diff options
author | Ben Hutchings <bhutchings@solarflare.com> | 2012-05-17 18:40:54 +0100 |
---|---|---|
committer | Ben Hutchings <bhutchings@solarflare.com> | 2012-08-24 20:10:11 +0100 |
commit | f7251a9ce936f1006fbfdef63dbe42ae5e0fee7c (patch) | |
tree | e0187036af8587b61f4f492a31f3429b33dc0ffb /drivers/net/ethernet/sfc | |
parent | 14bf718fb97efe9ff649c317e7d87a3617b13e7c (diff) | |
download | blackbird-op-linux-f7251a9ce936f1006fbfdef63dbe42ae5e0fee7c.tar.gz blackbird-op-linux-f7251a9ce936f1006fbfdef63dbe42ae5e0fee7c.zip |
sfc: Simplify TSO header buffer allocation
TSO header buffers contain a control structure immediately followed by
the packet headers, and are kept on a free list when not in use. This
complicates buffer management and tends to result in cache read misses
when we recycle such buffers (particularly if DMA-coherent memory
requires caches to be disabled).
Replace the free list with a simple mapping by descriptor index. We
know that there is always a payload descriptor between any two
descriptors with TSO header buffers, so we can allocate only one
such buffer for each two descriptors.
While we're at it, use a standard error code for allocation failure,
not -1.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Diffstat (limited to 'drivers/net/ethernet/sfc')
-rw-r--r-- | drivers/net/ethernet/sfc/net_driver.h | 13 | ||||
-rw-r--r-- | drivers/net/ethernet/sfc/nic.c | 2 | ||||
-rw-r--r-- | drivers/net/ethernet/sfc/tx.c | 315 |
3 files changed, 112 insertions, 218 deletions
diff --git a/drivers/net/ethernet/sfc/net_driver.h b/drivers/net/ethernet/sfc/net_driver.h index 28a6d6258692..a4fe9a786ef8 100644 --- a/drivers/net/ethernet/sfc/net_driver.h +++ b/drivers/net/ethernet/sfc/net_driver.h @@ -94,7 +94,8 @@ struct efx_special_buffer { * struct efx_tx_buffer - buffer state for a TX descriptor * @skb: When @flags & %EFX_TX_BUF_SKB, the associated socket buffer to be * freed when descriptor completes - * @tsoh: When @flags & %EFX_TX_BUF_TSOH, the associated TSO header structure. + * @heap_buf: When @flags & %EFX_TX_BUF_HEAP, the associated heap buffer to be + * freed when descriptor completes. * @dma_addr: DMA address of the fragment. * @flags: Flags for allocation and DMA mapping type * @len: Length of this fragment. @@ -104,7 +105,7 @@ struct efx_special_buffer { struct efx_tx_buffer { union { const struct sk_buff *skb; - struct efx_tso_header *tsoh; + void *heap_buf; }; dma_addr_t dma_addr; unsigned short flags; @@ -113,7 +114,7 @@ struct efx_tx_buffer { }; #define EFX_TX_BUF_CONT 1 /* not last descriptor of packet */ #define EFX_TX_BUF_SKB 2 /* buffer is last part of skb */ -#define EFX_TX_BUF_TSOH 4 /* buffer is TSO header */ +#define EFX_TX_BUF_HEAP 4 /* buffer was allocated with kmalloc() */ #define EFX_TX_BUF_MAP_SINGLE 8 /* buffer was mapped with dma_map_single() */ /** @@ -134,6 +135,7 @@ struct efx_tx_buffer { * @channel: The associated channel * @core_txq: The networking core TX queue structure * @buffer: The software buffer ring + * @tsoh_page: Array of pages of TSO header buffers * @txd: The hardware descriptor ring * @ptr_mask: The size of the ring minus 1. * @initialised: Has hardware queue been initialised? @@ -157,9 +159,6 @@ struct efx_tx_buffer { * variable indicates that the queue is full. This is to * avoid cache-line ping-pong between the xmit path and the * completion path. - * @tso_headers_free: A list of TSO headers allocated for this TX queue - * that are not in use, and so available for new TSO sends. The list - * is protected by the TX queue lock. * @tso_bursts: Number of times TSO xmit invoked by kernel * @tso_long_headers: Number of packets with headers too long for standard * blocks @@ -176,6 +175,7 @@ struct efx_tx_queue { struct efx_channel *channel; struct netdev_queue *core_txq; struct efx_tx_buffer *buffer; + struct efx_buffer *tsoh_page; struct efx_special_buffer txd; unsigned int ptr_mask; bool initialised; @@ -188,7 +188,6 @@ struct efx_tx_queue { unsigned int insert_count ____cacheline_aligned_in_smp; unsigned int write_count; unsigned int old_read_count; - struct efx_tso_header *tso_headers_free; unsigned int tso_bursts; unsigned int tso_long_headers; unsigned int tso_packets; diff --git a/drivers/net/ethernet/sfc/nic.c b/drivers/net/ethernet/sfc/nic.c index aa113709831d..cdff40b65729 100644 --- a/drivers/net/ethernet/sfc/nic.c +++ b/drivers/net/ethernet/sfc/nic.c @@ -298,7 +298,7 @@ efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer) /************************************************************************** * * Generic buffer handling - * These buffers are used for interrupt status and MAC stats + * These buffers are used for interrupt status, MAC stats, etc. * **************************************************************************/ diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c index 330d9111a339..61bc0ed718e3 100644 --- a/drivers/net/ethernet/sfc/tx.c +++ b/drivers/net/ethernet/sfc/tx.c @@ -47,51 +47,16 @@ static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue, 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); } - buffer->flags &= EFX_TX_BUF_TSOH; + buffer->len = 0; + buffer->flags = 0; } -/** - * 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; -}; - 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->flags & EFX_TX_BUF_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->flags &= ~EFX_TX_BUF_TSOH; - } -} - static inline unsigned efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr) @@ -245,7 +210,6 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) do { 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->flags); EFX_BUG_ON_PARANOID(buffer->len); EFX_BUG_ON_PARANOID(buffer->unmap_len); @@ -309,7 +273,6 @@ netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask; buffer = &tx_queue->buffer[insert_ptr]; efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl); - buffer->len = 0; } /* Free the fragment we were mid-way through pushing */ @@ -352,7 +315,6 @@ static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue, } efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl); - buffer->len = 0; ++tx_queue->read_count; read_ptr = tx_queue->read_count & tx_queue->ptr_mask; @@ -495,6 +457,21 @@ 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; @@ -516,14 +493,27 @@ int efx_probe_tx_queue(struct efx_tx_queue *tx_queue) if (!tx_queue->buffer) return -ENOMEM; + 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; @@ -559,7 +549,6 @@ 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->len = 0; ++tx_queue->read_count; } @@ -580,13 +569,12 @@ void efx_fini_tx_queue(struct efx_tx_queue *tx_queue) 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; @@ -594,6 +582,14 @@ 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; } @@ -616,17 +612,6 @@ void efx_remove_tx_queue(struct efx_tx_queue *tx_queue) #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) @@ -699,91 +684,43 @@ 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) { - struct device *dma_dev = &tx_queue->efx->pci_dev->dev; - struct efx_tso_header *tsoh; - dma_addr_t dma_addr; - u8 *base_kva, *kva; + u8 *result; - base_kva = dma_alloc_coherent(dma_dev, PAGE_SIZE, &dma_addr, GFP_ATOMIC); - 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; - } - - /* dma_alloc_coherent() 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 device *dma_dev) -{ - struct efx_tso_header **p; - unsigned long base_kva; - dma_addr_t base_dma; + EFX_BUG_ON_PARANOID(buffer->len); + EFX_BUG_ON_PARANOID(buffer->flags); + EFX_BUG_ON_PARANOID(buffer->unmap_len); - base_kva = (unsigned long)tsoh & PAGE_MASK; - base_dma = tsoh->dma_addr & PAGE_MASK; + if (likely(len <= TSOH_STD_SIZE - TSOH_OFFSET)) { + 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) + TSOH_OFFSET; - p = &tx_queue->tso_headers_free; - while (*p != NULL) { - if (((unsigned long)*p & PAGE_MASK) == base_kva) - *p = (*p)->next; - else - p = &(*p)->next; - } + if (unlikely(!page_buf->addr) && + efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE)) + return NULL; - dma_free_coherent(dma_dev, PAGE_SIZE, (void *)base_kva, base_dma); -} + 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 = dma_map_single(&tx_queue->efx->pci_dev->dev, - TSOH_BUFFER(tsoh), header_len, - DMA_TO_DEVICE); - if (unlikely(dma_mapping_error(&tx_queue->efx->pci_dev->dev, - tsoh->dma_addr))) { - kfree(tsoh); - return NULL; + buffer->heap_buf = kmalloc(TSOH_OFFSET + len, GFP_ATOMIC); + if (unlikely(!buffer->heap_buf)) + return NULL; + result = (u8 *)buffer->heap_buf + TSOH_OFFSET; + 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) -{ - dma_unmap_single(&tx_queue->efx->pci_dev->dev, - tsoh->dma_addr, tsoh->unmap_len, - DMA_TO_DEVICE); - kfree(tsoh); + return result; } /** @@ -814,7 +751,6 @@ static void efx_tx_queue_insert(struct efx_tx_queue *tx_queue, 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->flags); @@ -846,53 +782,42 @@ static void 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->flags); - buffer->len = len; - buffer->dma_addr = tsoh->dma_addr; - buffer->tsoh = tsoh; - buffer->flags = EFX_TX_BUF_TSOH | EFX_TX_BUF_CONT; + 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->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->flags & EFX_TX_BUF_SKB); - if (buffer->unmap_len) { - unmap_addr = (buffer->dma_addr + buffer->len - - buffer->unmap_len); - if (buffer->flags & EFX_TX_BUF_MAP_SINGLE) - dma_unmap_single(&tx_queue->efx->pci_dev->dev, - unmap_addr, buffer->unmap_len, - DMA_TO_DEVICE); - else - dma_unmap_page(&tx_queue->efx->pci_dev->dev, - unmap_addr, buffer->unmap_len, - DMA_TO_DEVICE); - buffer->unmap_len = 0; - } - buffer->len = 0; - buffer->flags = 0; + efx_dequeue_buffer(tx_queue, buffer, NULL, NULL); } } @@ -1014,35 +939,24 @@ static void 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 efx_tx_buffer *buffer = + &tx_queue->buffer[tx_queue->insert_count & tx_queue->ptr_mask]; struct tcphdr *tsoh_th; unsigned ip_length; u8 *header; + int rc; - /* 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; - } else { - tx_queue->tso_long_headers++; - tsoh = efx_tsoh_heap_alloc(tx_queue, st->header_len); - if (unlikely(!tsoh)) - return -1; - } + /* Allocate and insert a DMA-mapped header buffer. */ + header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len); + if (!header) + return -ENOMEM; - header = TSOH_BUFFER(tsoh); tsoh_th = (struct tcphdr *)(header + SKB_TCP_OFF(skb)); /* Copy and update the headers. */ @@ -1078,12 +992,13 @@ static int tso_start_new_packet(struct efx_tx_queue *tx_queue, tsoh_iph->payload_len = htons(ip_length - sizeof(*tsoh_iph)); } + rc = efx_tso_put_header(tx_queue, buffer, header); + if (unlikely(rc)) + return rc; + st->packet_space = skb_shinfo(skb)->gso_size; ++tx_queue->tso_packets; - /* Form a descriptor for this header. */ - efx_tso_put_header(tx_queue, tsoh, st->header_len); - return 0; } @@ -1182,23 +1097,3 @@ static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue, efx_enqueue_unwind(tx_queue); return NETDEV_TX_OK; } - - -/* - * 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->dev); -} |