diff options
Diffstat (limited to 'drivers/net/ethernet/sfc/tx.c')
| -rw-r--r-- | drivers/net/ethernet/sfc/tx.c | 1131 |
1 files changed, 628 insertions, 503 deletions
diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c index 72f0fbc73b1..ede8dcca0ff 100644 --- a/drivers/net/ethernet/sfc/tx.c +++ b/drivers/net/ethernet/sfc/tx.c @@ -1,7 +1,7 @@ /**************************************************************************** - * Driver for Solarflare Solarstorm network controllers and boards + * Driver for Solarflare network controllers and boards * Copyright 2005-2006 Fen Systems Ltd. - * Copyright 2005-2010 Solarflare Communications Inc. + * Copyright 2005-2013 Solarflare Communications Inc. * * This program is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License version 2 as published @@ -17,18 +17,46 @@ #include <net/ipv6.h> #include <linux/if_ether.h> #include <linux/highmem.h> +#include <linux/cache.h> #include "net_driver.h" #include "efx.h" +#include "io.h" #include "nic.h" #include "workarounds.h" +#include "ef10_regs.h" -/* - * TX descriptor ring full threshold - * - * The tx_queue descriptor ring fill-level must fall below this value - * before we restart the netif queue - */ -#define EFX_TXQ_THRESHOLD(_efx) ((_efx)->txq_entries / 2u) +#ifdef EFX_USE_PIO + +#define EFX_PIOBUF_SIZE_MAX ER_DZ_TX_PIOBUF_SIZE +#define EFX_PIOBUF_SIZE_DEF ALIGN(256, L1_CACHE_BYTES) +unsigned int efx_piobuf_size __read_mostly = EFX_PIOBUF_SIZE_DEF; + +#endif /* EFX_USE_PIO */ + +static inline unsigned int +efx_tx_queue_get_insert_index(const struct efx_tx_queue *tx_queue) +{ + return tx_queue->insert_count & tx_queue->ptr_mask; +} + +static inline struct efx_tx_buffer * +__efx_tx_queue_get_insert_buffer(const struct efx_tx_queue *tx_queue) +{ + return &tx_queue->buffer[efx_tx_queue_get_insert_index(tx_queue)]; +} + +static inline struct efx_tx_buffer * +efx_tx_queue_get_insert_buffer(const struct efx_tx_queue *tx_queue) +{ + struct efx_tx_buffer *buffer = + __efx_tx_queue_get_insert_buffer(tx_queue); + + EFX_BUG_ON_PARANOID(buffer->len); + EFX_BUG_ON_PARANOID(buffer->flags); + EFX_BUG_ON_PARANOID(buffer->unmap_len); + + return buffer; +} static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue, struct efx_tx_buffer *buffer, @@ -36,70 +64,34 @@ static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue, unsigned int *bytes_compl) { if (buffer->unmap_len) { - struct pci_dev *pci_dev = tx_queue->efx->pci_dev; - dma_addr_t unmap_addr = (buffer->dma_addr + buffer->len - - buffer->unmap_len); - if (buffer->unmap_single) - pci_unmap_single(pci_dev, unmap_addr, buffer->unmap_len, - PCI_DMA_TODEVICE); + struct device *dma_dev = &tx_queue->efx->pci_dev->dev; + dma_addr_t unmap_addr = buffer->dma_addr - buffer->dma_offset; + if (buffer->flags & EFX_TX_BUF_MAP_SINGLE) + dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len, + DMA_TO_DEVICE); else - pci_unmap_page(pci_dev, unmap_addr, buffer->unmap_len, - PCI_DMA_TODEVICE); + dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len, + DMA_TO_DEVICE); buffer->unmap_len = 0; - buffer->unmap_single = false; } - if (buffer->skb) { + if (buffer->flags & EFX_TX_BUF_SKB) { (*pkts_compl)++; (*bytes_compl) += buffer->skb->len; dev_kfree_skb_any((struct sk_buff *) buffer->skb); - buffer->skb = NULL; netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev, "TX queue %d transmission id %x complete\n", tx_queue->queue, tx_queue->read_count); + } else if (buffer->flags & EFX_TX_BUF_HEAP) { + kfree(buffer->heap_buf); } -} -/** - * struct efx_tso_header - a DMA mapped buffer for packet headers - * @next: Linked list of free ones. - * The list is protected by the TX queue lock. - * @dma_unmap_len: Length to unmap for an oversize buffer, or 0. - * @dma_addr: The DMA address of the header below. - * - * This controls the memory used for a TSO header. Use TSOH_DATA() - * to find the packet header data. Use TSOH_SIZE() to calculate the - * total size required for a given packet header length. TSO headers - * in the free list are exactly %TSOH_STD_SIZE bytes in size. - */ -struct efx_tso_header { - union { - struct efx_tso_header *next; - size_t unmap_len; - }; - dma_addr_t dma_addr; -}; + buffer->len = 0; + buffer->flags = 0; +} static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue, struct sk_buff *skb); -static void efx_fini_tso(struct efx_tx_queue *tx_queue); -static void efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, - struct efx_tso_header *tsoh); - -static void efx_tsoh_free(struct efx_tx_queue *tx_queue, - struct efx_tx_buffer *buffer) -{ - if (buffer->tsoh) { - if (likely(!buffer->tsoh->unmap_len)) { - buffer->tsoh->next = tx_queue->tso_headers_free; - tx_queue->tso_headers_free = buffer->tsoh; - } else { - efx_tsoh_heap_free(tx_queue, buffer->tsoh); - } - buffer->tsoh = NULL; - } -} - static inline unsigned efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr) @@ -110,7 +102,7 @@ efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr) * little benefit from using descriptors that cross those * boundaries and we keep things simple by not doing so. */ - unsigned len = (~dma_addr & 0xfff) + 1; + unsigned len = (~dma_addr & (EFX_PAGE_SIZE - 1)) + 1; /* Work around hardware bug for unaligned buffers. */ if (EFX_WORKAROUND_5391(efx) && (dma_addr & 0xf)) @@ -119,6 +111,228 @@ efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr) return len; } +unsigned int efx_tx_max_skb_descs(struct efx_nic *efx) +{ + /* Header and payload descriptor for each output segment, plus + * one for every input fragment boundary within a segment + */ + unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS; + + /* Possibly one more per segment for the alignment workaround, + * or for option descriptors + */ + if (EFX_WORKAROUND_5391(efx) || efx_nic_rev(efx) >= EFX_REV_HUNT_A0) + max_descs += EFX_TSO_MAX_SEGS; + + /* Possibly more for PCIe page boundaries within input fragments */ + if (PAGE_SIZE > EFX_PAGE_SIZE) + max_descs += max_t(unsigned int, MAX_SKB_FRAGS, + DIV_ROUND_UP(GSO_MAX_SIZE, EFX_PAGE_SIZE)); + + return max_descs; +} + +/* Get partner of a TX queue, seen as part of the same net core queue */ +static struct efx_tx_queue *efx_tx_queue_partner(struct efx_tx_queue *tx_queue) +{ + if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD) + return tx_queue - EFX_TXQ_TYPE_OFFLOAD; + else + return tx_queue + EFX_TXQ_TYPE_OFFLOAD; +} + +static void efx_tx_maybe_stop_queue(struct efx_tx_queue *txq1) +{ + /* We need to consider both queues that the net core sees as one */ + struct efx_tx_queue *txq2 = efx_tx_queue_partner(txq1); + struct efx_nic *efx = txq1->efx; + unsigned int fill_level; + + fill_level = max(txq1->insert_count - txq1->old_read_count, + txq2->insert_count - txq2->old_read_count); + if (likely(fill_level < efx->txq_stop_thresh)) + return; + + /* We used the stale old_read_count above, which gives us a + * pessimistic estimate of the fill level (which may even + * validly be >= efx->txq_entries). Now try again using + * read_count (more likely to be a cache miss). + * + * If we read read_count and then conditionally stop the + * queue, it is possible for the completion path to race with + * us and complete all outstanding descriptors in the middle, + * after which there will be no more completions to wake it. + * Therefore we stop the queue first, then read read_count + * (with a memory barrier to ensure the ordering), then + * restart the queue if the fill level turns out to be low + * enough. + */ + netif_tx_stop_queue(txq1->core_txq); + smp_mb(); + txq1->old_read_count = ACCESS_ONCE(txq1->read_count); + txq2->old_read_count = ACCESS_ONCE(txq2->read_count); + + fill_level = max(txq1->insert_count - txq1->old_read_count, + txq2->insert_count - txq2->old_read_count); + EFX_BUG_ON_PARANOID(fill_level >= efx->txq_entries); + if (likely(fill_level < efx->txq_stop_thresh)) { + smp_mb(); + if (likely(!efx->loopback_selftest)) + netif_tx_start_queue(txq1->core_txq); + } +} + +#ifdef EFX_USE_PIO + +struct efx_short_copy_buffer { + int used; + u8 buf[L1_CACHE_BYTES]; +}; + +/* Copy in explicit 64-bit writes. */ +static void efx_memcpy_64(void __iomem *dest, void *src, size_t len) +{ + u64 *src64 = src; + u64 __iomem *dest64 = dest; + size_t l64 = len / 8; + size_t i; + + for (i = 0; i < l64; i++) + writeq(src64[i], &dest64[i]); +} + +/* Copy to PIO, respecting that writes to PIO buffers must be dword aligned. + * Advances piobuf pointer. Leaves additional data in the copy buffer. + */ +static void efx_memcpy_toio_aligned(struct efx_nic *efx, u8 __iomem **piobuf, + u8 *data, int len, + struct efx_short_copy_buffer *copy_buf) +{ + int block_len = len & ~(sizeof(copy_buf->buf) - 1); + + efx_memcpy_64(*piobuf, data, block_len); + *piobuf += block_len; + len -= block_len; + + if (len) { + data += block_len; + BUG_ON(copy_buf->used); + BUG_ON(len > sizeof(copy_buf->buf)); + memcpy(copy_buf->buf, data, len); + copy_buf->used = len; + } +} + +/* Copy to PIO, respecting dword alignment, popping data from copy buffer first. + * Advances piobuf pointer. Leaves additional data in the copy buffer. + */ +static void efx_memcpy_toio_aligned_cb(struct efx_nic *efx, u8 __iomem **piobuf, + u8 *data, int len, + struct efx_short_copy_buffer *copy_buf) +{ + if (copy_buf->used) { + /* if the copy buffer is partially full, fill it up and write */ + int copy_to_buf = + min_t(int, sizeof(copy_buf->buf) - copy_buf->used, len); + + memcpy(copy_buf->buf + copy_buf->used, data, copy_to_buf); + copy_buf->used += copy_to_buf; + + /* if we didn't fill it up then we're done for now */ + if (copy_buf->used < sizeof(copy_buf->buf)) + return; + + efx_memcpy_64(*piobuf, copy_buf->buf, sizeof(copy_buf->buf)); + *piobuf += sizeof(copy_buf->buf); + data += copy_to_buf; + len -= copy_to_buf; + copy_buf->used = 0; + } + + efx_memcpy_toio_aligned(efx, piobuf, data, len, copy_buf); +} + +static void efx_flush_copy_buffer(struct efx_nic *efx, u8 __iomem *piobuf, + struct efx_short_copy_buffer *copy_buf) +{ + /* if there's anything in it, write the whole buffer, including junk */ + if (copy_buf->used) + efx_memcpy_64(piobuf, copy_buf->buf, sizeof(copy_buf->buf)); +} + +/* Traverse skb structure and copy fragments in to PIO buffer. + * Advances piobuf pointer. + */ +static void efx_skb_copy_bits_to_pio(struct efx_nic *efx, struct sk_buff *skb, + u8 __iomem **piobuf, + struct efx_short_copy_buffer *copy_buf) +{ + int i; + + efx_memcpy_toio_aligned(efx, piobuf, skb->data, skb_headlen(skb), + copy_buf); + + for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) { + skb_frag_t *f = &skb_shinfo(skb)->frags[i]; + u8 *vaddr; + + vaddr = kmap_atomic(skb_frag_page(f)); + + efx_memcpy_toio_aligned_cb(efx, piobuf, vaddr + f->page_offset, + skb_frag_size(f), copy_buf); + kunmap_atomic(vaddr); + } + + EFX_BUG_ON_PARANOID(skb_shinfo(skb)->frag_list); +} + +static struct efx_tx_buffer * +efx_enqueue_skb_pio(struct efx_tx_queue *tx_queue, struct sk_buff *skb) +{ + struct efx_tx_buffer *buffer = + efx_tx_queue_get_insert_buffer(tx_queue); + u8 __iomem *piobuf = tx_queue->piobuf; + + /* Copy to PIO buffer. Ensure the writes are padded to the end + * of a cache line, as this is required for write-combining to be + * effective on at least x86. + */ + + if (skb_shinfo(skb)->nr_frags) { + /* The size of the copy buffer will ensure all writes + * are the size of a cache line. + */ + struct efx_short_copy_buffer copy_buf; + + copy_buf.used = 0; + + efx_skb_copy_bits_to_pio(tx_queue->efx, skb, + &piobuf, ©_buf); + efx_flush_copy_buffer(tx_queue->efx, piobuf, ©_buf); + } else { + /* Pad the write to the size of a cache line. + * We can do this because we know the skb_shared_info sruct is + * after the source, and the destination buffer is big enough. + */ + BUILD_BUG_ON(L1_CACHE_BYTES > + SKB_DATA_ALIGN(sizeof(struct skb_shared_info))); + efx_memcpy_64(tx_queue->piobuf, skb->data, + ALIGN(skb->len, L1_CACHE_BYTES)); + } + + EFX_POPULATE_QWORD_5(buffer->option, + ESF_DZ_TX_DESC_IS_OPT, 1, + ESF_DZ_TX_OPTION_TYPE, ESE_DZ_TX_OPTION_DESC_PIO, + ESF_DZ_TX_PIO_CONT, 0, + ESF_DZ_TX_PIO_BYTE_CNT, skb->len, + ESF_DZ_TX_PIO_BUF_ADDR, + tx_queue->piobuf_offset); + ++tx_queue->pio_packets; + ++tx_queue->insert_count; + return buffer; +} +#endif /* EFX_USE_PIO */ + /* * Add a socket buffer to a TX queue * @@ -132,21 +346,20 @@ efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr) * This function is split out from efx_hard_start_xmit to allow the * loopback test to direct packets via specific TX queues. * - * Returns NETDEV_TX_OK or NETDEV_TX_BUSY + * Returns NETDEV_TX_OK. * You must hold netif_tx_lock() to call this function. */ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) { struct efx_nic *efx = tx_queue->efx; - struct pci_dev *pci_dev = efx->pci_dev; + struct device *dma_dev = &efx->pci_dev->dev; struct efx_tx_buffer *buffer; skb_frag_t *fragment; - unsigned int len, unmap_len = 0, fill_level, insert_ptr; + unsigned int len, unmap_len = 0; dma_addr_t dma_addr, unmap_addr = 0; unsigned int dma_len; - bool unmap_single; - int q_space, i = 0; - netdev_tx_t rc = NETDEV_TX_OK; + unsigned short dma_flags; + int i = 0; EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count); @@ -164,20 +377,28 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) return NETDEV_TX_OK; } - fill_level = tx_queue->insert_count - tx_queue->old_read_count; - q_space = efx->txq_entries - 1 - fill_level; + /* Consider using PIO for short packets */ +#ifdef EFX_USE_PIO + if (skb->len <= efx_piobuf_size && tx_queue->piobuf && + efx_nic_tx_is_empty(tx_queue) && + efx_nic_tx_is_empty(efx_tx_queue_partner(tx_queue))) { + buffer = efx_enqueue_skb_pio(tx_queue, skb); + dma_flags = EFX_TX_BUF_OPTION; + goto finish_packet; + } +#endif - /* Map for DMA. Use pci_map_single rather than pci_map_page + /* Map for DMA. Use dma_map_single rather than dma_map_page * since this is more efficient on machines with sparse * memory. */ - unmap_single = true; - dma_addr = pci_map_single(pci_dev, skb->data, len, PCI_DMA_TODEVICE); + dma_flags = EFX_TX_BUF_MAP_SINGLE; + dma_addr = dma_map_single(dma_dev, skb->data, len, PCI_DMA_TODEVICE); /* Process all fragments */ while (1) { - if (unlikely(pci_dma_mapping_error(pci_dev, dma_addr))) - goto pci_err; + if (unlikely(dma_mapping_error(dma_dev, dma_addr))) + goto dma_err; /* Store fields for marking in the per-fragment final * descriptor */ @@ -186,40 +407,7 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) /* Add to TX queue, splitting across DMA boundaries */ do { - if (unlikely(q_space-- <= 0)) { - /* It might be that completions have - * happened since the xmit path last - * checked. Update the xmit path's - * copy of read_count. - */ - netif_tx_stop_queue(tx_queue->core_txq); - /* This memory barrier protects the - * change of queue state from the access - * of read_count. */ - smp_mb(); - tx_queue->old_read_count = - ACCESS_ONCE(tx_queue->read_count); - fill_level = (tx_queue->insert_count - - tx_queue->old_read_count); - q_space = efx->txq_entries - 1 - fill_level; - if (unlikely(q_space-- <= 0)) { - rc = NETDEV_TX_BUSY; - goto unwind; - } - smp_mb(); - if (likely(!efx->loopback_selftest)) - netif_tx_start_queue( - tx_queue->core_txq); - } - - insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask; - buffer = &tx_queue->buffer[insert_ptr]; - efx_tsoh_free(tx_queue, buffer); - EFX_BUG_ON_PARANOID(buffer->tsoh); - EFX_BUG_ON_PARANOID(buffer->skb); - EFX_BUG_ON_PARANOID(buffer->len); - EFX_BUG_ON_PARANOID(!buffer->continuation); - EFX_BUG_ON_PARANOID(buffer->unmap_len); + buffer = efx_tx_queue_get_insert_buffer(tx_queue); dma_len = efx_max_tx_len(efx, dma_addr); if (likely(dma_len >= len)) @@ -228,14 +416,16 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) /* Fill out per descriptor fields */ buffer->len = dma_len; buffer->dma_addr = dma_addr; + buffer->flags = EFX_TX_BUF_CONT; len -= dma_len; dma_addr += dma_len; ++tx_queue->insert_count; } while (len); /* Transfer ownership of the unmapping to the final buffer */ - buffer->unmap_single = unmap_single; + buffer->flags = EFX_TX_BUF_CONT | dma_flags; buffer->unmap_len = unmap_len; + buffer->dma_offset = buffer->dma_addr - unmap_addr; unmap_len = 0; /* Get address and size of next fragment */ @@ -245,23 +435,28 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) len = skb_frag_size(fragment); i++; /* Map for DMA */ - unmap_single = false; - dma_addr = skb_frag_dma_map(&pci_dev->dev, fragment, 0, len, + dma_flags = 0; + dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len, DMA_TO_DEVICE); } /* Transfer ownership of the skb to the final buffer */ +#ifdef EFX_USE_PIO +finish_packet: +#endif buffer->skb = skb; - buffer->continuation = false; + buffer->flags = EFX_TX_BUF_SKB | dma_flags; netdev_tx_sent_queue(tx_queue->core_txq, skb->len); /* Pass off to hardware */ efx_nic_push_buffers(tx_queue); + efx_tx_maybe_stop_queue(tx_queue); + return NETDEV_TX_OK; - pci_err: + dma_err: netif_err(efx, tx_err, efx->net_dev, " TX queue %d could not map skb with %d bytes %d " "fragments for DMA\n", tx_queue->queue, skb->len, @@ -270,28 +465,25 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) /* Mark the packet as transmitted, and free the SKB ourselves */ dev_kfree_skb_any(skb); - unwind: /* Work backwards until we hit the original insert pointer value */ while (tx_queue->insert_count != tx_queue->write_count) { unsigned int pkts_compl = 0, bytes_compl = 0; --tx_queue->insert_count; - insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask; - buffer = &tx_queue->buffer[insert_ptr]; + buffer = __efx_tx_queue_get_insert_buffer(tx_queue); efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl); - buffer->len = 0; } /* Free the fragment we were mid-way through pushing */ if (unmap_len) { - if (unmap_single) - pci_unmap_single(pci_dev, unmap_addr, unmap_len, - PCI_DMA_TODEVICE); + if (dma_flags & EFX_TX_BUF_MAP_SINGLE) + dma_unmap_single(dma_dev, unmap_addr, unmap_len, + DMA_TO_DEVICE); else - pci_unmap_page(pci_dev, unmap_addr, unmap_len, - PCI_DMA_TODEVICE); + dma_unmap_page(dma_dev, unmap_addr, unmap_len, + DMA_TO_DEVICE); } - return rc; + return NETDEV_TX_OK; } /* Remove packets from the TX queue @@ -312,7 +504,9 @@ static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue, while (read_ptr != stop_index) { struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr]; - if (unlikely(buffer->len == 0)) { + + if (!(buffer->flags & EFX_TX_BUF_OPTION) && + unlikely(buffer->len == 0)) { netif_err(efx, tx_err, efx->net_dev, "TX queue %d spurious TX completion id %x\n", tx_queue->queue, read_ptr); @@ -321,8 +515,6 @@ static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue, } efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl); - buffer->continuation = true; - buffer->len = 0; ++tx_queue->read_count; read_ptr = tx_queue->read_count & tx_queue->ptr_mask; @@ -339,7 +531,7 @@ static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue, * OS to free the skb. */ netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb, - struct net_device *net_dev) + struct net_device *net_dev) { struct efx_nic *efx = netdev_priv(net_dev); struct efx_tx_queue *tx_queue; @@ -347,6 +539,12 @@ netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb, EFX_WARN_ON_PARANOID(!netif_device_present(net_dev)); + /* PTP "event" packet */ + if (unlikely(efx_xmit_with_hwtstamp(skb)) && + unlikely(efx_ptp_is_ptp_tx(efx, skb))) { + return efx_ptp_tx(efx, skb); + } + index = skb_get_queue_mapping(skb); type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0; if (index >= efx->n_tx_channels) { @@ -431,6 +629,7 @@ void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index) { unsigned fill_level; struct efx_nic *efx = tx_queue->efx; + struct efx_tx_queue *txq2; unsigned int pkts_compl = 0, bytes_compl = 0; EFX_BUG_ON_PARANOID(index > tx_queue->ptr_mask); @@ -438,18 +637,22 @@ void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index) efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl); netdev_tx_completed_queue(tx_queue->core_txq, pkts_compl, bytes_compl); - /* See if we need to restart the netif queue. This barrier - * separates the update of read_count from the test of the - * queue state. */ + if (pkts_compl > 1) + ++tx_queue->merge_events; + + /* See if we need to restart the netif queue. This memory + * barrier ensures that we write read_count (inside + * efx_dequeue_buffers()) before reading the queue status. + */ smp_mb(); if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) && likely(efx->port_enabled) && likely(netif_device_present(efx->net_dev))) { - fill_level = tx_queue->insert_count - tx_queue->read_count; - if (fill_level < EFX_TXQ_THRESHOLD(efx)) { - EFX_BUG_ON_PARANOID(!efx_dev_registered(efx)); + txq2 = efx_tx_queue_partner(tx_queue); + fill_level = max(tx_queue->insert_count - tx_queue->read_count, + txq2->insert_count - txq2->read_count); + if (fill_level <= efx->txq_wake_thresh) netif_tx_wake_queue(tx_queue->core_txq); - } } /* Check whether the hardware queue is now empty */ @@ -463,11 +666,26 @@ void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index) } } +/* Size of page-based TSO header buffers. Larger blocks must be + * allocated from the heap. + */ +#define TSOH_STD_SIZE 128 +#define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE) + +/* At most half the descriptors in the queue at any time will refer to + * a TSO header buffer, since they must always be followed by a + * payload descriptor referring to an skb. + */ +static unsigned int efx_tsoh_page_count(struct efx_tx_queue *tx_queue) +{ + return DIV_ROUND_UP(tx_queue->ptr_mask + 1, 2 * TSOH_PER_PAGE); +} + int efx_probe_tx_queue(struct efx_tx_queue *tx_queue) { struct efx_nic *efx = tx_queue->efx; unsigned int entries; - int i, rc; + int rc; /* Create the smallest power-of-two aligned ring */ entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE); @@ -483,17 +701,28 @@ int efx_probe_tx_queue(struct efx_tx_queue *tx_queue) GFP_KERNEL); if (!tx_queue->buffer) return -ENOMEM; - for (i = 0; i <= tx_queue->ptr_mask; ++i) - tx_queue->buffer[i].continuation = true; + + if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD) { + tx_queue->tsoh_page = + kcalloc(efx_tsoh_page_count(tx_queue), + sizeof(tx_queue->tsoh_page[0]), GFP_KERNEL); + if (!tx_queue->tsoh_page) { + rc = -ENOMEM; + goto fail1; + } + } /* Allocate hardware ring */ rc = efx_nic_probe_tx(tx_queue); if (rc) - goto fail; + goto fail2; return 0; - fail: +fail2: + kfree(tx_queue->tsoh_page); + tx_queue->tsoh_page = NULL; +fail1: kfree(tx_queue->buffer); tx_queue->buffer = NULL; return rc; @@ -517,10 +746,13 @@ void efx_init_tx_queue(struct efx_tx_queue *tx_queue) tx_queue->initialised = true; } -void efx_release_tx_buffers(struct efx_tx_queue *tx_queue) +void efx_fini_tx_queue(struct efx_tx_queue *tx_queue) { struct efx_tx_buffer *buffer; + netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev, + "shutting down TX queue %d\n", tx_queue->queue); + if (!tx_queue->buffer) return; @@ -529,35 +761,16 @@ void efx_release_tx_buffers(struct efx_tx_queue *tx_queue) unsigned int pkts_compl = 0, bytes_compl = 0; buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask]; efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl); - buffer->continuation = true; - buffer->len = 0; ++tx_queue->read_count; } netdev_tx_reset_queue(tx_queue->core_txq); } -void efx_fini_tx_queue(struct efx_tx_queue *tx_queue) -{ - if (!tx_queue->initialised) - return; - - netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev, - "shutting down TX queue %d\n", tx_queue->queue); - - tx_queue->initialised = false; - - /* Flush TX queue, remove descriptor ring */ - efx_nic_fini_tx(tx_queue); - - efx_release_tx_buffers(tx_queue); - - /* Free up TSO header cache */ - efx_fini_tso(tx_queue); -} - void efx_remove_tx_queue(struct efx_tx_queue *tx_queue) { + int i; + if (!tx_queue->buffer) return; @@ -565,6 +778,14 @@ void efx_remove_tx_queue(struct efx_tx_queue *tx_queue) "destroying TX queue %d\n", tx_queue->queue); efx_nic_remove_tx(tx_queue); + if (tx_queue->tsoh_page) { + for (i = 0; i < efx_tsoh_page_count(tx_queue); i++) + efx_nic_free_buffer(tx_queue->efx, + &tx_queue->tsoh_page[i]); + kfree(tx_queue->tsoh_page); + tx_queue->tsoh_page = NULL; + } + kfree(tx_queue->buffer); tx_queue->buffer = NULL; } @@ -578,31 +799,7 @@ void efx_remove_tx_queue(struct efx_tx_queue *tx_queue) * Requires TX checksum offload support. */ -/* Number of bytes inserted at the start of a TSO header buffer, - * similar to NET_IP_ALIGN. - */ -#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS -#define TSOH_OFFSET 0 -#else -#define TSOH_OFFSET NET_IP_ALIGN -#endif - -#define TSOH_BUFFER(tsoh) ((u8 *)(tsoh + 1) + TSOH_OFFSET) - -/* Total size of struct efx_tso_header, buffer and padding */ -#define TSOH_SIZE(hdr_len) \ - (sizeof(struct efx_tso_header) + TSOH_OFFSET + hdr_len) - -/* Size of blocks on free list. Larger blocks must be allocated from - * the heap. - */ -#define TSOH_STD_SIZE 128 - #define PTR_DIFF(p1, p2) ((u8 *)(p1) - (u8 *)(p2)) -#define ETH_HDR_LEN(skb) (skb_network_header(skb) - (skb)->data) -#define SKB_TCP_OFF(skb) PTR_DIFF(tcp_hdr(skb), (skb)->data) -#define SKB_IPV4_OFF(skb) PTR_DIFF(ip_hdr(skb), (skb)->data) -#define SKB_IPV6_OFF(skb) PTR_DIFF(ipv6_hdr(skb), (skb)->data) /** * struct tso_state - TSO state for an SKB @@ -614,10 +811,15 @@ void efx_remove_tx_queue(struct efx_tx_queue *tx_queue) * @in_len: Remaining length in current SKB fragment * @unmap_len: Length of SKB fragment * @unmap_addr: DMA address of SKB fragment - * @unmap_single: DMA single vs page mapping flag + * @dma_flags: TX buffer flags for DMA mapping - %EFX_TX_BUF_MAP_SINGLE or 0 * @protocol: Network protocol (after any VLAN header) + * @ip_off: Offset of IP header + * @tcp_off: Offset of TCP header * @header_len: Number of bytes of header - * @full_packet_size: Number of bytes to put in each outgoing segment + * @ip_base_len: IPv4 tot_len or IPv6 payload_len, before TCP payload + * @header_dma_addr: Header DMA address, when using option descriptors + * @header_unmap_len: Header DMA mapped length, or 0 if not using option + * descriptors * * The state used during segmentation. It is put into this data structure * just to make it easy to pass into inline functions. @@ -626,7 +828,7 @@ struct tso_state { /* Output position */ unsigned out_len; unsigned seqnum; - unsigned ipv4_id; + u16 ipv4_id; unsigned packet_space; /* Input position */ @@ -634,11 +836,15 @@ struct tso_state { unsigned in_len; unsigned unmap_len; dma_addr_t unmap_addr; - bool unmap_single; + unsigned short dma_flags; __be16 protocol; + unsigned int ip_off; + unsigned int tcp_off; unsigned header_len; - int full_packet_size; + unsigned int ip_base_len; + dma_addr_t header_dma_addr; + unsigned int header_unmap_len; }; @@ -653,17 +859,8 @@ static __be16 efx_tso_check_protocol(struct sk_buff *skb) EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto != protocol); if (protocol == htons(ETH_P_8021Q)) { - /* Find the encapsulated protocol; reset network header - * and transport header based on that. */ struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data; protocol = veh->h_vlan_encapsulated_proto; - skb_set_network_header(skb, sizeof(*veh)); - if (protocol == htons(ETH_P_IP)) - skb_set_transport_header(skb, sizeof(*veh) + - 4 * ip_hdr(skb)->ihl); - else if (protocol == htons(ETH_P_IPV6)) - skb_set_transport_header(skb, sizeof(*veh) + - sizeof(struct ipv6hdr)); } if (protocol == htons(ETH_P_IP)) { @@ -679,92 +876,44 @@ static __be16 efx_tso_check_protocol(struct sk_buff *skb) return protocol; } - -/* - * Allocate a page worth of efx_tso_header structures, and string them - * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM. - */ -static int efx_tsoh_block_alloc(struct efx_tx_queue *tx_queue) +static u8 *efx_tsoh_get_buffer(struct efx_tx_queue *tx_queue, + struct efx_tx_buffer *buffer, unsigned int len) { + u8 *result; - struct pci_dev *pci_dev = tx_queue->efx->pci_dev; - struct efx_tso_header *tsoh; - dma_addr_t dma_addr; - u8 *base_kva, *kva; - - base_kva = pci_alloc_consistent(pci_dev, PAGE_SIZE, &dma_addr); - if (base_kva == NULL) { - netif_err(tx_queue->efx, tx_err, tx_queue->efx->net_dev, - "Unable to allocate page for TSO headers\n"); - return -ENOMEM; - } - - /* pci_alloc_consistent() allocates pages. */ - EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1u)); - - for (kva = base_kva; kva < base_kva + PAGE_SIZE; kva += TSOH_STD_SIZE) { - tsoh = (struct efx_tso_header *)kva; - tsoh->dma_addr = dma_addr + (TSOH_BUFFER(tsoh) - base_kva); - tsoh->next = tx_queue->tso_headers_free; - tx_queue->tso_headers_free = tsoh; - } - - return 0; -} - - -/* Free up a TSO header, and all others in the same page. */ -static void efx_tsoh_block_free(struct efx_tx_queue *tx_queue, - struct efx_tso_header *tsoh, - struct pci_dev *pci_dev) -{ - struct efx_tso_header **p; - unsigned long base_kva; - dma_addr_t base_dma; - - base_kva = (unsigned long)tsoh & PAGE_MASK; - base_dma = tsoh->dma_addr & PAGE_MASK; - - p = &tx_queue->tso_headers_free; - while (*p != NULL) { - if (((unsigned long)*p & PAGE_MASK) == base_kva) - *p = (*p)->next; - else - p = &(*p)->next; - } + EFX_BUG_ON_PARANOID(buffer->len); + EFX_BUG_ON_PARANOID(buffer->flags); + EFX_BUG_ON_PARANOID(buffer->unmap_len); - pci_free_consistent(pci_dev, PAGE_SIZE, (void *)base_kva, base_dma); -} + if (likely(len <= TSOH_STD_SIZE - NET_IP_ALIGN)) { + unsigned index = + (tx_queue->insert_count & tx_queue->ptr_mask) / 2; + struct efx_buffer *page_buf = + &tx_queue->tsoh_page[index / TSOH_PER_PAGE]; + unsigned offset = + TSOH_STD_SIZE * (index % TSOH_PER_PAGE) + NET_IP_ALIGN; + + if (unlikely(!page_buf->addr) && + efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE, + GFP_ATOMIC)) + return NULL; + + result = (u8 *)page_buf->addr + offset; + buffer->dma_addr = page_buf->dma_addr + offset; + buffer->flags = EFX_TX_BUF_CONT; + } else { + tx_queue->tso_long_headers++; -static struct efx_tso_header * -efx_tsoh_heap_alloc(struct efx_tx_queue *tx_queue, size_t header_len) -{ - struct efx_tso_header *tsoh; - - tsoh = kmalloc(TSOH_SIZE(header_len), GFP_ATOMIC | GFP_DMA); - if (unlikely(!tsoh)) - return NULL; - - tsoh->dma_addr = pci_map_single(tx_queue->efx->pci_dev, - TSOH_BUFFER(tsoh), header_len, - PCI_DMA_TODEVICE); - if (unlikely(pci_dma_mapping_error(tx_queue->efx->pci_dev, - tsoh->dma_addr))) { - kfree(tsoh); - return NULL; + buffer->heap_buf = kmalloc(NET_IP_ALIGN + len, GFP_ATOMIC); + if (unlikely(!buffer->heap_buf)) + return NULL; + result = (u8 *)buffer->heap_buf + NET_IP_ALIGN; + buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_HEAP; } - tsoh->unmap_len = header_len; - return tsoh; -} + buffer->len = len; -static void -efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, struct efx_tso_header *tsoh) -{ - pci_unmap_single(tx_queue->efx->pci_dev, - tsoh->dma_addr, tsoh->unmap_len, - PCI_DMA_TODEVICE); - kfree(tsoh); + return result; } /** @@ -774,62 +923,26 @@ efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, struct efx_tso_header *tsoh) * @len: Length of fragment * @final_buffer: The final buffer inserted into the queue * - * Push descriptors onto the TX queue. Return 0 on success or 1 if - * @tx_queue full. + * Push descriptors onto the TX queue. */ -static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue, - dma_addr_t dma_addr, unsigned len, - struct efx_tx_buffer **final_buffer) +static void efx_tx_queue_insert(struct efx_tx_queue *tx_queue, + dma_addr_t dma_addr, unsigned len, + struct efx_tx_buffer **final_buffer) { struct efx_tx_buffer *buffer; struct efx_nic *efx = tx_queue->efx; - unsigned dma_len, fill_level, insert_ptr; - int q_space; + unsigned dma_len; EFX_BUG_ON_PARANOID(len <= 0); - fill_level = tx_queue->insert_count - tx_queue->old_read_count; - /* -1 as there is no way to represent all descriptors used */ - q_space = efx->txq_entries - 1 - fill_level; - while (1) { - if (unlikely(q_space-- <= 0)) { - /* It might be that completions have happened - * since the xmit path last checked. Update - * the xmit path's copy of read_count. - */ - netif_tx_stop_queue(tx_queue->core_txq); - /* This memory barrier protects the change of - * queue state from the access of read_count. */ - smp_mb(); - tx_queue->old_read_count = - ACCESS_ONCE(tx_queue->read_count); - fill_level = (tx_queue->insert_count - - tx_queue->old_read_count); - q_space = efx->txq_entries - 1 - fill_level; - if (unlikely(q_space-- <= 0)) { - *final_buffer = NULL; - return 1; - } - smp_mb(); - netif_tx_start_queue(tx_queue->core_txq); - } - - insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask; - buffer = &tx_queue->buffer[insert_ptr]; + buffer = efx_tx_queue_get_insert_buffer(tx_queue); ++tx_queue->insert_count; EFX_BUG_ON_PARANOID(tx_queue->insert_count - tx_queue->read_count >= efx->txq_entries); - efx_tsoh_free(tx_queue, buffer); - EFX_BUG_ON_PARANOID(buffer->len); - EFX_BUG_ON_PARANOID(buffer->unmap_len); - EFX_BUG_ON_PARANOID(buffer->skb); - EFX_BUG_ON_PARANOID(!buffer->continuation); - EFX_BUG_ON_PARANOID(buffer->tsoh); - buffer->dma_addr = dma_addr; dma_len = efx_max_tx_len(efx, dma_addr); @@ -838,7 +951,8 @@ static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue, if (dma_len >= len) break; - buffer->len = dma_len; /* Don't set the other members */ + buffer->len = dma_len; + buffer->flags = EFX_TX_BUF_CONT; dma_addr += dma_len; len -= dma_len; } @@ -846,7 +960,6 @@ static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue, EFX_BUG_ON_PARANOID(!len); buffer->len = len; *final_buffer = buffer; - return 0; } @@ -857,82 +970,102 @@ static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue, * a single fragment, and we know it doesn't cross a page boundary. It * also allows us to not worry about end-of-packet etc. */ -static void efx_tso_put_header(struct efx_tx_queue *tx_queue, - struct efx_tso_header *tsoh, unsigned len) +static int efx_tso_put_header(struct efx_tx_queue *tx_queue, + struct efx_tx_buffer *buffer, u8 *header) { - struct efx_tx_buffer *buffer; - - buffer = &tx_queue->buffer[tx_queue->insert_count & tx_queue->ptr_mask]; - efx_tsoh_free(tx_queue, buffer); - EFX_BUG_ON_PARANOID(buffer->len); - EFX_BUG_ON_PARANOID(buffer->unmap_len); - EFX_BUG_ON_PARANOID(buffer->skb); - EFX_BUG_ON_PARANOID(!buffer->continuation); - EFX_BUG_ON_PARANOID(buffer->tsoh); - buffer->len = len; - buffer->dma_addr = tsoh->dma_addr; - buffer->tsoh = tsoh; + if (unlikely(buffer->flags & EFX_TX_BUF_HEAP)) { + buffer->dma_addr = dma_map_single(&tx_queue->efx->pci_dev->dev, + header, buffer->len, + DMA_TO_DEVICE); + if (unlikely(dma_mapping_error(&tx_queue->efx->pci_dev->dev, + buffer->dma_addr))) { + kfree(buffer->heap_buf); + buffer->len = 0; + buffer->flags = 0; + return -ENOMEM; + } + buffer->unmap_len = buffer->len; + buffer->dma_offset = 0; + buffer->flags |= EFX_TX_BUF_MAP_SINGLE; + } ++tx_queue->insert_count; + return 0; } -/* Remove descriptors put into a tx_queue. */ +/* Remove buffers put into a tx_queue. None of the buffers must have + * an skb attached. + */ static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue) { struct efx_tx_buffer *buffer; - dma_addr_t unmap_addr; /* Work backwards until we hit the original insert pointer value */ while (tx_queue->insert_count != tx_queue->write_count) { --tx_queue->insert_count; - buffer = &tx_queue->buffer[tx_queue->insert_count & - tx_queue->ptr_mask]; - efx_tsoh_free(tx_queue, buffer); - EFX_BUG_ON_PARANOID(buffer->skb); - if (buffer->unmap_len) { - unmap_addr = (buffer->dma_addr + buffer->len - - buffer->unmap_len); - if (buffer->unmap_single) - pci_unmap_single(tx_queue->efx->pci_dev, - unmap_addr, buffer->unmap_len, - PCI_DMA_TODEVICE); - else - pci_unmap_page(tx_queue->efx->pci_dev, - unmap_addr, buffer->unmap_len, - PCI_DMA_TODEVICE); - buffer->unmap_len = 0; - } - buffer->len = 0; - buffer->continuation = true; + buffer = __efx_tx_queue_get_insert_buffer(tx_queue); + efx_dequeue_buffer(tx_queue, buffer, NULL, NULL); } } /* Parse the SKB header and initialise state. */ -static void tso_start(struct tso_state *st, const struct sk_buff *skb) +static int tso_start(struct tso_state *st, struct efx_nic *efx, + const struct sk_buff *skb) { - /* All ethernet/IP/TCP headers combined size is TCP header size - * plus offset of TCP header relative to start of packet. - */ - st->header_len = ((tcp_hdr(skb)->doff << 2u) - + PTR_DIFF(tcp_hdr(skb), skb->data)); - st->full_packet_size = st->header_len + skb_shinfo(skb)->gso_size; + bool use_opt_desc = efx_nic_rev(efx) >= EFX_REV_HUNT_A0; + struct device *dma_dev = &efx->pci_dev->dev; + unsigned int header_len, in_len; + dma_addr_t dma_addr; - if (st->protocol == htons(ETH_P_IP)) + st->ip_off = skb_network_header(skb) - skb->data; + st->tcp_off = skb_transport_header(skb) - skb->data; + header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u); + in_len = skb_headlen(skb) - header_len; + st->header_len = header_len; + st->in_len = in_len; + if (st->protocol == htons(ETH_P_IP)) { + st->ip_base_len = st->header_len - st->ip_off; st->ipv4_id = ntohs(ip_hdr(skb)->id); - else + } else { + st->ip_base_len = st->header_len - st->tcp_off; st->ipv4_id = 0; + } st->seqnum = ntohl(tcp_hdr(skb)->seq); EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg); EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn); EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst); - st->packet_space = st->full_packet_size; - st->out_len = skb->len - st->header_len; - st->unmap_len = 0; - st->unmap_single = false; + st->out_len = skb->len - header_len; + + if (!use_opt_desc) { + st->header_unmap_len = 0; + + if (likely(in_len == 0)) { + st->dma_flags = 0; + st->unmap_len = 0; + return 0; + } + + dma_addr = dma_map_single(dma_dev, skb->data + header_len, + in_len, DMA_TO_DEVICE); + st->dma_flags = EFX_TX_BUF_MAP_SINGLE; + st->dma_addr = dma_addr; + st->unmap_addr = dma_addr; + st->unmap_len = in_len; + } else { + dma_addr = dma_map_single(dma_dev, skb->data, + skb_headlen(skb), DMA_TO_DEVICE); + st->header_dma_addr = dma_addr; + st->header_unmap_len = skb_headlen(skb); + st->dma_flags = 0; + st->dma_addr = dma_addr + header_len; + st->unmap_len = 0; + } + + return unlikely(dma_mapping_error(dma_dev, dma_addr)) ? -ENOMEM : 0; } static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx, @@ -941,7 +1074,7 @@ static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx, st->unmap_addr = skb_frag_dma_map(&efx->pci_dev->dev, frag, 0, skb_frag_size(frag), DMA_TO_DEVICE); if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) { - st->unmap_single = false; + st->dma_flags = 0; st->unmap_len = skb_frag_size(frag); st->in_len = skb_frag_size(frag); st->dma_addr = st->unmap_addr; @@ -950,24 +1083,6 @@ static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx, return -ENOMEM; } -static int tso_get_head_fragment(struct tso_state *st, struct efx_nic *efx, - const struct sk_buff *skb) -{ - int hl = st->header_len; - int len = skb_headlen(skb) - hl; - - st->unmap_addr = pci_map_single(efx->pci_dev, skb->data + hl, - len, PCI_DMA_TODEVICE); - if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) { - st->unmap_single = true; - st->unmap_len = len; - st->in_len = len; - st->dma_addr = st->unmap_addr; - return 0; - } - return -ENOMEM; -} - /** * tso_fill_packet_with_fragment - form descriptors for the current fragment @@ -976,20 +1091,19 @@ static int tso_get_head_fragment(struct tso_state *st, struct efx_nic *efx, * @st: TSO state * * Form descriptors for the current fragment, until we reach the end - * of fragment or end-of-packet. Return 0 on success, 1 if not enough - * space in @tx_queue. + * of fragment or end-of-packet. */ -static int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue, - const struct sk_buff *skb, - struct tso_state *st) +static void tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue, + const struct sk_buff *skb, + struct tso_state *st) { struct efx_tx_buffer *buffer; - int n, end_of_packet, rc; + int n; if (st->in_len == 0) - return 0; + return; if (st->packet_space == 0) - return 0; + return; EFX_BUG_ON_PARANOID(st->in_len <= 0); EFX_BUG_ON_PARANOID(st->packet_space <= 0); @@ -1000,25 +1114,25 @@ static int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue, st->out_len -= n; st->in_len -= n; - rc = efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer); - if (likely(rc == 0)) { - if (st->out_len == 0) - /* Transfer ownership of the skb */ - buffer->skb = skb; + efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer); - end_of_packet = st->out_len == 0 || st->packet_space == 0; - buffer->continuation = !end_of_packet; + if (st->out_len == 0) { + /* Transfer ownership of the skb */ + buffer->skb = skb; + buffer->flags = EFX_TX_BUF_SKB; + } else if (st->packet_space != 0) { + buffer->flags = EFX_TX_BUF_CONT; + } - if (st->in_len == 0) { - /* Transfer ownership of the pci mapping */ - buffer->unmap_len = st->unmap_len; - buffer->unmap_single = st->unmap_single; - st->unmap_len = 0; - } + if (st->in_len == 0) { + /* Transfer ownership of the DMA mapping */ + buffer->unmap_len = st->unmap_len; + buffer->dma_offset = buffer->unmap_len - buffer->len; + buffer->flags |= st->dma_flags; + st->unmap_len = 0; } st->dma_addr += n; - return rc; } @@ -1029,75 +1143,107 @@ static int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue, * @st: TSO state * * Generate a new header and prepare for the new packet. Return 0 on - * success, or -1 if failed to alloc header. + * success, or -%ENOMEM if failed to alloc header. */ static int tso_start_new_packet(struct efx_tx_queue *tx_queue, const struct sk_buff *skb, struct tso_state *st) { - struct efx_tso_header *tsoh; - struct tcphdr *tsoh_th; - unsigned ip_length; - u8 *header; - - /* Allocate a DMA-mapped header buffer. */ - if (likely(TSOH_SIZE(st->header_len) <= TSOH_STD_SIZE)) { - if (tx_queue->tso_headers_free == NULL) { - if (efx_tsoh_block_alloc(tx_queue)) - return -1; - } - EFX_BUG_ON_PARANOID(!tx_queue->tso_headers_free); - tsoh = tx_queue->tso_headers_free; - tx_queue->tso_headers_free = tsoh->next; - tsoh->unmap_len = 0; + struct efx_tx_buffer *buffer = + efx_tx_queue_get_insert_buffer(tx_queue); + bool is_last = st->out_len <= skb_shinfo(skb)->gso_size; + u8 tcp_flags_clear; + + if (!is_last) { + st->packet_space = skb_shinfo(skb)->gso_size; + tcp_flags_clear = 0x09; /* mask out FIN and PSH */ } else { - tx_queue->tso_long_headers++; - tsoh = efx_tsoh_heap_alloc(tx_queue, st->header_len); - if (unlikely(!tsoh)) - return -1; + st->packet_space = st->out_len; + tcp_flags_clear = 0x00; } - header = TSOH_BUFFER(tsoh); - tsoh_th = (struct tcphdr *)(header + SKB_TCP_OFF(skb)); + if (!st->header_unmap_len) { + /* Allocate and insert a DMA-mapped header buffer. */ + struct tcphdr *tsoh_th; + unsigned ip_length; + u8 *header; + int rc; - /* Copy and update the headers. */ - memcpy(header, skb->data, st->header_len); + header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len); + if (!header) + return -ENOMEM; - tsoh_th->seq = htonl(st->seqnum); - st->seqnum += skb_shinfo(skb)->gso_size; - if (st->out_len > skb_shinfo(skb)->gso_size) { - /* This packet will not finish the TSO burst. */ - ip_length = st->full_packet_size - ETH_HDR_LEN(skb); - tsoh_th->fin = 0; - tsoh_th->psh = 0; - } else { - /* This packet will be the last in the TSO burst. */ - ip_length = st->header_len - ETH_HDR_LEN(skb) + st->out_len; - tsoh_th->fin = tcp_hdr(skb)->fin; - tsoh_th->psh = tcp_hdr(skb)->psh; - } + tsoh_th = (struct tcphdr *)(header + st->tcp_off); - if (st->protocol == htons(ETH_P_IP)) { - struct iphdr *tsoh_iph = - (struct iphdr *)(header + SKB_IPV4_OFF(skb)); + /* Copy and update the headers. */ + memcpy(header, skb->data, st->header_len); + + tsoh_th->seq = htonl(st->seqnum); + ((u8 *)tsoh_th)[13] &= ~tcp_flags_clear; - tsoh_iph->tot_len = htons(ip_length); + ip_length = st->ip_base_len + st->packet_space; - /* Linux leaves suitable gaps in the IP ID space for us to fill. */ - tsoh_iph->id = htons(st->ipv4_id); - st->ipv4_id++; + if (st->protocol == htons(ETH_P_IP)) { + struct iphdr *tsoh_iph = + (struct iphdr *)(header + st->ip_off); + + tsoh_iph->tot_len = htons(ip_length); + tsoh_iph->id = htons(st->ipv4_id); + } else { + struct ipv6hdr *tsoh_iph = + (struct ipv6hdr *)(header + st->ip_off); + + tsoh_iph->payload_len = htons(ip_length); + } + + rc = efx_tso_put_header(tx_queue, buffer, header); + if (unlikely(rc)) + return rc; } else { - struct ipv6hdr *tsoh_iph = - (struct ipv6hdr *)(header + SKB_IPV6_OFF(skb)); + /* Send the original headers with a TSO option descriptor + * in front + */ + u8 tcp_flags = ((u8 *)tcp_hdr(skb))[13] & ~tcp_flags_clear; - tsoh_iph->payload_len = htons(ip_length - sizeof(*tsoh_iph)); + buffer->flags = EFX_TX_BUF_OPTION; + buffer->len = 0; + buffer->unmap_len = 0; + EFX_POPULATE_QWORD_5(buffer->option, + ESF_DZ_TX_DESC_IS_OPT, 1, + ESF_DZ_TX_OPTION_TYPE, + ESE_DZ_TX_OPTION_DESC_TSO, + ESF_DZ_TX_TSO_TCP_FLAGS, tcp_flags, + ESF_DZ_TX_TSO_IP_ID, st->ipv4_id, + ESF_DZ_TX_TSO_TCP_SEQNO, st->seqnum); + ++tx_queue->insert_count; + + /* We mapped the headers in tso_start(). Unmap them + * when the last segment is completed. + */ + buffer = efx_tx_queue_get_insert_buffer(tx_queue); + buffer->dma_addr = st->header_dma_addr; + buffer->len = st->header_len; + if (is_last) { + buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_MAP_SINGLE; + buffer->unmap_len = st->header_unmap_len; + buffer->dma_offset = 0; + /* Ensure we only unmap them once in case of a + * later DMA mapping error and rollback + */ + st->header_unmap_len = 0; + } else { + buffer->flags = EFX_TX_BUF_CONT; + buffer->unmap_len = 0; + } + ++tx_queue->insert_count; } - st->packet_space = skb_shinfo(skb)->gso_size; - ++tx_queue->tso_packets; + st->seqnum += skb_shinfo(skb)->gso_size; + + /* Linux leaves suitable gaps in the IP ID space for us to fill. */ + ++st->ipv4_id; - /* Form a descriptor for this header. */ - efx_tso_put_header(tx_queue, tsoh, st->header_len); + ++tx_queue->tso_packets; return 0; } @@ -1112,13 +1258,13 @@ static int tso_start_new_packet(struct efx_tx_queue *tx_queue, * * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if * @skb was not enqueued. In all cases @skb is consumed. Return - * %NETDEV_TX_OK or %NETDEV_TX_BUSY. + * %NETDEV_TX_OK. */ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue, struct sk_buff *skb) { struct efx_nic *efx = tx_queue->efx; - int frag_i, rc, rc2 = NETDEV_TX_OK; + int frag_i, rc; struct tso_state state; /* Find the packet protocol and sanity-check it */ @@ -1126,12 +1272,11 @@ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue, EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count); - tso_start(&state, skb); + rc = tso_start(&state, efx, skb); + if (rc) + goto mem_err; - /* Assume that skb header area contains exactly the headers, and - * all payload is in the frag list. - */ - if (skb_headlen(skb) == state.header_len) { + if (likely(state.in_len == 0)) { /* Grab the first payload fragment. */ EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1); frag_i = 0; @@ -1140,9 +1285,7 @@ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue, if (rc) goto mem_err; } else { - rc = tso_get_head_fragment(&state, efx, skb); - if (rc) - goto mem_err; + /* Payload starts in the header area. */ frag_i = -1; } @@ -1150,11 +1293,7 @@ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue, goto mem_err; while (1) { - rc = tso_fill_packet_with_fragment(tx_queue, skb, &state); - if (unlikely(rc)) { - rc2 = NETDEV_TX_BUSY; - goto unwind; - } + tso_fill_packet_with_fragment(tx_queue, skb, &state); /* Move onto the next fragment? */ if (state.in_len == 0) { @@ -1178,45 +1317,31 @@ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue, /* Pass off to hardware */ efx_nic_push_buffers(tx_queue); + efx_tx_maybe_stop_queue(tx_queue); + tx_queue->tso_bursts++; return NETDEV_TX_OK; mem_err: netif_err(efx, tx_err, efx->net_dev, - "Out of memory for TSO headers, or PCI mapping error\n"); + "Out of memory for TSO headers, or DMA mapping error\n"); dev_kfree_skb_any(skb); - unwind: /* Free the DMA mapping we were in the process of writing out */ if (state.unmap_len) { - if (state.unmap_single) - pci_unmap_single(efx->pci_dev, state.unmap_addr, - state.unmap_len, PCI_DMA_TODEVICE); + if (state.dma_flags & EFX_TX_BUF_MAP_SINGLE) + dma_unmap_single(&efx->pci_dev->dev, state.unmap_addr, + state.unmap_len, DMA_TO_DEVICE); else - pci_unmap_page(efx->pci_dev, state.unmap_addr, - state.unmap_len, PCI_DMA_TODEVICE); + dma_unmap_page(&efx->pci_dev->dev, state.unmap_addr, + state.unmap_len, DMA_TO_DEVICE); } - efx_enqueue_unwind(tx_queue); - return rc2; -} - + /* Free the header DMA mapping, if using option descriptors */ + if (state.header_unmap_len) + dma_unmap_single(&efx->pci_dev->dev, state.header_dma_addr, + state.header_unmap_len, DMA_TO_DEVICE); -/* - * Free up all TSO datastructures associated with tx_queue. This - * routine should be called only once the tx_queue is both empty and - * will no longer be used. - */ -static void efx_fini_tso(struct efx_tx_queue *tx_queue) -{ - unsigned i; - - if (tx_queue->buffer) { - for (i = 0; i <= tx_queue->ptr_mask; ++i) - efx_tsoh_free(tx_queue, &tx_queue->buffer[i]); - } - - while (tx_queue->tso_headers_free != NULL) - efx_tsoh_block_free(tx_queue, tx_queue->tso_headers_free, - tx_queue->efx->pci_dev); + efx_enqueue_unwind(tx_queue); + return NETDEV_TX_OK; } |