_DMA_TODEVICE);
+		sgp->len[j] = cpu_to_be32(frag->length);
 		sgp->addr[j] = cpu_to_be64(mapping);
 		j ^= 1;
 		if (j == 0)

@@ -1270,7 +1270,7 @@ static void deferred_unmap_destructor(st
 
 	si = skb_shinfo(skb);
 	for (i = 0; i < si->nr_frags; i++)
-		pci_unmap_page(dui->pdev, *p++, si->frags[i].size,
+		pci_unmap_page(dui->pdev, *p++, si->frags[i].length,
 			       PCI_DMA_TODEVICE);
 }
 

@@ -1728,9 +1728,8 @@ static void skb_data_init(struct sk_buff
 	} else {
 		skb_copy_to_linear_data(skb, p->va, SKB_DATA_SIZE);
 		skb_shinfo(skb)->frags[0].page = p->frag.page;
-		skb_shinfo(skb)->frags[0].page_offset =
-		    p->frag.page_offset + SKB_DATA_SIZE;
-		skb_shinfo(skb)->frags[0].size = len - SKB_DATA_SIZE;
+		skb_shinfo(skb)->frags[0].offset = p->frag.offset + SKB_DATA_SIZE;
+		skb_shinfo(skb)->frags[0].length = len - SKB_DATA_SIZE;
 		skb_shinfo(skb)->nr_frags = 1;
 		skb->data_len = len - SKB_DATA_SIZE;
 		skb->tail += SKB_DATA_SIZE;
drivers/net/e1000/e1000_main.c +4 -4
@@ -3049,11 +3049,11 @@ e1000_tx_map(struct e1000_adapter *adapt
 	}
 
 	for (f = 0; f < nr_frags; f++) {
-		struct skb_frag_struct *frag;
+		skb_frag_t *frag;
 
 		frag = &skb_shinfo(skb)->frags[f];
-		len = frag->size;
-		offset = frag->page_offset;
+		len = frag->length;
+		offset = frag->offset;
 
 		while (len) {
 			buffer_info = &tx_ring->buffer_info[i];

@@ -3358,7 +3358,7 @@ e1000_xmit_frame(struct sk_buff *skb, st
 
 	nr_frags = skb_shinfo(skb)->nr_frags;
 	for (f = 0; f < nr_frags; f++)
-		count += TXD_USE_COUNT(skb_shinfo(skb)->frags[f].size,
+		count += TXD_USE_COUNT(skb_shinfo(skb)->frags[f].length,
 				       max_txd_pwr);
 	if (adapter->pcix_82544)
 		count += nr_frags;
drivers/net/ehea/ehea_main.c +3 -4
@@ -1390,7 +1390,7 @@ static inline void write_swqe2_data(stru
 			sg1entry->l_key = lkey;
 			sg1entry->len = frag->size;
 			tmp_addr =  (u64)(page_address(frag->page)
-					  + frag->page_offset);
+					  + frag->offset);
 			sg1entry->vaddr = tmp_addr;
 			swqe->descriptors++;
 			sg1entry_contains_frag_data = 1;

@@ -1404,8 +1404,7 @@ static inline void write_swqe2_data(stru
 			sgentry->l_key = lkey;
 			sgentry->len = frag->size;
 
-			tmp_addr = (u64)(page_address(frag->page)
-					 + frag->page_offset);
+			tmp_addr = (u64)(page_address(frag->page) + frag->offset);
 			sgentry->vaddr = tmp_addr;
 			swqe->descriptors++;
 		}

@@ -1789,7 +1788,7 @@ static void ehea_xmit3(struct sk_buff *s
 		for (i = 0; i < nfrags; i++) {
 			frag = &skb_shinfo(skb)->frags[i];
 			memcpy(imm_data,
-			       page_address(frag->page) + frag->page_offset,
+			       page_address(frag->page) + frag->offset,
 			       frag->size);
 			imm_data += frag->size;
 		}
drivers/net/forcedeth.c +12 -10
@@ -1649,8 +1649,8 @@ static int nv_start_xmit(struct sk_buff 
 
 	/* add fragments to entries count */
 	for (i = 0; i < fragments; i++) {
-		entries += (skb_shinfo(skb)->frags[i].size >> NV_TX2_TSO_MAX_SHIFT) +
-			   ((skb_shinfo(skb)->frags[i].size & (NV_TX2_TSO_MAX_SIZE-1)) ? 1 : 0);
+		entries += (skb_shinfo(skb)->frags[i].length >> NV_TX2_TSO_MAX_SHIFT) +
+			   ((skb_shinfo(skb)->frags[i].length & (NV_TX2_TSO_MAX_SIZE-1)) ? 1 : 0);
 	}
 
 	empty_slots = nv_get_empty_tx_slots(np);

@@ -1687,15 +1687,16 @@ static int nv_start_xmit(struct sk_buff 
 	/* setup the fragments */
 	for (i = 0; i < fragments; i++) {
 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
-		u32 size = frag->size;
+		u32 size = frag->length;
 		offset = 0;
 
 		do {
 			prev_tx = put_tx;
 			prev_tx_ctx = np->put_tx_ctx;
 			bcnt = (size > NV_TX2_TSO_MAX_SIZE) ? NV_TX2_TSO_MAX_SIZE : size;
-			np->put_tx_ctx->dma = pci_map_page(np->pci_dev, frag->page, frag->page_offset+offset, bcnt,
-							   PCI_DMA_TODEVICE);
+			np->put_tx_ctx->dma = pci_map_page(np->pci_dev, frag->page,
+							   frag->offset + offset,
+							   bcnt, PCI_DMA_TODEVICE);
 			np->put_tx_ctx->dma_len = bcnt;
 			put_tx->buf = cpu_to_le32(np->put_tx_ctx->dma);
 			put_tx->flaglen = cpu_to_le32((bcnt-1) | tx_flags);

@@ -1765,8 +1766,8 @@ static int nv_start_xmit_optimized(struc
 
 	/* add fragments to entries count */
 	for (i = 0; i < fragments; i++) {
-		entries += (skb_shinfo(skb)->frags[i].size >> NV_TX2_TSO_MAX_SHIFT) +
-			   ((skb_shinfo(skb)->frags[i].size & (NV_TX2_TSO_MAX_SIZE-1)) ? 1 : 0);
+		entries += (skb_shinfo(skb)->frags[i].length >> NV_TX2_TSO_MAX_SHIFT) +
+			   ((skb_shinfo(skb)->frags[i].length & (NV_TX2_TSO_MAX_SIZE-1)) ? 1 : 0);
 	}
 
 	empty_slots = nv_get_empty_tx_slots(np);

@@ -1804,15 +1805,16 @@ static int nv_start_xmit_optimized(struc
 	/* setup the fragments */
 	for (i = 0; i < fragments; i++) {
 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
-		u32 size = frag->size;
+		u32 size = frag->length;
 		offset = 0;
 
 		do {
 			prev_tx = put_tx;
 			prev_tx_ctx = np->put_tx_ctx;
 			bcnt = (size > NV_TX2_TSO_MAX_SIZE) ? NV_TX2_TSO_MAX_SIZE : size;
-			np->put_tx_ctx->dma = pci_map_page(np->pci_dev, frag->page, frag->page_offset+offset, bcnt,
-							   PCI_DMA_TODEVICE);
+			np->put_tx_ctx->dma = pci_map_page(np->pci_dev, frag->page,
+							   frag->offset + offset,
+							   bcnt, PCI_DMA_TODEVICE);
 			np->put_tx_ctx->dma_len = bcnt;
 			put_tx->bufhigh = cpu_to_le64(np->put_tx_ctx->dma) >> 32;
 			put_tx->buflow = cpu_to_le64(np->put_tx_ctx->dma) & 0x0FFFFFFFF;
drivers/net/ibm_emac/ibm_emac_core.c +1 -1
@@ -1165,7 +1165,7 @@ static int emac_start_xmit_sg(struct sk_
 		if (unlikely(dev->tx_cnt + mal_tx_chunks(len) >= NUM_TX_BUFF))
 			goto undo_frame;
 
-		pd = dma_map_page(dev->ldev, frag->page, frag->page_offset, len,
+		pd = dma_map_page(dev->ldev, frag->page, frag->offset, len,
 				  DMA_TO_DEVICE);
 
 		slot = emac_xmit_split(dev, slot, pd, len, i == nr_frags - 1,
drivers/net/ibmveth.c +1 -1
@@ -698,7 +698,7 @@ static int ibmveth_start_xmit(struct sk_
 		skb_frag_t *frag = &skb_shinfo(skb)->frags[curfrag];
 		desc[curfrag+1].fields.address
 			= dma_map_single(&adapter->vdev->dev,
-				page_address(frag->page) + frag->page_offset,
+				page_address(frag->page) + frag->offset,
 				frag->size, DMA_TO_DEVICE);
 		desc[curfrag+1].fields.length = frag->size;
 		desc[curfrag+1].fields.valid  = 1;
drivers/net/ixgb/ixgb_main.c +3 -3
@@ -1314,10 +1314,10 @@ ixgb_tx_map(struct ixgb_adapter *adapter
 	}
 
 	for(f = 0; f < nr_frags; f++) {
-		struct skb_frag_struct *frag;
+		skb_frag_t *frag;
 
 		frag = &skb_shinfo(skb)->frags[f];
-		len = frag->size;
+		len = frag->length;
 		offset = 0;
 
 		while(len) {

@@ -1334,7 +1334,7 @@ ixgb_tx_map(struct ixgb_adapter *adapter
 			buffer_info->dma =
 				pci_map_page(adapter->pdev,
 					frag->page,
-					frag->page_offset + offset,
+					frag->offset + offset,
 					size,
 					PCI_DMA_TODEVICE);
 			buffer_info->time_stamp = jiffies;
drivers/net/mv643xx_eth.c +1 -1
@@ -1114,7 +1114,7 @@ static void eth_tx_fill_frag_descs(struc
 		desc->l4i_chk = 0;
 		desc->byte_cnt = this_frag->size;
 		desc->buf_ptr = dma_map_page(NULL, this_frag->page,
-						this_frag->page_offset,
+						this_frag->offset,
 						this_frag->size,
 						DMA_TO_DEVICE);
 	}
drivers/net/myri10ge/myri10ge.c +13 -13
@@ -877,9 +877,9 @@ static inline void myri10ge_vlan_ip_csum
 
 static inline void
 myri10ge_rx_skb_build(struct sk_buff *skb, u8 * va,
-		      struct skb_frag_struct *rx_frags, int len, int hlen)
+		      skb_frag_t *rx_frags, int len, int hlen)
 {
-	struct skb_frag_struct *skb_frags;
+	skb_frag_t *skb_frags;
 
 	skb->len = skb->data_len = len;
 	skb->truesize = len + sizeof(struct sk_buff);

@@ -888,7 +888,7 @@ myri10ge_rx_skb_build(struct sk_buff *sk
 	skb_frags = skb_shinfo(skb)->frags;
 	while (len > 0) {
 		memcpy(skb_frags, rx_frags, sizeof(*skb_frags));
-		len -= rx_frags->size;
+		len -= rx_frags->length;
 		skb_frags++;
 		rx_frags++;
 		skb_shinfo(skb)->nr_frags++;

@@ -899,8 +899,8 @@ myri10ge_rx_skb_build(struct sk_buff *sk
 	 * the beginning of the packet in skb_headlen(), move it
 	 * manually */
 	skb_copy_to_linear_data(skb, va, hlen);
-	skb_shinfo(skb)->frags[0].page_offset += hlen;
-	skb_shinfo(skb)->frags[0].size -= hlen;
+	skb_shinfo(skb)->frags[0].offset += hlen;
+	skb_shinfo(skb)->frags[0].length -= hlen;
 	skb->data_len -= hlen;
 	skb->tail += hlen;
 	skb_pull(skb, MXGEFW_PAD);

@@ -994,7 +994,7 @@ myri10ge_rx_done(struct myri10ge_priv *m
 		 int bytes, int len, __wsum csum)
 {
 	struct sk_buff *skb;
-	struct skb_frag_struct rx_frags[MYRI10GE_MAX_FRAGS_PER_FRAME];
+	skb_frag_t rx_frags[MYRI10GE_MAX_FRAGS_PER_FRAME];
 	int i, idx, hlen, remainder;
 	struct pci_dev *pdev = mgp->pdev;
 	struct net_device *dev = mgp->dev;

@@ -1008,11 +1008,11 @@ myri10ge_rx_done(struct myri10ge_priv *m
 	for (i = 0, remainder = len; remainder > 0; i++) {
 		myri10ge_unmap_rx_page(pdev, &rx->info[idx], bytes);
 		rx_frags[i].page = rx->info[idx].page;
-		rx_frags[i].page_offset = rx->info[idx].page_offset;
+		rx_frags[i].offset = rx->info[idx].page_offset;
 		if (remainder < MYRI10GE_ALLOC_SIZE)
-			rx_frags[i].size = remainder;
+			rx_frags[i].length = remainder;
 		else
-			rx_frags[i].size = MYRI10GE_ALLOC_SIZE;
+			rx_frags[i].length = MYRI10GE_ALLOC_SIZE;
 		rx->cnt++;
 		idx = rx->cnt & rx->mask;
 		remainder -= MYRI10GE_ALLOC_SIZE;

@@ -1034,7 +1034,7 @@ myri10ge_rx_done(struct myri10ge_priv *m
 
 	/* Attach the pages to the skb, and trim off any padding */
 	myri10ge_rx_skb_build(skb, va, rx_frags, len, hlen);
-	if (skb_shinfo(skb)->frags[0].size <= 0) {
+	if (skb_shinfo(skb)->frags[0].length <= 0) {
 		put_page(skb_shinfo(skb)->frags[0].page);
 		skb_shinfo(skb)->nr_frags = 0;
 	}

@@ -2026,7 +2026,7 @@ static int myri10ge_xmit(struct sk_buff 
 	struct myri10ge_priv *mgp = netdev_priv(dev);
 	struct mcp_kreq_ether_send *req;
 	struct myri10ge_tx_buf *tx = &mgp->tx;
-	struct skb_frag_struct *frag;
+	skb_frag_t *frag;
 	dma_addr_t bus;
 	u32 low;
 	__be32 high_swapped;

@@ -2214,8 +2214,8 @@ again:
 		idx = (count + tx->req) & tx->mask;
 		frag = &skb_shinfo(skb)->frags[frag_idx];
 		frag_idx++;
-		len = frag->size;
-		bus = pci_map_page(mgp->pdev, frag->page, frag->page_offset,
+		len = frag->length;
+		bus = pci_map_page(mgp->pdev, frag->page, frag->offset,
 				   len, PCI_DMA_TODEVICE);
 		pci_unmap_addr_set(&tx->info[idx], bus, bus);
 		pci_unmap_len_set(&tx->info[idx], len, len);
drivers/net/netxen/netxen_nic_main.c +3 -3
@@ -996,7 +996,7 @@ static int netxen_nic_xmit_frame(struct 
 	hwdesc->addr_buffer1 = cpu_to_le64(buffrag->dma);
 
 	for (i = 1, k = 1; i < frag_count; i++, k++) {
-		struct skb_frag_struct *frag;
+		skb_frag_t *frag;
 		int len, temp_len;
 		unsigned long offset;
 		dma_addr_t temp_dma;

@@ -1010,8 +1010,8 @@ static int netxen_nic_xmit_frame(struct 
 			memset(hwdesc, 0, sizeof(struct cmd_desc_type0));
 		}
 		frag = &skb_shinfo(skb)->frags[i - 1];
-		len = frag->size;
-		offset = frag->page_offset;
+		len = frag->length;
+		offset = frag->offset;
 
 		temp_len = len;
 		temp_dma = pci_map_page(adapter->pdev, frag->page, offset,
drivers/net/ns83820.c +4 -5
@@ -1187,13 +1187,12 @@ again:
 		if (!nr_frags)
 			break;
 
-		buf = pci_map_page(dev->pci_dev, frag->page,
-				   frag->page_offset,
-				   frag->size, PCI_DMA_TODEVICE);
+		buf = pci_map_page(dev->pci_dev, frag->page, frag->offset,
+				   frag->length, PCI_DMA_TODEVICE);
 		dprintk("frag: buf=%08Lx  page=%08lx offset=%08lx\n",
 			(long long)buf, (long) page_to_pfn(frag->page),
-			frag->page_offset);
-		len = frag->size;
+			frag->offset);
+		len = frag->length;
 		frag++;
 		nr_frags--;
 	}
drivers/net/qla3xxx.c +3 -3
@@ -2548,7 +2548,7 @@ static int ql_send_map(struct ql3_adapte
 
 			map =
 			    pci_map_page(qdev->pdev, frag->page,
-					 frag->page_offset, frag->size,
+					 frag->offset, frag->length,
 					 PCI_DMA_TODEVICE);
 
 			err = pci_dma_mapping_error(map);

@@ -2560,10 +2560,10 @@ static int ql_send_map(struct ql3_adapte
 
 			oal_entry->dma_lo = cpu_to_le32(LS_64BITS(map));
 			oal_entry->dma_hi = cpu_to_le32(MS_64BITS(map));
-			oal_entry->len = cpu_to_le32(frag->size);
+			oal_entry->len = cpu_to_le32(frag->length);
 			pci_unmap_addr_set(&tx_cb->map[seg], mapaddr, map);
 			pci_unmap_len_set(&tx_cb->map[seg], maplen,
-					  frag->size);
+					  frag->length);
 		}
 		/* Terminate the last segment. */
 		oal_entry->len =
drivers/net/r8169.c +2 -2
@@ -2243,8 +2243,8 @@ static int rtl8169_xmit_frags(struct rtl
 		entry = (entry + 1) % NUM_TX_DESC;
 
 		txd = tp->TxDescArray + entry;
-		len = frag->size;
-		addr = ((void *) page_address(frag->page)) + frag->page_offset;
+		len = frag->length;
+		addr = ((void *) page_address(frag->page)) + frag->offset;
 		mapping = pci_map_single(tp->pci_dev, addr, len, PCI_DMA_TODEVICE);
 
 		/* anti gcc 2.95.3 bugware (sic) */
drivers/net/s2io.c +10 -8
@@ -2141,9 +2141,9 @@ static struct sk_buff *s2io_txdl_getskb(
 			skb_frag_t *frag = &skb_shinfo(skb)->frags[j];
 			if (!txds->Buffer_Pointer)
 				break;
-			pci_unmap_page(nic->pdev, (dma_addr_t)
-					txds->Buffer_Pointer,
-				       frag->size, PCI_DMA_TODEVICE);
+			pci_unmap_page(nic->pdev,
+				       (dma_addr_t) txds->Buffer_Pointer,
+				       frag->length, PCI_DMA_TODEVICE);
 		}
 	}
 	memset(txdlp,0, (sizeof(struct TxD) * fifo_data->max_txds));

@@ -4087,13 +4087,15 @@ static int s2io_xmit(struct sk_buff *skb
 	for (i = 0; i < frg_cnt; i++) {
 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 		/* A '0' length fragment will be ignored */
-		if (!frag->size)
+		if (!frag->length)
 			continue;
 		txdp++;
-		txdp->Buffer_Pointer = (u64) pci_map_page
-		    (sp->pdev, frag->page, frag->page_offset,
-		     frag->size, PCI_DMA_TODEVICE);
-		txdp->Control_1 = TXD_BUFFER0_SIZE(frag->size);
+		txdp->Buffer_Pointer = (u64) pci_map_page(sp->pdev,
+							  frag->page,
+							  frag->offset,
+							  frag->length,
+							  PCI_DMA_TODEVICE);
+		txdp->Control_1 = TXD_BUFFER0_SIZE(frag->length);
 		if (offload_type == SKB_GSO_UDP)
 			txdp->Control_1 |= TXD_UFO_EN;
 	}
drivers/net/sk98lin/skge.c +4 -4
@@ -1721,15 +1721,15 @@ struct sk_buff	*pMessage)	/* pointer to 
 		*/
 		PhysAddr = (SK_U64) pci_map_page(pAC->PciDev,
 						 sk_frag->page,
-						 sk_frag->page_offset,
-						 sk_frag->size,
+						 sk_frag->offset,
+						 sk_frag->length,
 						 PCI_DMA_TODEVICE);
 
 		pTxd->VDataLow  = (SK_U32) (PhysAddr & 0xffffffff);
 		pTxd->VDataHigh = (SK_U32) (PhysAddr >> 32);
 		pTxd->pMBuf     = pMessage;
 		
-		pTxd->TBControl = Control | BMU_OWN | sk_frag->size;
+		pTxd->TBControl = Control | BMU_OWN | sk_frag->length;
 
 		/* 
 		** Do we have the last fragment? 

@@ -1745,7 +1745,7 @@ struct sk_buff	*pMessage)	/* pointer to 
 		pTxdLst = pTxd;
 		pTxd    = pTxd->pNextTxd;
 		pTxPort->TxdRingFree--;
-		BytesSend += sk_frag->size;
+		BytesSend += sk_frag->length;
 	}
 
 	/*
drivers/net/skge.c +4 -4
@@ -2684,8 +2684,8 @@ static int skge_xmit_frame(struct sk_buf
 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 
-			map = pci_map_page(hw->pdev, frag->page, frag->page_offset,
-					   frag->size, PCI_DMA_TODEVICE);
+			map = pci_map_page(hw->pdev, frag->page, frag->offset,
+					   frag->length, PCI_DMA_TODEVICE);
 
 			e = e->next;
 			e->skb = skb;

@@ -2695,9 +2695,9 @@ static int skge_xmit_frame(struct sk_buf
 			tf->dma_lo = map;
 			tf->dma_hi = (u64) map >> 32;
 			pci_unmap_addr_set(e, mapaddr, map);
-			pci_unmap_len_set(e, maplen, frag->size);
+			pci_unmap_len_set(e, maplen, frag->length);
 
-			tf->control = BMU_OWN | BMU_SW | control | frag->size;
+			tf->control = BMU_OWN | BMU_SW | control | frag->length;
 		}
 		tf->control |= BMU_EOF | BMU_IRQ_EOF;
 	}
drivers/net/sky2.c +8 -8
@@ -912,8 +912,8 @@ static void sky2_rx_map_skb(struct pci_d
 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
 		re->frag_addr[i] = pci_map_page(pdev,
 						skb_shinfo(skb)->frags[i].page,
-						skb_shinfo(skb)->frags[i].page_offset,
-						skb_shinfo(skb)->frags[i].size,
+						skb_shinfo(skb)->frags[i].offset,
+						skb_shinfo(skb)->frags[i].length,
 						PCI_DMA_FROMDEVICE);
 }
 

@@ -927,7 +927,7 @@ static void sky2_rx_unmap_skb(struct pci
 
 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
 		pci_unmap_page(pdev, re->frag_addr[i],
-			       skb_shinfo(skb)->frags[i].size,
+			       skb_shinfo(skb)->frags[i].length,
 			       PCI_DMA_FROMDEVICE);
 }
 

@@ -1457,8 +1457,8 @@ static int sky2_xmit_frame(struct sk_buf
 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 
-		mapping = pci_map_page(hw->pdev, frag->page, frag->page_offset,
-				       frag->size, PCI_DMA_TODEVICE);
+		mapping = pci_map_page(hw->pdev, frag->page, frag->offset,
+				       frag->length, PCI_DMA_TODEVICE);
 		addr64 = high32(mapping);
 		if (addr64 != sky2->tx_addr64) {
 			le = get_tx_le(sky2);

@@ -1470,14 +1470,14 @@ static int sky2_xmit_frame(struct sk_buf
 
 		le = get_tx_le(sky2);
 		le->addr = cpu_to_le32((u32) mapping);
-		le->length = cpu_to_le16(frag->size);
+		le->length = cpu_to_le16(frag->length);
 		le->ctrl = ctrl;
 		le->opcode = OP_BUFFER | HW_OWNER;
 
 		re = tx_le_re(sky2, le);
 		re->skb = skb;
 		pci_unmap_addr_set(re, mapaddr, mapping);
-		pci_unmap_len_set(re, maplen, frag->size);
+		pci_unmap_len_set(re, maplen, frag->length);
 	}
 
 	le->ctrl |= EOP;

@@ -2002,7 +2002,7 @@ static void skb_put_frags(struct sk_buff
 		} else {
 			size = min(length, (unsigned) PAGE_SIZE);
 
-			frag->size = size;
+			frag->length = size;
 			skb->data_len += size;
 			skb->truesize += size;
 			skb->len += size;
drivers/net/starfire.c +6 -3
@@ -1262,9 +1262,12 @@ static int start_tx(struct sk_buff *skb,
 				pci_map_single(np->pci_dev, skb->data, skb_first_frag_len(skb), PCI_DMA_TODEVICE);
 		} else {
 			skb_frag_t *this_frag = &skb_shinfo(skb)->frags[i - 1];
-			status |= this_frag->size;
+			status |= this_frag->length;
 			np->tx_info[entry].mapping =
-				pci_map_single(np->pci_dev, page_address(this_frag->page) + this_frag->page_offset, this_frag->size, PCI_DMA_TODEVICE);
+				pci_map_single(np->pci_dev,
+					       page_address(this_frag->page) + this_frag->offset,
+					       this_frag->length,
+					       PCI_DMA_TODEVICE);
 		}
 
 		np->tx_ring[entry].addr = cpu_to_dma(np->tx_info[entry].mapping);

@@ -1362,7 +1365,7 @@ static irqreturn_t intr_handler(int irq,
 					for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 						pci_unmap_single(np->pci_dev,
 								 np->tx_info[entry].mapping,
-								 skb_shinfo(skb)->frags[i].size,
+								 skb_shinfo(skb)->frags[i].length,
 								 PCI_DMA_TODEVICE);
 						np->dirty_tx++;
 						entry++;
drivers/net/sungem.c +2 -2
@@ -1101,10 +1101,10 @@ static int gem_start_xmit(struct sk_buff
 			dma_addr_t mapping;
 			u64 this_ctrl;
 
-			len = this_frag->size;
+			len = this_frag->length;
 			mapping = pci_map_page(gp->pdev,
 					       this_frag->page,
-					       this_frag->page_offset,
+					       this_frag->offset,
 					       len, PCI_DMA_TODEVICE);
 			this_ctrl = ctrl;
 			if (frag == skb_shinfo(skb)->nr_frags - 1)
drivers/net/sunhme.c +2 -2
@@ -2313,10 +2313,10 @@ static int happy_meal_start_xmit(struct 
 			skb_frag_t *this_frag = &skb_shinfo(skb)->frags[frag];
 			u32 len, mapping, this_txflags;
 
-			len = this_frag->size;
+			len = this_frag->length;
 			mapping = hme_dma_map(hp,
 					      ((void *) page_address(this_frag->page) +
-					       this_frag->page_offset),
+					       this_frag->offset),
 					      len, DMA_TODEVICE);
 			this_txflags = tx_flags;
 			if (frag == skb_shinfo(skb)->nr_frags - 1)
drivers/net/tg3.c +7 -7
@@ -3095,7 +3095,7 @@ static void tg3_tx(struct tg3 *tp)
 
 			pci_unmap_page(tp->pdev,
 				       pci_unmap_addr(ri, mapping),
-				       skb_shinfo(skb)->frags[i].size,
+				       skb_shinfo(skb)->frags[i].length,
 				       PCI_DMA_TODEVICE);
 
 			sw_idx = NEXT_TX(sw_idx);

@@ -3835,7 +3835,7 @@ static int tigon3_dma_hwbug_workaround(s
 		if (i == 0)
 			len = skb_headlen(skb);
 		else
-			len = skb_shinfo(skb)->frags[i-1].size;
+			len = skb_shinfo(skb)->frags[i-1].length;
 		pci_unmap_single(tp->pdev,
 				 pci_unmap_addr(&tp->tx_buffers[entry], mapping),
 				 len, PCI_DMA_TODEVICE);

@@ -3962,10 +3962,10 @@ static int tg3_start_xmit(struct sk_buff
 		for (i = 0; i <= last; i++) {
 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 
-			len = frag->size;
+			len = frag->length;
 			mapping = pci_map_page(tp->pdev,
 					       frag->page,
-					       frag->page_offset,
+					       frag->offset,
 					       len, PCI_DMA_TODEVICE);
 
 			tp->tx_buffers[entry].skb = NULL;

@@ -4144,10 +4144,10 @@ static int tg3_start_xmit_dma_bug(struct
 		for (i = 0; i <= last; i++) {
 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 
-			len = frag->size;
+			len = frag->length;
 			mapping = pci_map_page(tp->pdev,
 					       frag->page,
-					       frag->page_offset,
+					       frag->offset,
 					       len, PCI_DMA_TODEVICE);
 
 			tp->tx_buffers[entry].skb = NULL;

@@ -4321,7 +4321,7 @@ static void tg3_free_rings(struct tg3 *t
 			txp = &tp->tx_buffers[i & (TG3_TX_RING_SIZE - 1)];
 			pci_unmap_page(tp->pdev,
 				       pci_unmap_addr(txp, mapping),
-				       skb_shinfo(skb)->frags[j].size,
+				       skb_shinfo(skb)->frags[j].length,
 				       PCI_DMA_TODEVICE);
 			i++;
 		}
drivers/net/tsi108_eth.c +1 -1
@@ -715,7 +715,7 @@ static int tsi108_send_packet(struct sk_
 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];
 
 			data->txring[tx].buf0 =
-			    dma_map_page(NULL, frag->page, frag->page_offset,
+			    dma_map_page(NULL, frag->page, frag->offset,
 					    frag->size, DMA_TO_DEVICE);
 			data->txring[tx].len = frag->size;
 		}
drivers/net/typhoon.c +1 -1
@@ -874,7 +874,7 @@ typhoon_start_tx(struct sk_buff *skb, st
 
 			len = frag->size;
 			frag_addr = (void *) page_address(frag->page) +
-						frag->page_offset;
+						frag->offset;
 			skb_dma = pci_map_single(tp->tx_pdev, frag_addr, len,
 					 PCI_DMA_TODEVICE);
 			txd->flags = TYPHOON_FRAG_DESC | TYPHOON_DESC_VALID;
drivers/net/via-velocity.c +1 -1
@@ -1966,7 +1966,7 @@ static int velocity_xmit(struct sk_buff 
 
 			for (i = 0; i < nfrags; i++) {
 				skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
-				void *addr = ((void *) page_address(frag->page + frag->page_offset));
+				void *addr = ((void *) page_address(frag->page + frag->offset));
 
 				tdinfo->skb_dma[i + 1] = pci_map_single(vptr->pdev, addr, frag->size, PCI_DMA_TODEVICE);
net/appletalk/ddp.c +2 -2
@@ -957,7 +957,7 @@ static unsigned long atalk_sum_skb(const
 
 		BUG_TRAP(start <= offset + len);
 
-		end = start + skb_shinfo(skb)->frags[i].size;
+		end = start + skb_shinfo(skb)->frags[i].length;
 		if ((copy = end - offset) > 0) {
 			u8 *vaddr;
 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];

@@ -965,7 +965,7 @@ static unsigned long atalk_sum_skb(const
 			if (copy > len)
 				copy = len;
 			vaddr = kmap_skb_frag(frag);
-			sum = atalk_sum_partial(vaddr + frag->page_offset +
+			sum = atalk_sum_partial(vaddr + frag->offset +
 						  offset - start, copy, sum);
 			kunmap_skb_frag(vaddr);
net/core/datagram.c +4 -5
@@ -267,7 +267,7 @@ int skb_copy_datagram_iovec(const struct
 
 		BUG_TRAP(start <= offset + len);
 
-		end = start + skb_shinfo(skb)->frags[i].size;
+		end = start + skb_shinfo(skb)->frags[i].length;
 		if ((copy = end - offset) > 0) {
 			int err;
 			u8  *vaddr;

@@ -277,7 +277,7 @@ int skb_copy_datagram_iovec(const struct
 			if (copy > len)
 				copy = len;
 			vaddr = kmap(page);
-			err = memcpy_toiovec(to, vaddr + frag->page_offset +
+			err = memcpy_toiovec(to, vaddr + frag->offset +
 					     offset - start, copy);
 			kunmap(page);
 			if (err)

@@ -348,7 +348,7 @@ static int skb_copy_and_csum_datagram(co
 
 		BUG_TRAP(start <= offset + len);
 
-		end = start + skb_shinfo(skb)->frags[i].size;
+		end = start + skb_shinfo(skb)->frags[i].length;
 		if ((copy = end - offset) > 0) {
 			__wsum csum2;
 			int err = 0;

@@ -359,8 +359,7 @@ static int skb_copy_and_csum_datagram(co
 			if (copy > len)
 				copy = len;
 			vaddr = kmap(page);
-			csum2 = csum_and_copy_to_user(vaddr +
-							frag->page_offset +
+			csum2 = csum_and_copy_to_user(vaddr + frag->offset +
 							offset - start,
 						      to, copy, 0, &err);
 			kunmap(page);
net/core/skbuff.c +32 -32

211 lines not on this page · raw message

net/core/sock.c +4 -4
@@ -1215,10 +1215,10 @@ static struct sk_buff *sock_alloc_send_p
 
 					frag = &skb_shinfo(skb)->frags[i];
 					frag->page = page;
-					frag->page_offset = 0;
-					frag->size = (data_len >= PAGE_SIZE ?
-						      PAGE_SIZE :
-						      data_len);
+					frag->offset = 0;
+					frag->length = (data_len >= PAGE_SIZE ?
+							PAGE_SIZE :
+							data_len);
 					data_len -= PAGE_SIZE;
 				}
net/core/user_dma.c +1 -1
@@ -83,7 +83,7 @@ int dma_skb_copy_datagram_iovec(struct d
 				copy = len;
 
 			cookie = dma_memcpy_pg_to_iovec(chan, to, pinned_list, page,
-					frag->page_offset + offset - start, copy);
+					frag->offset + offset - start, copy);
 			if (cookie < 0)
 				goto fault;
 			len -= copy;
drivers/net/cxgb3/adapter.h +1 -1
@@ -75,7 +75,7 @@ struct rx_desc;
 struct rx_sw_desc;
 
 struct sge_fl_page {
-	struct skb_frag_struct frag;
+	skb_frag_t frag;
 	unsigned char *va;
 };
net/core/pktgen.c +21 -21

92 lines not on this page · raw message

net/ipv4/ip_fragment.c +2 -2
@@ -647,8 +647,8 @@ static struct sk_buff *ip_frag_reasm(str
 		head->next = clone;
 		skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
 		skb_shinfo(head)->frag_list = NULL;
-		for (i=0; i<skb_shinfo(head)->nr_frags; i++)
-			plen += skb_shinfo(head)->frags[i].size;
+		for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
+			plen += skb_shinfo(head)->frags[i].length;
 		clone->len = clone->data_len = head->data_len - plen;
 		head->data_len -= clone->len;
 		head->len -= clone->len;
net/ipv4/ip_output.c +6 -3
@@ -1019,12 +1019,15 @@ alloc_new_skb:
 				err = -EMSGSIZE;
 				goto error;
 			}
-			if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
+			if (getfrag(from,
+				    page_address(frag->page)
+				      + frag->offset + frag->length,
+				    offset, copy, skb->len, skb) < 0) {
 				err = -EFAULT;
 				goto error;
 			}
 			sk->sk_sndmsg_off += copy;
-			frag->size += copy;
+			frag->length += copy;
 			skb->len += copy;
 			skb->data_len += copy;
 		}

@@ -1152,7 +1155,7 @@ ssize_t	ip_append_page(struct sock *sk, 
 		if (len > size)
 			len = size;
 		if (skb_can_coalesce(skb, i, page, offset)) {
-			skb_shinfo(skb)->frags[i-1].size += len;
+			skb_shinfo(skb)->frags[i-1].length += len;
 		} else if (i < MAX_SKB_FRAGS) {
 			get_page(page);
 			skb_fill_page_desc(skb, i, page, offset, len);
net/ipv4/tcp.c +4 -5
@@ -558,7 +558,7 @@ new_segment:
 			goto wait_for_memory;
 
 		if (can_coalesce) {
-			skb_shinfo(skb)->frags[i - 1].size += copy;
+			skb_shinfo(skb)->frags[i - 1].length += copy;
 		} else {
 			get_page(page);
 			skb_fill_page_desc(skb, i, page, offset, copy);

@@ -799,10 +799,9 @@ new_segment:
 				}
 
 				/* Update the skb. */
-				if (merge) {
-					skb_shinfo(skb)->frags[i - 1].size +=
-									copy;
-				} else {
+				if (merge)
+					skb_shinfo(skb)->frags[i - 1].length += copy;
+				else {
 					skb_fill_page_desc(skb, i, page, off, copy);
 					if (TCP_PAGE(sk)) {
 						get_page(page);
net/ipv4/tcp_output.c +4 -4
@@ -720,14 +720,14 @@ static void __pskb_trim_head(struct sk_b
 	eat = len;
 	k = 0;
 	for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
-		if (skb_shinfo(skb)->frags[i].size <= eat) {
+		if (skb_shinfo(skb)->frags[i].length <= eat) {
 			put_page(skb_shinfo(skb)->frags[i].page);
-			eat -= skb_shinfo(skb)->frags[i].size;
+			eat -= skb_shinfo(skb)->frags[i].length;
 		} else {
 			skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
 			if (eat) {
-				skb_shinfo(skb)->frags[k].page_offset += eat;
-				skb_shinfo(skb)->frags[k].size -= eat;
+				skb_shinfo(skb)->frags[k].offset += eat;
+				skb_shinfo(skb)->frags[k].length -= eat;
 				eat = 0;
 			}
 			k++;
net/ipv6/ip6_output.c +5 -2
@@ -1314,12 +1314,15 @@ alloc_new_skb:
 				err = -EMSGSIZE;
 				goto error;
 			}
-			if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
+			if (getfrag(from,
+				    page_address(frag->page)
+				      + frag->offset + frag->length,
+				    offset, copy, skb->len, skb) < 0) {
 				err = -EFAULT;
 				goto error;
 			}
 			sk->sk_sndmsg_off += copy;
-			frag->size += copy;
+			frag->length += copy;
 			skb->len += copy;
 			skb->data_len += copy;
 		}
net/ipv6/netfilter/nf_conntrack_reasm.c +1 -1
@@ -612,7 +612,7 @@ nf_ct_frag6_reasm(struct nf_ct_frag6_que
 		skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
 		skb_shinfo(head)->frag_list = NULL;
 		for (i=0; i<skb_shinfo(head)->nr_frags; i++)
-			plen += skb_shinfo(head)->frags[i].size;
+			plen += skb_shinfo(head)->frags[i].length;
 		clone->len = clone->data_len = head->data_len - plen;
 		head->data_len -= clone->len;
 		head->len -= clone->len;
net/ipv6/reassembly.c +1 -1
@@ -634,7 +634,7 @@ static int ip6_frag_reasm(struct frag_qu
 		skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
 		skb_shinfo(head)->frag_list = NULL;
 		for (i=0; i<skb_shinfo(head)->nr_frags; i++)
-			plen += skb_shinfo(head)->frags[i].size;
+			plen += skb_shinfo(head)->frags[i].length;
 		clone->len = clone->data_len = head->data_len - plen;
 		head->data_len -= clone->len;
 		head->len -= clone->len;
net/xfrm/xfrm_algo.c +2 -2
@@ -570,7 +570,7 @@ int skb_icv_walk(const struct sk_buff *s
 
 		BUG_TRAP(start <= offset + len);
 
-		end = start + skb_shinfo(skb)->frags[i].size;
+		end = start + skb_shinfo(skb)->frags[i].length;
 		if ((copy = end - offset) > 0) {
 			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 

@@ -578,7 +578,7 @@ int skb_icv_walk(const struct sk_buff *s
 				copy = len;
 
 			sg.page = frag->page;
-			sg.offset = frag->page_offset + offset-start;
+			sg.offset = frag->offset + offset-start;
 			sg.length = copy;
 
 			err = icv_update(desc, &sg, copy);
Signature · 1 line
Stephen Hemminger <[email protected]>
author
#2 lore raw

RFC 2/2 shrink size of scatterlist on common i386/x86-64

In reply to Stephen Hemminger

The scatterlist only needs 16 bits for length/offset because PAGE_SIZE is 4K

Signed-off-by Stephen Hemminger <[email protected]>
include/asm-i386/scatterlist.h +2 -2
@@ -5,9 +5,9 @@
 
 struct scatterlist {
     struct page		*page;
-    unsigned int	offset;
     dma_addr_t		dma_address;
-    unsigned int	length;
+    u16			offset;
+    u16			length;
 };
 
 /* These macros should be used after a pci_map_sg call has been done
include/asm-x86_64/scatterlist.h +2 -2
@@ -5,10 +5,10 @@
 
 struct scatterlist {
     struct page		*page;
-    unsigned int	offset;
-    unsigned int	length;
     dma_addr_t		dma_address;
     unsigned int        dma_length;
+    u16			offset;
+    u16			length;
 };
 
 #define ISA_DMA_THRESHOLD (0x00ffffff)
Signature · 1 line
Stephen Hemminger <[email protected]>
#3 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

include/asm-i386/scatterlist.h line 11

Quoted 12 lines
include/asm-i386/scatterlist.h +2 -2
@@ -5,9 +5,9 @@
 
 struct scatterlist {
     struct page		*page;
-    unsigned int	offset;
     dma_addr_t		dma_address;
-    unsigned int	length;
+    u16			offset;
+    u16			length;
 };

Actually this struct layout could be even better, since pointers are 32 bits but dma_addr_t may be 64 bits... having

struct scatterlist {
	dma_addr_t	dma_address;
	struct page    *page;
	u16		offset;
	u16		length;
};

would allow struct scatterlist to be 16 bytes. Seems like a good thing...

- R.

maintainer
#4 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

In reply to Stephen Hemminger

From: Stephen Hemminger <[email protected]>
Date: Thu, 05 Jul 2007 16:14:14 -0700

Quoted 4 lines

The scatterlist only needs 16 bits for length/offset because PAGE_SIZE is 4K

Signed-off-by Stephen Hemminger <[email protected]>

Unfortunately I don't think this can be done, even on i386.

It is legal to use order!=0 pages and multi-page areas in a single scatterlist entry.

author
#5 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

Quoted 6 lines · On Thu, 05 Jul 2007 16:43:08 -0700 (PDT) David Miller <[email protected]>…

From: Stephen Hemminger <[email protected]>
Date: Thu, 05 Jul 2007 16:14:14 -0700

Quoted 3 lines

The scatterlist only needs 16 bits for length/offset because PAGE_SIZE is 4K

Quoted 7 lines
Quoted 1 line
Signed-off-by Stephen Hemminger <[email protected]>

Unfortunately I don't think this can be done, even on i386.

It is legal to use order!=0 pages and multi-page areas in a single scatterlist entry.

Okay, but then using SG lists makes skbuff's much bigger.

fraglist scatterlist per skbuff
32 bit 8 20 +12 * 18 = +216!
64 bit 16 32 +16 * 18 = +288

So never mind...

I'll do a fraglist to scatter list set of routines, but not sure if it's worth it.

Signature · 1 line
Stephen Hemminger <[email protected]>
maintainer
#6 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

From: Stephen Hemminger <[email protected]>
Date: Thu, 5 Jul 2007 17:00:51 -0700

Quoted 17 lines
Quoted 6 lines · On Thu, 05 Jul 2007 16:43:08 -0700 (PDT) David Miller <[email protected]>…

From: Stephen Hemminger <[email protected]>
Date: Thu, 05 Jul 2007 16:14:14 -0700

Quoted 3 lines

The scatterlist only needs 16 bits for length/offset because PAGE_SIZE is 4K

Quoted 7 lines
Quoted 1 line
Signed-off-by Stephen Hemminger <[email protected]>

Unfortunately I don't think this can be done, even on i386.

It is legal to use order!=0 pages and multi-page areas in a single scatterlist entry.

Quoted 7 lines

Okay, but then using SG lists makes skbuff's much bigger.

fraglist scatterlist per skbuff
32 bit 8 20 +12 * 18 = +216!
64 bit 16 32 +16 * 18 = +288

So never mind...

I know, this is why nobody ever really tries to tackle this.

Quoted 2 lines

I'll do a fraglist to scatter list set of routines, but not sure if it's worth it.

It's better to add dma_map_skb() et al. interfaces to be honest.

Also even with the scatterlist idea, we'd still need to do two map calls, one for skb->data and one for the page vector.

#7 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

Quoted 2 lines · David Miller

Also even with the scatterlist idea, we'd still need to do two map calls, one for skb->data and one for the page vector.

We could make skb->shinfo(skb)->frags[0] the first segment and deprecate skb->data and skb->len.

OK, you can stop hitting me now...
Rusty.

maintainer
#8 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

From: Rusty Russell <[email protected]>
Date: Fri, 06 Jul 2007 17:43:46 +1000

Quoted 8 lines · David Miller
Quoted 2 lines

Also even with the scatterlist idea, we'd still need to do two map calls, one for skb->data and one for the page vector.

We could make skb->shinfo(skb)->frags[0] the first segment and deprecate skb->data and skb->len.

OK, you can stop hitting me now...

The fly in that ointment is that referencing the backing page of kmalloc() data is not safe currently, especially with SLUB.

Jens Axboe ran into this while trying to implement sendfile() via splice() several weeks ago, it's an ongoing battle that even C. Lameter is even involved in now :-)

#9 lore raw

RE: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

In reply to David Miller
Quoted 10 lines · David Miller
Quoted 7 lines

Okay, but then using SG lists makes skbuff's much bigger.

fraglist scatterlist per skbuff
32 bit 8 20 +12 * 18 = +216!
64 bit 16 32 +16 * 18 = +288

So never mind...

I know, this is why nobody ever really tries to tackle this.

Quoted 7 lines
Quoted 2 lines

I'll do a fraglist to scatter list set of routines, but not sure if it's worth it.

It's better to add dma_map_skb() et al. interfaces to be honest.

Also even with the scatterlist idea, we'd still need to do two map calls, one for skb->data and one for the page vector.

FWIW, I tried this about a year ago to try to improve e1000 performance on pSeries. I was hoping to simplify the driver transmit code and make IOMMU mapping easier. This was on 2.6.16 or thereabouts.

Net result: zilch. No performance increase, no noticeable CPU utilization
benefits. Nothing. So I dropped it.

Slightly off topic:
The real problem that I saw on pSeries is lock contention for the IOMMU. It's architected with a single table per slot, which is great in that two boards in separate slots won't have lock contention. However, this all goes out the window when you drop a quad-port gigabit adapter in there.
The time spent waiting for the IOMMU table lock goes up exponentially as you activate each additional port.

In my opinion, IOMMU table locking is the major issue with this type of architecture. Since both Intel and AMD are touting IOMMUs for virtual- ization support, this is an issue that's going to need a lot of scrutiny.

-Mitch

maintainer
#10 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

From: "Williams, Mitch A" <[email protected]>
Date: Fri, 6 Jul 2007 10:14:56 -0700

Quoted 4 lines

In my opinion, IOMMU table locking is the major issue with this type of architecture. Since both Intel and AMD are touting IOMMUs for virtual- ization support, this is an issue that's going to need a lot of scrutiny.

For the allocation of IOMMU entries themselves you can play tricks using atomic operations on 64-bit words of the allocator bitmap to avoid locking that.

You can use per-cpu salts to determine where to start the search and avoid hitting the same cachelines as other cpus working on the same table.

But you'll need to lock in order to flush the IOMMU tlb I'm afraid. The way to mitigate that is to only flush the IOMMU tlb once per allocator generation.

#11 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

Quoted 11 lines · David Miller

From: "Williams, Mitch A" <[email protected]>
Date: Fri, 6 Jul 2007 10:14:56 -0700

Quoted 4 lines

In my opinion, IOMMU table locking is the major issue with this type of architecture. Since both Intel and AMD are touting IOMMUs for virtual- ization support, this is an issue that's going to need a lot of scrutiny.

For the allocation of IOMMU entries themselves you can play tricks using atomic operations on 64-bit words of the allocator bitmap to avoid locking that.

Hmm, any pointers?

Quoted 7 lines

You can use per-cpu salts to determine where to start the search and avoid hitting the same cachelines as other cpus working on the same table.

But you'll need to lock in order to flush the IOMMU tlb I'm afraid. The way to mitigate that is to only flush the IOMMU tlb once per allocator generation.

That works, but isn't optimal when you have an isolation-capable IOMMU and you want the full isolation properties of the IOMMU. If you only flush the IOTLB when the allocator wraps around, a stale entry in the IOTLB can allow a DMA to go through for an IO entry that has already been unmapped. One way to mitigate that and still retain full isolation is to make sure no one else gets to use the frames that are the targets of the DMA until the translation has been flushed out of the IOTLB, but that requires pretty deep surgery.

Cheers,
Muli

maintainer
#12 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

From: Muli Ben-Yehuda <[email protected]>
Date: Sun, 8 Jul 2007 19:17:30 +0300

Quoted 9 lines · David Miller
Quoted 8 lines

From: "Williams, Mitch A" <[email protected]>
Date: Fri, 6 Jul 2007 10:14:56 -0700

Quoted 4 lines

In my opinion, IOMMU table locking is the major issue with this type of architecture. Since both Intel and AMD are touting IOMMUs for virtual- ization support, this is an issue that's going to need a lot of scrutiny.

Quoted 5 lines
Quoted 3 lines

For the allocation of IOMMU entries themselves you can play tricks using atomic operations on 64-bit words of the allocator bitmap to avoid locking that.

Hmm, any pointers?

I've never implemented it but it is certainly possible to do.

Quoted 8 lines

That works, but isn't optimal when you have an isolation-capable IOMMU and you want the full isolation properties of the IOMMU. If you only flush the IOTLB when the allocator wraps around, a stale entry in the IOTLB can allow a DMA to go through for an IO entry that has already been unmapped. One way to mitigate that and still retain full isolation is to make sure no one else gets to use the frames that are the targets of the DMA until the translation has been flushed out of the IOTLB, but that requires pretty deep surgery.

Virtualization sucks doesn't it? :-)

It's one of the worst aspects of all of this virtualization business. In my view it makes no sense to split up the physical hardware. Just give control nodes complete access to everything and instead of playing games partializing real hardware, just give virtual instances to everybody and be done with all of this complexity.

Anyways, hypervisors et al. have already decided to do this braindamage so you will need to find some way to make it go fast now won't you? :)

#13 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

Quoted 11 lines · David Miller
Quoted 9 lines

That works, but isn't optimal when you have an isolation-capable IOMMU and you want the full isolation properties of the IOMMU. If you only flush the IOTLB when the allocator wraps around, a stale entry in the IOTLB can allow a DMA to go through for an IO entry that has already been unmapped. One way to mitigate that and still retain full isolation is to make sure no one else gets to use the frames that are the targets of the DMA until the translation has been flushed out of the IOTLB, but that requires pretty deep surgery.

Virtualization sucks doesn't it? :-)

no comment :-) FWIW isolation capable IOMMUs are also useful to catch DMA errors when drivers program devices to DMA where they shouldn't. Sure, drivers can trash the kernel in other ways, but a printk beats random memory corruption any day of the week.

Quoted 6 lines

It's one of the worst aspects of all of this virtualization business. In my view it makes no sense to split up the physical hardware. Just give control nodes complete access to everything and instead of playing games partializing real hardware, just give virtual instances to everybody and be done with all of this complexity.

Virtual instances have a non-negligible performance cost and development cost - you need a virtual driver for any device or class of devices out there. Not to say that they aren't useful, just that direct access (aka "passthrough") has its uses too.

Quoted 3 lines

Anyways, hypervisors et al. have already decided to do this braindamage so you will need to find some way to make it go fast now won't you? :)

Working on it :-)

Cheers,
Muli

#14 lore raw

Re: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

In reply to Williams, Mitch A
Quoted 19 lines · Mitch A
Quoted 10 lines · David Miller
Quoted 7 lines

Okay, but then using SG lists makes skbuff's much bigger.

fraglist scatterlist per skbuff
32 bit 8 20 +12 * 18 = +216!
64 bit 16 32 +16 * 18 = +288

So never mind...

I know, this is why nobody ever really tries to tackle this.

Quoted 7 lines
Quoted 2 lines

I'll do a fraglist to scatter list set of routines, but not sure if it's worth it.

It's better to add dma_map_skb() et al. interfaces to be honest.

Also even with the scatterlist idea, we'd still need to do two map calls, one for skb->data and one for the page vector.

Quoted 8 lines

FWIW, I tried this about a year ago to try to improve e1000 performance on pSeries. I was hoping to simplify the driver transmit code and make IOMMU mapping easier. This was on 2.6.16 or thereabouts.

Net result: zilch. No performance increase, no noticeable CPU utilization
benefits. Nothing. So I dropped it.

Do you have pointers to the patches perchance?

Quoted 9 lines

Slightly off topic:
The real problem that I saw on pSeries is lock contention for the IOMMU. It's architected with a single table per slot, which is great in that two boards in separate slots won't have lock contention. However, this all goes out the window when you drop a quad-port gigabit adapter in there.
The time spent waiting for the IOMMU table lock goes up exponentially as you activate each additional port.

Quoted 4 lines

In my opinion, IOMMU table locking is the major issue with this type of architecture. Since both Intel and AMD are touting IOMMUs for virtual- ization support, this is an issue that's going to need a lot of scrutiny.

Agreed.

Cheers,
Muli

#15 lore raw

RE: [RFC 2/2] shrink size of scatterlist on common i386/x86-64

Quoted 5 lines · Muli Ben-Yehuda
Quoted 3 lines

Net result: zilch. No performance increase, no noticeable CPU utilization
benefits. Nothing. So I dropped it.

Do you have pointers to the patches perchance?

Muli, I've been looking for this code and it looks like it's gone. I was using a Power5 system that I had borrowed and the folks I borrowed it from have reformatted the drive. I never bothered to hang on to the patches because they were neither particularly pretty or particularly useful. Sorry.
-Mitch