Ben Hutchings | 8ceee66 | 2008-04-27 12:55:59 +0100 | [diff] [blame^] | 1 | /**************************************************************************** |
| 2 | * Driver for Solarflare Solarstorm network controllers and boards |
| 3 | * Copyright 2005-2006 Fen Systems Ltd. |
| 4 | * Copyright 2005-2008 Solarflare Communications Inc. |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify it |
| 7 | * under the terms of the GNU General Public License version 2 as published |
| 8 | * by the Free Software Foundation, incorporated herein by reference. |
| 9 | */ |
| 10 | |
| 11 | #include <linux/pci.h> |
| 12 | #include <linux/tcp.h> |
| 13 | #include <linux/ip.h> |
| 14 | #include <linux/in.h> |
| 15 | #include <linux/if_ether.h> |
| 16 | #include <linux/highmem.h> |
| 17 | #include "net_driver.h" |
| 18 | #include "tx.h" |
| 19 | #include "efx.h" |
| 20 | #include "falcon.h" |
| 21 | #include "workarounds.h" |
| 22 | |
| 23 | /* |
| 24 | * TX descriptor ring full threshold |
| 25 | * |
| 26 | * The tx_queue descriptor ring fill-level must fall below this value |
| 27 | * before we restart the netif queue |
| 28 | */ |
| 29 | #define EFX_NETDEV_TX_THRESHOLD(_tx_queue) \ |
| 30 | (_tx_queue->efx->type->txd_ring_mask / 2u) |
| 31 | |
| 32 | /* We want to be able to nest calls to netif_stop_queue(), since each |
| 33 | * channel can have an individual stop on the queue. |
| 34 | */ |
| 35 | void efx_stop_queue(struct efx_nic *efx) |
| 36 | { |
| 37 | spin_lock_bh(&efx->netif_stop_lock); |
| 38 | EFX_TRACE(efx, "stop TX queue\n"); |
| 39 | |
| 40 | atomic_inc(&efx->netif_stop_count); |
| 41 | netif_stop_queue(efx->net_dev); |
| 42 | |
| 43 | spin_unlock_bh(&efx->netif_stop_lock); |
| 44 | } |
| 45 | |
| 46 | /* Wake netif's TX queue |
| 47 | * We want to be able to nest calls to netif_stop_queue(), since each |
| 48 | * channel can have an individual stop on the queue. |
| 49 | */ |
| 50 | inline void efx_wake_queue(struct efx_nic *efx) |
| 51 | { |
| 52 | local_bh_disable(); |
| 53 | if (atomic_dec_and_lock(&efx->netif_stop_count, |
| 54 | &efx->netif_stop_lock)) { |
| 55 | EFX_TRACE(efx, "waking TX queue\n"); |
| 56 | netif_wake_queue(efx->net_dev); |
| 57 | spin_unlock(&efx->netif_stop_lock); |
| 58 | } |
| 59 | local_bh_enable(); |
| 60 | } |
| 61 | |
| 62 | static inline void efx_dequeue_buffer(struct efx_tx_queue *tx_queue, |
| 63 | struct efx_tx_buffer *buffer) |
| 64 | { |
| 65 | if (buffer->unmap_len) { |
| 66 | struct pci_dev *pci_dev = tx_queue->efx->pci_dev; |
| 67 | if (buffer->unmap_single) |
| 68 | pci_unmap_single(pci_dev, buffer->unmap_addr, |
| 69 | buffer->unmap_len, PCI_DMA_TODEVICE); |
| 70 | else |
| 71 | pci_unmap_page(pci_dev, buffer->unmap_addr, |
| 72 | buffer->unmap_len, PCI_DMA_TODEVICE); |
| 73 | buffer->unmap_len = 0; |
| 74 | buffer->unmap_single = 0; |
| 75 | } |
| 76 | |
| 77 | if (buffer->skb) { |
| 78 | dev_kfree_skb_any((struct sk_buff *) buffer->skb); |
| 79 | buffer->skb = NULL; |
| 80 | EFX_TRACE(tx_queue->efx, "TX queue %d transmission id %x " |
| 81 | "complete\n", tx_queue->queue, read_ptr); |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | |
| 86 | /* |
| 87 | * Add a socket buffer to a TX queue |
| 88 | * |
| 89 | * This maps all fragments of a socket buffer for DMA and adds them to |
| 90 | * the TX queue. The queue's insert pointer will be incremented by |
| 91 | * the number of fragments in the socket buffer. |
| 92 | * |
| 93 | * If any DMA mapping fails, any mapped fragments will be unmapped, |
| 94 | * the queue's insert pointer will be restored to its original value. |
| 95 | * |
| 96 | * Returns NETDEV_TX_OK or NETDEV_TX_BUSY |
| 97 | * You must hold netif_tx_lock() to call this function. |
| 98 | */ |
| 99 | static inline int efx_enqueue_skb(struct efx_tx_queue *tx_queue, |
| 100 | const struct sk_buff *skb) |
| 101 | { |
| 102 | struct efx_nic *efx = tx_queue->efx; |
| 103 | struct pci_dev *pci_dev = efx->pci_dev; |
| 104 | struct efx_tx_buffer *buffer; |
| 105 | skb_frag_t *fragment; |
| 106 | struct page *page; |
| 107 | int page_offset; |
| 108 | unsigned int len, unmap_len = 0, fill_level, insert_ptr, misalign; |
| 109 | dma_addr_t dma_addr, unmap_addr = 0; |
| 110 | unsigned int dma_len; |
| 111 | unsigned unmap_single; |
| 112 | int q_space, i = 0; |
| 113 | int rc = NETDEV_TX_OK; |
| 114 | |
| 115 | EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count); |
| 116 | |
| 117 | /* Get size of the initial fragment */ |
| 118 | len = skb_headlen(skb); |
| 119 | |
| 120 | fill_level = tx_queue->insert_count - tx_queue->old_read_count; |
| 121 | q_space = efx->type->txd_ring_mask - 1 - fill_level; |
| 122 | |
| 123 | /* Map for DMA. Use pci_map_single rather than pci_map_page |
| 124 | * since this is more efficient on machines with sparse |
| 125 | * memory. |
| 126 | */ |
| 127 | unmap_single = 1; |
| 128 | dma_addr = pci_map_single(pci_dev, skb->data, len, PCI_DMA_TODEVICE); |
| 129 | |
| 130 | /* Process all fragments */ |
| 131 | while (1) { |
| 132 | if (unlikely(pci_dma_mapping_error(dma_addr))) |
| 133 | goto pci_err; |
| 134 | |
| 135 | /* Store fields for marking in the per-fragment final |
| 136 | * descriptor */ |
| 137 | unmap_len = len; |
| 138 | unmap_addr = dma_addr; |
| 139 | |
| 140 | /* Add to TX queue, splitting across DMA boundaries */ |
| 141 | do { |
| 142 | if (unlikely(q_space-- <= 0)) { |
| 143 | /* It might be that completions have |
| 144 | * happened since the xmit path last |
| 145 | * checked. Update the xmit path's |
| 146 | * copy of read_count. |
| 147 | */ |
| 148 | ++tx_queue->stopped; |
| 149 | /* This memory barrier protects the |
| 150 | * change of stopped from the access |
| 151 | * of read_count. */ |
| 152 | smp_mb(); |
| 153 | tx_queue->old_read_count = |
| 154 | *(volatile unsigned *) |
| 155 | &tx_queue->read_count; |
| 156 | fill_level = (tx_queue->insert_count |
| 157 | - tx_queue->old_read_count); |
| 158 | q_space = (efx->type->txd_ring_mask - 1 - |
| 159 | fill_level); |
| 160 | if (unlikely(q_space-- <= 0)) |
| 161 | goto stop; |
| 162 | smp_mb(); |
| 163 | --tx_queue->stopped; |
| 164 | } |
| 165 | |
| 166 | insert_ptr = (tx_queue->insert_count & |
| 167 | efx->type->txd_ring_mask); |
| 168 | buffer = &tx_queue->buffer[insert_ptr]; |
| 169 | EFX_BUG_ON_PARANOID(buffer->skb); |
| 170 | EFX_BUG_ON_PARANOID(buffer->len); |
| 171 | EFX_BUG_ON_PARANOID(buffer->continuation != 1); |
| 172 | EFX_BUG_ON_PARANOID(buffer->unmap_len); |
| 173 | |
| 174 | dma_len = (((~dma_addr) & efx->type->tx_dma_mask) + 1); |
| 175 | if (likely(dma_len > len)) |
| 176 | dma_len = len; |
| 177 | |
| 178 | misalign = (unsigned)dma_addr & efx->type->bug5391_mask; |
| 179 | if (misalign && dma_len + misalign > 512) |
| 180 | dma_len = 512 - misalign; |
| 181 | |
| 182 | /* Fill out per descriptor fields */ |
| 183 | buffer->len = dma_len; |
| 184 | buffer->dma_addr = dma_addr; |
| 185 | len -= dma_len; |
| 186 | dma_addr += dma_len; |
| 187 | ++tx_queue->insert_count; |
| 188 | } while (len); |
| 189 | |
| 190 | /* Transfer ownership of the unmapping to the final buffer */ |
| 191 | buffer->unmap_addr = unmap_addr; |
| 192 | buffer->unmap_single = unmap_single; |
| 193 | buffer->unmap_len = unmap_len; |
| 194 | unmap_len = 0; |
| 195 | |
| 196 | /* Get address and size of next fragment */ |
| 197 | if (i >= skb_shinfo(skb)->nr_frags) |
| 198 | break; |
| 199 | fragment = &skb_shinfo(skb)->frags[i]; |
| 200 | len = fragment->size; |
| 201 | page = fragment->page; |
| 202 | page_offset = fragment->page_offset; |
| 203 | i++; |
| 204 | /* Map for DMA */ |
| 205 | unmap_single = 0; |
| 206 | dma_addr = pci_map_page(pci_dev, page, page_offset, len, |
| 207 | PCI_DMA_TODEVICE); |
| 208 | } |
| 209 | |
| 210 | /* Transfer ownership of the skb to the final buffer */ |
| 211 | buffer->skb = skb; |
| 212 | buffer->continuation = 0; |
| 213 | |
| 214 | /* Pass off to hardware */ |
| 215 | falcon_push_buffers(tx_queue); |
| 216 | |
| 217 | return NETDEV_TX_OK; |
| 218 | |
| 219 | pci_err: |
| 220 | EFX_ERR_RL(efx, " TX queue %d could not map skb with %d bytes %d " |
| 221 | "fragments for DMA\n", tx_queue->queue, skb->len, |
| 222 | skb_shinfo(skb)->nr_frags + 1); |
| 223 | |
| 224 | /* Mark the packet as transmitted, and free the SKB ourselves */ |
| 225 | dev_kfree_skb_any((struct sk_buff *)skb); |
| 226 | goto unwind; |
| 227 | |
| 228 | stop: |
| 229 | rc = NETDEV_TX_BUSY; |
| 230 | |
| 231 | if (tx_queue->stopped == 1) |
| 232 | efx_stop_queue(efx); |
| 233 | |
| 234 | unwind: |
| 235 | /* Work backwards until we hit the original insert pointer value */ |
| 236 | while (tx_queue->insert_count != tx_queue->write_count) { |
| 237 | --tx_queue->insert_count; |
| 238 | insert_ptr = tx_queue->insert_count & efx->type->txd_ring_mask; |
| 239 | buffer = &tx_queue->buffer[insert_ptr]; |
| 240 | efx_dequeue_buffer(tx_queue, buffer); |
| 241 | buffer->len = 0; |
| 242 | } |
| 243 | |
| 244 | /* Free the fragment we were mid-way through pushing */ |
| 245 | if (unmap_len) |
| 246 | pci_unmap_page(pci_dev, unmap_addr, unmap_len, |
| 247 | PCI_DMA_TODEVICE); |
| 248 | |
| 249 | return rc; |
| 250 | } |
| 251 | |
| 252 | /* Remove packets from the TX queue |
| 253 | * |
| 254 | * This removes packets from the TX queue, up to and including the |
| 255 | * specified index. |
| 256 | */ |
| 257 | static inline void efx_dequeue_buffers(struct efx_tx_queue *tx_queue, |
| 258 | unsigned int index) |
| 259 | { |
| 260 | struct efx_nic *efx = tx_queue->efx; |
| 261 | unsigned int stop_index, read_ptr; |
| 262 | unsigned int mask = tx_queue->efx->type->txd_ring_mask; |
| 263 | |
| 264 | stop_index = (index + 1) & mask; |
| 265 | read_ptr = tx_queue->read_count & mask; |
| 266 | |
| 267 | while (read_ptr != stop_index) { |
| 268 | struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr]; |
| 269 | if (unlikely(buffer->len == 0)) { |
| 270 | EFX_ERR(tx_queue->efx, "TX queue %d spurious TX " |
| 271 | "completion id %x\n", tx_queue->queue, |
| 272 | read_ptr); |
| 273 | efx_schedule_reset(efx, RESET_TYPE_TX_SKIP); |
| 274 | return; |
| 275 | } |
| 276 | |
| 277 | efx_dequeue_buffer(tx_queue, buffer); |
| 278 | buffer->continuation = 1; |
| 279 | buffer->len = 0; |
| 280 | |
| 281 | ++tx_queue->read_count; |
| 282 | read_ptr = tx_queue->read_count & mask; |
| 283 | } |
| 284 | } |
| 285 | |
| 286 | /* Initiate a packet transmission on the specified TX queue. |
| 287 | * Note that returning anything other than NETDEV_TX_OK will cause the |
| 288 | * OS to free the skb. |
| 289 | * |
| 290 | * This function is split out from efx_hard_start_xmit to allow the |
| 291 | * loopback test to direct packets via specific TX queues. It is |
| 292 | * therefore a non-static inline, so as not to penalise performance |
| 293 | * for non-loopback transmissions. |
| 294 | * |
| 295 | * Context: netif_tx_lock held |
| 296 | */ |
| 297 | inline int efx_xmit(struct efx_nic *efx, |
| 298 | struct efx_tx_queue *tx_queue, struct sk_buff *skb) |
| 299 | { |
| 300 | int rc; |
| 301 | |
| 302 | /* Map fragments for DMA and add to TX queue */ |
| 303 | rc = efx_enqueue_skb(tx_queue, skb); |
| 304 | if (unlikely(rc != NETDEV_TX_OK)) |
| 305 | goto out; |
| 306 | |
| 307 | /* Update last TX timer */ |
| 308 | efx->net_dev->trans_start = jiffies; |
| 309 | |
| 310 | out: |
| 311 | return rc; |
| 312 | } |
| 313 | |
| 314 | /* Initiate a packet transmission. We use one channel per CPU |
| 315 | * (sharing when we have more CPUs than channels). On Falcon, the TX |
| 316 | * completion events will be directed back to the CPU that transmitted |
| 317 | * the packet, which should be cache-efficient. |
| 318 | * |
| 319 | * Context: non-blocking. |
| 320 | * Note that returning anything other than NETDEV_TX_OK will cause the |
| 321 | * OS to free the skb. |
| 322 | */ |
| 323 | int efx_hard_start_xmit(struct sk_buff *skb, struct net_device *net_dev) |
| 324 | { |
| 325 | struct efx_nic *efx = net_dev->priv; |
| 326 | return efx_xmit(efx, &efx->tx_queue[0], skb); |
| 327 | } |
| 328 | |
| 329 | void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index) |
| 330 | { |
| 331 | unsigned fill_level; |
| 332 | struct efx_nic *efx = tx_queue->efx; |
| 333 | |
| 334 | EFX_BUG_ON_PARANOID(index > efx->type->txd_ring_mask); |
| 335 | |
| 336 | efx_dequeue_buffers(tx_queue, index); |
| 337 | |
| 338 | /* See if we need to restart the netif queue. This barrier |
| 339 | * separates the update of read_count from the test of |
| 340 | * stopped. */ |
| 341 | smp_mb(); |
| 342 | if (unlikely(tx_queue->stopped)) { |
| 343 | fill_level = tx_queue->insert_count - tx_queue->read_count; |
| 344 | if (fill_level < EFX_NETDEV_TX_THRESHOLD(tx_queue)) { |
| 345 | EFX_BUG_ON_PARANOID(!NET_DEV_REGISTERED(efx)); |
| 346 | |
| 347 | /* Do this under netif_tx_lock(), to avoid racing |
| 348 | * with efx_xmit(). */ |
| 349 | netif_tx_lock(efx->net_dev); |
| 350 | if (tx_queue->stopped) { |
| 351 | tx_queue->stopped = 0; |
| 352 | efx_wake_queue(efx); |
| 353 | } |
| 354 | netif_tx_unlock(efx->net_dev); |
| 355 | } |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | int efx_probe_tx_queue(struct efx_tx_queue *tx_queue) |
| 360 | { |
| 361 | struct efx_nic *efx = tx_queue->efx; |
| 362 | unsigned int txq_size; |
| 363 | int i, rc; |
| 364 | |
| 365 | EFX_LOG(efx, "creating TX queue %d\n", tx_queue->queue); |
| 366 | |
| 367 | /* Allocate software ring */ |
| 368 | txq_size = (efx->type->txd_ring_mask + 1) * sizeof(*tx_queue->buffer); |
| 369 | tx_queue->buffer = kzalloc(txq_size, GFP_KERNEL); |
| 370 | if (!tx_queue->buffer) { |
| 371 | rc = -ENOMEM; |
| 372 | goto fail1; |
| 373 | } |
| 374 | for (i = 0; i <= efx->type->txd_ring_mask; ++i) |
| 375 | tx_queue->buffer[i].continuation = 1; |
| 376 | |
| 377 | /* Allocate hardware ring */ |
| 378 | rc = falcon_probe_tx(tx_queue); |
| 379 | if (rc) |
| 380 | goto fail2; |
| 381 | |
| 382 | return 0; |
| 383 | |
| 384 | fail2: |
| 385 | kfree(tx_queue->buffer); |
| 386 | tx_queue->buffer = NULL; |
| 387 | fail1: |
| 388 | tx_queue->used = 0; |
| 389 | |
| 390 | return rc; |
| 391 | } |
| 392 | |
| 393 | int efx_init_tx_queue(struct efx_tx_queue *tx_queue) |
| 394 | { |
| 395 | EFX_LOG(tx_queue->efx, "initialising TX queue %d\n", tx_queue->queue); |
| 396 | |
| 397 | tx_queue->insert_count = 0; |
| 398 | tx_queue->write_count = 0; |
| 399 | tx_queue->read_count = 0; |
| 400 | tx_queue->old_read_count = 0; |
| 401 | BUG_ON(tx_queue->stopped); |
| 402 | |
| 403 | /* Set up TX descriptor ring */ |
| 404 | return falcon_init_tx(tx_queue); |
| 405 | } |
| 406 | |
| 407 | void efx_release_tx_buffers(struct efx_tx_queue *tx_queue) |
| 408 | { |
| 409 | struct efx_tx_buffer *buffer; |
| 410 | |
| 411 | if (!tx_queue->buffer) |
| 412 | return; |
| 413 | |
| 414 | /* Free any buffers left in the ring */ |
| 415 | while (tx_queue->read_count != tx_queue->write_count) { |
| 416 | buffer = &tx_queue->buffer[tx_queue->read_count & |
| 417 | tx_queue->efx->type->txd_ring_mask]; |
| 418 | efx_dequeue_buffer(tx_queue, buffer); |
| 419 | buffer->continuation = 1; |
| 420 | buffer->len = 0; |
| 421 | |
| 422 | ++tx_queue->read_count; |
| 423 | } |
| 424 | } |
| 425 | |
| 426 | void efx_fini_tx_queue(struct efx_tx_queue *tx_queue) |
| 427 | { |
| 428 | EFX_LOG(tx_queue->efx, "shutting down TX queue %d\n", tx_queue->queue); |
| 429 | |
| 430 | /* Flush TX queue, remove descriptor ring */ |
| 431 | falcon_fini_tx(tx_queue); |
| 432 | |
| 433 | efx_release_tx_buffers(tx_queue); |
| 434 | |
| 435 | /* Release queue's stop on port, if any */ |
| 436 | if (tx_queue->stopped) { |
| 437 | tx_queue->stopped = 0; |
| 438 | efx_wake_queue(tx_queue->efx); |
| 439 | } |
| 440 | } |
| 441 | |
| 442 | void efx_remove_tx_queue(struct efx_tx_queue *tx_queue) |
| 443 | { |
| 444 | EFX_LOG(tx_queue->efx, "destroying TX queue %d\n", tx_queue->queue); |
| 445 | falcon_remove_tx(tx_queue); |
| 446 | |
| 447 | kfree(tx_queue->buffer); |
| 448 | tx_queue->buffer = NULL; |
| 449 | tx_queue->used = 0; |
| 450 | } |
| 451 | |
| 452 | |