aboutsummaryrefslogtreecommitdiffstats
path: root/drivers
diff options
context:
space:
mode:
authorAyaz Abdulla <aabdulla@nvidia.com>2007-01-21 18:10:52 -0500
committerJeff Garzik <jeff@garzik.org>2007-02-05 16:58:48 -0500
commitb01867cbd1853995946c8c838eff93a0885d8bc6 (patch)
tree3e161800a0eaf8fb1af43f36750be62be0c5cfa3 /drivers
parent445583b89d71b48cf8c64e26acc5a710248feed7 (diff)
forcedeth: rx data path optimization
This patch optimizes the rx data paths and cleans up the code. Signed-Off-By: Ayaz Abdulla <aabdulla@nvidia.com> Signed-off-by: Jeff Garzik <jeff@garzik.org>
Diffstat (limited to 'drivers')
-rw-r--r--drivers/net/forcedeth.c330
1 files changed, 148 insertions, 182 deletions
diff --git a/drivers/net/forcedeth.c b/drivers/net/forcedeth.c
index 27b6bf846000..fd91071bbc82 100644
--- a/drivers/net/forcedeth.c
+++ b/drivers/net/forcedeth.c
@@ -1317,9 +1317,9 @@ static int nv_alloc_rx(struct net_device *dev)
1317 np->put_rx.orig->buf = cpu_to_le32(np->put_rx_ctx->dma); 1317 np->put_rx.orig->buf = cpu_to_le32(np->put_rx_ctx->dma);
1318 wmb(); 1318 wmb();
1319 np->put_rx.orig->flaglen = cpu_to_le32(np->rx_buf_sz | NV_RX_AVAIL); 1319 np->put_rx.orig->flaglen = cpu_to_le32(np->rx_buf_sz | NV_RX_AVAIL);
1320 if (np->put_rx.orig++ == np->last_rx.orig) 1320 if (unlikely(np->put_rx.orig++ == np->last_rx.orig))
1321 np->put_rx.orig = np->first_rx.orig; 1321 np->put_rx.orig = np->first_rx.orig;
1322 if (np->put_rx_ctx++ == np->last_rx_ctx) 1322 if (unlikely(np->put_rx_ctx++ == np->last_rx_ctx))
1323 np->put_rx_ctx = np->first_rx_ctx; 1323 np->put_rx_ctx = np->first_rx_ctx;
1324 } else { 1324 } else {
1325 return 1; 1325 return 1;
@@ -1349,9 +1349,9 @@ static int nv_alloc_rx_optimized(struct net_device *dev)
1349 np->put_rx.ex->buflow = cpu_to_le64(np->put_rx_ctx->dma) & 0x0FFFFFFFF; 1349 np->put_rx.ex->buflow = cpu_to_le64(np->put_rx_ctx->dma) & 0x0FFFFFFFF;
1350 wmb(); 1350 wmb();
1351 np->put_rx.ex->flaglen = cpu_to_le32(np->rx_buf_sz | NV_RX2_AVAIL); 1351 np->put_rx.ex->flaglen = cpu_to_le32(np->rx_buf_sz | NV_RX2_AVAIL);
1352 if (np->put_rx.ex++ == np->last_rx.ex) 1352 if (unlikely(np->put_rx.ex++ == np->last_rx.ex))
1353 np->put_rx.ex = np->first_rx.ex; 1353 np->put_rx.ex = np->first_rx.ex;
1354 if (np->put_rx_ctx++ == np->last_rx_ctx) 1354 if (unlikely(np->put_rx_ctx++ == np->last_rx_ctx))
1355 np->put_rx_ctx = np->first_rx_ctx; 1355 np->put_rx_ctx = np->first_rx_ctx;
1356 } else { 1356 } else {
1357 return 1; 1357 return 1;
@@ -2046,24 +2046,17 @@ static int nv_rx_process(struct net_device *dev, int limit)
2046{ 2046{
2047 struct fe_priv *np = netdev_priv(dev); 2047 struct fe_priv *np = netdev_priv(dev);
2048 u32 flags; 2048 u32 flags;
2049 u32 vlanflags = 0; 2049 u32 rx_processed_cnt = 0;
2050 int count; 2050 struct sk_buff *skb;
2051 2051 int len;
2052 for (count = 0; count < limit; ++count) {
2053 struct sk_buff *skb;
2054 int len;
2055 2052
2056 if (np->get_rx.orig == np->put_rx.orig) 2053 while((np->get_rx.orig != np->put_rx.orig) &&
2057 break; /* we scanned the whole ring - do not continue */ 2054 !((flags = le32_to_cpu(np->get_rx.orig->flaglen)) & NV_RX_AVAIL) &&
2058 flags = le32_to_cpu(np->get_rx.orig->flaglen); 2055 (rx_processed_cnt++ < limit)) {
2059 len = nv_descr_getlength(np->get_rx.orig, np->desc_ver);
2060 2056
2061 dprintk(KERN_DEBUG "%s: nv_rx_process: flags 0x%x.\n", 2057 dprintk(KERN_DEBUG "%s: nv_rx_process: flags 0x%x.\n",
2062 dev->name, flags); 2058 dev->name, flags);
2063 2059
2064 if (flags & NV_RX_AVAIL)
2065 break; /* still owned by hardware, */
2066
2067 /* 2060 /*
2068 * the packet is for us - immediately tear down the pci mapping. 2061 * the packet is for us - immediately tear down the pci mapping.
2069 * TODO: check if a prefetch of the first cacheline improves 2062 * TODO: check if a prefetch of the first cacheline improves
@@ -2087,99 +2080,80 @@ static int nv_rx_process(struct net_device *dev, int limit)
2087 } 2080 }
2088 /* look at what we actually got: */ 2081 /* look at what we actually got: */
2089 if (np->desc_ver == DESC_VER_1) { 2082 if (np->desc_ver == DESC_VER_1) {
2090 if (!(flags & NV_RX_DESCRIPTORVALID)) { 2083 if (likely(flags & NV_RX_DESCRIPTORVALID)) {
2091 dev_kfree_skb(skb); 2084 len = flags & LEN_MASK_V1;
2092 goto next_pkt; 2085 if (unlikely(flags & NV_RX_ERROR)) {
2093 } 2086 if (flags & NV_RX_ERROR4) {
2094 2087 len = nv_getlen(dev, skb->data, len);
2095 if (flags & NV_RX_ERROR) { 2088 if (len < 0) {
2096 if (flags & NV_RX_MISSEDFRAME) { 2089 np->stats.rx_errors++;
2097 np->stats.rx_missed_errors++; 2090 dev_kfree_skb(skb);
2098 np->stats.rx_errors++; 2091 goto next_pkt;
2099 dev_kfree_skb(skb); 2092 }
2100 goto next_pkt; 2093 }
2101 } 2094 /* framing errors are soft errors */
2102 if (flags & (NV_RX_ERROR1|NV_RX_ERROR2|NV_RX_ERROR3)) { 2095 else if (flags & NV_RX_FRAMINGERR) {
2103 np->stats.rx_errors++; 2096 if (flags & NV_RX_SUBSTRACT1) {
2104 dev_kfree_skb(skb); 2097 len--;
2105 goto next_pkt; 2098 }
2106 } 2099 }
2107 if (flags & NV_RX_CRCERR) { 2100 /* the rest are hard errors */
2108 np->stats.rx_crc_errors++; 2101 else {
2109 np->stats.rx_errors++; 2102 if (flags & NV_RX_MISSEDFRAME)
2110 dev_kfree_skb(skb); 2103 np->stats.rx_missed_errors++;
2111 goto next_pkt; 2104 if (flags & NV_RX_CRCERR)
2112 } 2105 np->stats.rx_crc_errors++;
2113 if (flags & NV_RX_OVERFLOW) { 2106 if (flags & NV_RX_OVERFLOW)
2114 np->stats.rx_over_errors++; 2107 np->stats.rx_over_errors++;
2115 np->stats.rx_errors++;
2116 dev_kfree_skb(skb);
2117 goto next_pkt;
2118 }
2119 if (flags & NV_RX_ERROR4) {
2120 len = nv_getlen(dev, skb->data, len);
2121 if (len < 0) {
2122 np->stats.rx_errors++; 2108 np->stats.rx_errors++;
2123 dev_kfree_skb(skb); 2109 dev_kfree_skb(skb);
2124 goto next_pkt; 2110 goto next_pkt;
2125 } 2111 }
2126 } 2112 }
2127 /* framing errors are soft errors. */ 2113 } else {
2128 if (flags & NV_RX_FRAMINGERR) {
2129 if (flags & NV_RX_SUBSTRACT1) {
2130 len--;
2131 }
2132 }
2133 }
2134 } else {
2135 if (!(flags & NV_RX2_DESCRIPTORVALID)) {
2136 dev_kfree_skb(skb); 2114 dev_kfree_skb(skb);
2137 goto next_pkt; 2115 goto next_pkt;
2138 } 2116 }
2139 2117 } else {
2140 if (flags & NV_RX2_ERROR) { 2118 if (likely(flags & NV_RX2_DESCRIPTORVALID)) {
2141 if (flags & (NV_RX2_ERROR1|NV_RX2_ERROR2|NV_RX2_ERROR3)) { 2119 len = flags & LEN_MASK_V2;
2142 np->stats.rx_errors++; 2120 if (unlikely(flags & NV_RX2_ERROR)) {
2143 dev_kfree_skb(skb); 2121 if (flags & NV_RX2_ERROR4) {
2144 goto next_pkt; 2122 len = nv_getlen(dev, skb->data, len);
2145 } 2123 if (len < 0) {
2146 if (flags & NV_RX2_CRCERR) { 2124 np->stats.rx_errors++;
2147 np->stats.rx_crc_errors++; 2125 dev_kfree_skb(skb);
2148 np->stats.rx_errors++; 2126 goto next_pkt;
2149 dev_kfree_skb(skb); 2127 }
2150 goto next_pkt; 2128 }
2151 } 2129 /* framing errors are soft errors */
2152 if (flags & NV_RX2_OVERFLOW) { 2130 else if (flags & NV_RX2_FRAMINGERR) {
2153 np->stats.rx_over_errors++; 2131 if (flags & NV_RX2_SUBSTRACT1) {
2154 np->stats.rx_errors++; 2132 len--;
2155 dev_kfree_skb(skb); 2133 }
2156 goto next_pkt; 2134 }
2157 } 2135 /* the rest are hard errors */
2158 if (flags & NV_RX2_ERROR4) { 2136 else {
2159 len = nv_getlen(dev, skb->data, len); 2137 if (flags & NV_RX2_CRCERR)
2160 if (len < 0) { 2138 np->stats.rx_crc_errors++;
2139 if (flags & NV_RX2_OVERFLOW)
2140 np->stats.rx_over_errors++;
2161 np->stats.rx_errors++; 2141 np->stats.rx_errors++;
2162 dev_kfree_skb(skb); 2142 dev_kfree_skb(skb);
2163 goto next_pkt; 2143 goto next_pkt;
2164 } 2144 }
2165 } 2145 }
2166 /* framing errors are soft errors */ 2146 if ((flags & NV_RX2_CHECKSUMMASK) == NV_RX2_CHECKSUMOK2)/*ip and tcp */ {
2167 if (flags & NV_RX2_FRAMINGERR) {
2168 if (flags & NV_RX2_SUBSTRACT1) {
2169 len--;
2170 }
2171 }
2172 }
2173 if (np->rx_csum) {
2174 flags &= NV_RX2_CHECKSUMMASK;
2175 if (flags == NV_RX2_CHECKSUMOK1 ||
2176 flags == NV_RX2_CHECKSUMOK2 ||
2177 flags == NV_RX2_CHECKSUMOK3) {
2178 dprintk(KERN_DEBUG "%s: hw checksum hit!.\n", dev->name);
2179 skb->ip_summed = CHECKSUM_UNNECESSARY; 2147 skb->ip_summed = CHECKSUM_UNNECESSARY;
2180 } else { 2148 } else {
2181 dprintk(KERN_DEBUG "%s: hwchecksum miss!.\n", dev->name); 2149 if ((flags & NV_RX2_CHECKSUMMASK) == NV_RX2_CHECKSUMOK1 ||
2150 (flags & NV_RX2_CHECKSUMMASK) == NV_RX2_CHECKSUMOK3) {
2151 skb->ip_summed = CHECKSUM_UNNECESSARY;
2152 }
2182 } 2153 }
2154 } else {
2155 dev_kfree_skb(skb);
2156 goto next_pkt;
2183 } 2157 }
2184 } 2158 }
2185 /* got a valid packet - forward it to the network core */ 2159 /* got a valid packet - forward it to the network core */
@@ -2188,29 +2162,21 @@ static int nv_rx_process(struct net_device *dev, int limit)
2188 dprintk(KERN_DEBUG "%s: nv_rx_process: %d bytes, proto %d accepted.\n", 2162 dprintk(KERN_DEBUG "%s: nv_rx_process: %d bytes, proto %d accepted.\n",
2189 dev->name, len, skb->protocol); 2163 dev->name, len, skb->protocol);
2190#ifdef CONFIG_FORCEDETH_NAPI 2164#ifdef CONFIG_FORCEDETH_NAPI
2191 if (np->vlangrp && (vlanflags & NV_RX3_VLAN_TAG_PRESENT)) 2165 netif_receive_skb(skb);
2192 vlan_hwaccel_receive_skb(skb, np->vlangrp,
2193 vlanflags & NV_RX3_VLAN_TAG_MASK);
2194 else
2195 netif_receive_skb(skb);
2196#else 2166#else
2197 if (np->vlangrp && (vlanflags & NV_RX3_VLAN_TAG_PRESENT)) 2167 netif_rx(skb);
2198 vlan_hwaccel_rx(skb, np->vlangrp,
2199 vlanflags & NV_RX3_VLAN_TAG_MASK);
2200 else
2201 netif_rx(skb);
2202#endif 2168#endif
2203 dev->last_rx = jiffies; 2169 dev->last_rx = jiffies;
2204 np->stats.rx_packets++; 2170 np->stats.rx_packets++;
2205 np->stats.rx_bytes += len; 2171 np->stats.rx_bytes += len;
2206next_pkt: 2172next_pkt:
2207 if (np->get_rx.orig++ == np->last_rx.orig) 2173 if (unlikely(np->get_rx.orig++ == np->last_rx.orig))
2208 np->get_rx.orig = np->first_rx.orig; 2174 np->get_rx.orig = np->first_rx.orig;
2209 if (np->get_rx_ctx++ == np->last_rx_ctx) 2175 if (unlikely(np->get_rx_ctx++ == np->last_rx_ctx))
2210 np->get_rx_ctx = np->first_rx_ctx; 2176 np->get_rx_ctx = np->first_rx_ctx;
2211 } 2177 }
2212 2178
2213 return count; 2179 return rx_processed_cnt;
2214} 2180}
2215 2181
2216static int nv_rx_process_optimized(struct net_device *dev, int limit) 2182static int nv_rx_process_optimized(struct net_device *dev, int limit)
@@ -2218,24 +2184,17 @@ static int nv_rx_process_optimized(struct net_device *dev, int limit)
2218 struct fe_priv *np = netdev_priv(dev); 2184 struct fe_priv *np = netdev_priv(dev);
2219 u32 flags; 2185 u32 flags;
2220 u32 vlanflags = 0; 2186 u32 vlanflags = 0;
2221 int count; 2187 u32 rx_processed_cnt = 0;
2222 2188 struct sk_buff *skb;
2223 for (count = 0; count < limit; ++count) { 2189 int len;
2224 struct sk_buff *skb;
2225 int len;
2226 2190
2227 if (np->get_rx.ex == np->put_rx.ex) 2191 while((np->get_rx.ex != np->put_rx.ex) &&
2228 break; /* we scanned the whole ring - do not continue */ 2192 !((flags = le32_to_cpu(np->get_rx.ex->flaglen)) & NV_RX2_AVAIL) &&
2229 flags = le32_to_cpu(np->get_rx.ex->flaglen); 2193 (rx_processed_cnt++ < limit)) {
2230 len = nv_descr_getlength_ex(np->get_rx.ex, np->desc_ver);
2231 vlanflags = le32_to_cpu(np->get_rx.ex->buflow);
2232 2194
2233 dprintk(KERN_DEBUG "%s: nv_rx_process_optimized: flags 0x%x.\n", 2195 dprintk(KERN_DEBUG "%s: nv_rx_process_optimized: flags 0x%x.\n",
2234 dev->name, flags); 2196 dev->name, flags);
2235 2197
2236 if (flags & NV_RX_AVAIL)
2237 break; /* still owned by hardware, */
2238
2239 /* 2198 /*
2240 * the packet is for us - immediately tear down the pci mapping. 2199 * the packet is for us - immediately tear down the pci mapping.
2241 * TODO: check if a prefetch of the first cacheline improves 2200 * TODO: check if a prefetch of the first cacheline improves
@@ -2258,84 +2217,91 @@ static int nv_rx_process_optimized(struct net_device *dev, int limit)
2258 dprintk("\n"); 2217 dprintk("\n");
2259 } 2218 }
2260 /* look at what we actually got: */ 2219 /* look at what we actually got: */
2261 if (!(flags & NV_RX2_DESCRIPTORVALID)) { 2220 if (likely(flags & NV_RX2_DESCRIPTORVALID)) {
2262 dev_kfree_skb(skb); 2221 len = flags & LEN_MASK_V2;
2263 goto next_pkt; 2222 if (unlikely(flags & NV_RX2_ERROR)) {
2264 } 2223 if (flags & NV_RX2_ERROR4) {
2265 2224 len = nv_getlen(dev, skb->data, len);
2266 if (flags & NV_RX2_ERROR) { 2225 if (len < 0) {
2267 if (flags & (NV_RX2_ERROR1|NV_RX2_ERROR2|NV_RX2_ERROR3)) { 2226 np->stats.rx_errors++;
2268 np->stats.rx_errors++; 2227 dev_kfree_skb(skb);
2269 dev_kfree_skb(skb); 2228 goto next_pkt;
2270 goto next_pkt; 2229 }
2271 } 2230 }
2272 if (flags & NV_RX2_CRCERR) { 2231 /* framing errors are soft errors */
2273 np->stats.rx_crc_errors++; 2232 else if (flags & NV_RX2_FRAMINGERR) {
2274 np->stats.rx_errors++; 2233 if (flags & NV_RX2_SUBSTRACT1) {
2275 dev_kfree_skb(skb); 2234 len--;
2276 goto next_pkt; 2235 }
2277 } 2236 }
2278 if (flags & NV_RX2_OVERFLOW) { 2237 /* the rest are hard errors */
2279 np->stats.rx_over_errors++; 2238 else {
2280 np->stats.rx_errors++; 2239 if (flags & NV_RX2_CRCERR)
2281 dev_kfree_skb(skb); 2240 np->stats.rx_crc_errors++;
2282 goto next_pkt; 2241 if (flags & NV_RX2_OVERFLOW)
2283 } 2242 np->stats.rx_over_errors++;
2284 if (flags & NV_RX2_ERROR4) {
2285 len = nv_getlen(dev, skb->data, len);
2286 if (len < 0) {
2287 np->stats.rx_errors++; 2243 np->stats.rx_errors++;
2288 dev_kfree_skb(skb); 2244 dev_kfree_skb(skb);
2289 goto next_pkt; 2245 goto next_pkt;
2290 } 2246 }
2291 } 2247 }
2292 /* framing errors are soft errors */ 2248
2293 if (flags & NV_RX2_FRAMINGERR) { 2249 if ((flags & NV_RX2_CHECKSUMMASK) == NV_RX2_CHECKSUMOK2)/*ip and tcp */ {
2294 if (flags & NV_RX2_SUBSTRACT1) {
2295 len--;
2296 }
2297 }
2298 }
2299 if (np->rx_csum) {
2300 flags &= NV_RX2_CHECKSUMMASK;
2301 if (flags == NV_RX2_CHECKSUMOK1 ||
2302 flags == NV_RX2_CHECKSUMOK2 ||
2303 flags == NV_RX2_CHECKSUMOK3) {
2304 dprintk(KERN_DEBUG "%s: hw checksum hit!.\n", dev->name);
2305 skb->ip_summed = CHECKSUM_UNNECESSARY; 2250 skb->ip_summed = CHECKSUM_UNNECESSARY;
2306 } else { 2251 } else {
2307 dprintk(KERN_DEBUG "%s: hwchecksum miss!.\n", dev->name); 2252 if ((flags & NV_RX2_CHECKSUMMASK) == NV_RX2_CHECKSUMOK1 ||
2253 (flags & NV_RX2_CHECKSUMMASK) == NV_RX2_CHECKSUMOK3) {
2254 skb->ip_summed = CHECKSUM_UNNECESSARY;
2255 }
2308 } 2256 }
2309 } 2257
2310 /* got a valid packet - forward it to the network core */ 2258 /* got a valid packet - forward it to the network core */
2311 skb_put(skb, len); 2259 skb_put(skb, len);
2312 skb->protocol = eth_type_trans(skb, dev); 2260 skb->protocol = eth_type_trans(skb, dev);
2313 dprintk(KERN_DEBUG "%s: nv_rx_process: %d bytes, proto %d accepted.\n", 2261 prefetch(skb->data);
2314 dev->name, len, skb->protocol); 2262
2263 dprintk(KERN_DEBUG "%s: nv_rx_process_optimized: %d bytes, proto %d accepted.\n",
2264 dev->name, len, skb->protocol);
2265
2266 if (likely(!np->vlangrp)) {
2315#ifdef CONFIG_FORCEDETH_NAPI 2267#ifdef CONFIG_FORCEDETH_NAPI
2316 if (np->vlangrp && (vlanflags & NV_RX3_VLAN_TAG_PRESENT)) 2268 netif_receive_skb(skb);
2317 vlan_hwaccel_receive_skb(skb, np->vlangrp,
2318 vlanflags & NV_RX3_VLAN_TAG_MASK);
2319 else
2320 netif_receive_skb(skb);
2321#else 2269#else
2322 if (np->vlangrp && (vlanflags & NV_RX3_VLAN_TAG_PRESENT)) 2270 netif_rx(skb);
2323 vlan_hwaccel_rx(skb, np->vlangrp,
2324 vlanflags & NV_RX3_VLAN_TAG_MASK);
2325 else
2326 netif_rx(skb);
2327#endif 2271#endif
2328 dev->last_rx = jiffies; 2272 } else {
2329 np->stats.rx_packets++; 2273 vlanflags = le32_to_cpu(np->get_rx.ex->buflow);
2330 np->stats.rx_bytes += len; 2274 if (vlanflags & NV_RX3_VLAN_TAG_PRESENT) {
2275#ifdef CONFIG_FORCEDETH_NAPI
2276 vlan_hwaccel_receive_skb(skb, np->vlangrp,
2277 vlanflags & NV_RX3_VLAN_TAG_MASK);
2278#else
2279 vlan_hwaccel_rx(skb, np->vlangrp,
2280 vlanflags & NV_RX3_VLAN_TAG_MASK);
2281#endif
2282 } else {
2283#ifdef CONFIG_FORCEDETH_NAPI
2284 netif_receive_skb(skb);
2285#else
2286 netif_rx(skb);
2287#endif
2288 }
2289 }
2290
2291 dev->last_rx = jiffies;
2292 np->stats.rx_packets++;
2293 np->stats.rx_bytes += len;
2294 } else {
2295 dev_kfree_skb(skb);
2296 }
2331next_pkt: 2297next_pkt:
2332 if (np->get_rx.ex++ == np->last_rx.ex) 2298 if (unlikely(np->get_rx.ex++ == np->last_rx.ex))
2333 np->get_rx.ex = np->first_rx.ex; 2299 np->get_rx.ex = np->first_rx.ex;
2334 if (np->get_rx_ctx++ == np->last_rx_ctx) 2300 if (unlikely(np->get_rx_ctx++ == np->last_rx_ctx))
2335 np->get_rx_ctx = np->first_rx_ctx; 2301 np->get_rx_ctx = np->first_rx_ctx;
2336 } 2302 }
2337 2303
2338 return count; 2304 return rx_processed_cnt;
2339} 2305}
2340 2306
2341static void set_bufsize(struct net_device *dev) 2307static void set_bufsize(struct net_device *dev)
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-28 10:17:57 -0500