#include "nic.h"
#include "workarounds.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)
-
static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
struct efx_tx_buffer *buffer,
unsigned int *pkts_compl,
struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
dma_addr_t unmap_addr = (buffer->dma_addr + buffer->len -
buffer->unmap_len);
- if (buffer->unmap_single)
+ if (buffer->flags & EFX_TX_BUF_MAP_SINGLE)
dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
DMA_TO_DEVICE);
else
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);
}
+
+ buffer->flags &= EFX_TX_BUF_TSOH;
}
/**
static void efx_tsoh_free(struct efx_tx_queue *tx_queue,
struct efx_tx_buffer *buffer)
{
- if (buffer->tsoh) {
+ 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->tsoh = NULL;
+ buffer->flags &= ~EFX_TX_BUF_TSOH;
}
}
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 */
+ if (EFX_WORKAROUND_5391(efx))
+ 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);
+ }
+}
+
/*
* Add a socket buffer to a TX queue
*
* 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 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, insert_ptr;
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);
return NETDEV_TX_OK;
}
- fill_level = tx_queue->insert_count - tx_queue->old_read_count;
- q_space = efx->txq_entries - 1 - fill_level;
-
/* 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_flags = EFX_TX_BUF_MAP_SINGLE;
dma_addr = dma_map_single(dma_dev, skb->data, len, PCI_DMA_TODEVICE);
/* Process all fragments */
/* 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->flags);
EFX_BUG_ON_PARANOID(buffer->len);
- EFX_BUG_ON_PARANOID(!buffer->continuation);
EFX_BUG_ON_PARANOID(buffer->unmap_len);
dma_len = efx_max_tx_len(efx, dma_addr);
/* 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;
unmap_len = 0;
len = skb_frag_size(fragment);
i++;
/* Map for DMA */
- unmap_single = false;
+ 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 */
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;
dma_err:
/* 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;
/* Free the fragment we were mid-way through pushing */
if (unmap_len) {
- if (unmap_single)
+ if (dma_flags & EFX_TX_BUF_MAP_SINGLE)
dma_unmap_single(dma_dev, unmap_addr, unmap_len,
DMA_TO_DEVICE);
else
DMA_TO_DEVICE);
}
- return rc;
+ return NETDEV_TX_OK;
}
/* Remove packets from the TX queue
}
efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
- buffer->continuation = true;
buffer->len = 0;
++tx_queue->read_count;
{
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);
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. */
+ /* 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))
+ 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);
}
{
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);
GFP_KERNEL);
if (!tx_queue->buffer)
return -ENOMEM;
- for (i = 0; i <= tx_queue->ptr_mask; ++i)
- tx_queue->buffer[i].continuation = true;
/* Allocate hardware ring */
rc = efx_nic_probe_tx(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;
* @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)
* @header_len: Number of bytes of header
* @full_packet_size: Number of bytes to put in each outgoing segment
unsigned in_len;
unsigned unmap_len;
dma_addr_t unmap_addr;
- bool unmap_single;
+ unsigned short dma_flags;
__be16 protocol;
unsigned header_len;
* @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, insert_ptr;
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];
++tx_queue->insert_count;
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);
+ EFX_BUG_ON_PARANOID(buffer->flags);
buffer->dma_addr = dma_addr;
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;
}
EFX_BUG_ON_PARANOID(!len);
buffer->len = len;
*final_buffer = buffer;
- return 0;
}
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);
+ 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;
++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);
+ 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->unmap_single)
+ 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);
buffer->unmap_len = 0;
}
buffer->len = 0;
- buffer->continuation = true;
+ buffer->flags = 0;
}
}
st->out_len = skb->len - st->header_len;
st->unmap_len = 0;
- st->unmap_single = false;
+ st->dma_flags = 0;
}
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;
st->unmap_addr = dma_map_single(&efx->pci_dev->dev, skb->data + hl,
len, DMA_TO_DEVICE);
if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) {
- st->unmap_single = true;
+ st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
st->unmap_len = len;
st->in_len = len;
st->dma_addr = st->unmap_addr;
* @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);
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 DMA 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->flags |= st->dma_flags;
+ st->unmap_len = 0;
}
st->dma_addr += n;
- return rc;
}
*
* 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 */
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) {
/* 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;
"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)
+ 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
}
efx_enqueue_unwind(tx_queue);
- return rc2;
+ return NETDEV_TX_OK;
}
projects / cascardo / linux.git / blobdiff