Merge branch 'for-davem' of git://git.kernel.org/pub/scm/linux/kernel/git/bwh/sfc-next
Ben Hutchings says:
====================
1. Change the TX path to stop queues earlier and avoid returning
NETDEV_TX_BUSY.
2. Remove some inefficiencies in soft-TSO.
3. Fix various bugs involving device state transitions and/or reset
scheduling by error handlers.
4. Take advantage of my previous change to operstate initialisation.
5. Miscellaneous cleanup.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/drivers/net/ethernet/sfc/efx.c b/drivers/net/ethernet/sfc/efx.c
index 65a8d49..a606db4 100644
--- a/drivers/net/ethernet/sfc/efx.c
+++ b/drivers/net/ethernet/sfc/efx.c
@@ -202,11 +202,21 @@
#define EFX_ASSERT_RESET_SERIALISED(efx) \
do { \
- if ((efx->state == STATE_RUNNING) || \
+ if ((efx->state == STATE_READY) || \
(efx->state == STATE_DISABLED)) \
ASSERT_RTNL(); \
} while (0)
+static int efx_check_disabled(struct efx_nic *efx)
+{
+ if (efx->state == STATE_DISABLED) {
+ netif_err(efx, drv, efx->net_dev,
+ "device is disabled due to earlier errors\n");
+ return -EIO;
+ }
+ return 0;
+}
+
/**************************************************************************
*
* Event queue processing
@@ -630,6 +640,16 @@
efx->rx_buffer_order = get_order(efx->rx_buffer_len +
sizeof(struct efx_rx_page_state));
+ /* We must keep at least one descriptor in a TX ring empty.
+ * We could avoid this when the queue size does not exactly
+ * match the hardware ring size, but it's not that important.
+ * Therefore we stop the queue when one more skb might fill
+ * the ring completely. We wake it when half way back to
+ * empty.
+ */
+ efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
+ efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
+
/* Initialise the channels */
efx_for_each_channel(channel, efx) {
efx_for_each_channel_tx_queue(tx_queue, channel)
@@ -730,7 +750,11 @@
struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
u32 old_rxq_entries, old_txq_entries;
unsigned i, next_buffer_table = 0;
- int rc = 0;
+ int rc;
+
+ rc = efx_check_disabled(efx);
+ if (rc)
+ return rc;
/* Not all channels should be reallocated. We must avoid
* reallocating their buffer table entries.
@@ -1365,6 +1389,8 @@
{
struct efx_channel *channel;
+ BUG_ON(efx->state == STATE_DISABLED);
+
if (efx->legacy_irq)
efx->legacy_irq_enabled = true;
efx_nic_enable_interrupts(efx);
@@ -1382,6 +1408,9 @@
{
struct efx_channel *channel;
+ if (efx->state == STATE_DISABLED)
+ return;
+
efx_mcdi_mode_poll(efx);
efx_nic_disable_interrupts(efx);
@@ -1533,22 +1562,21 @@
return rc;
}
-/* Called after previous invocation(s) of efx_stop_all, restarts the port,
- * kernel transmit queues and NAPI processing, and ensures that the port is
- * scheduled to be reconfigured. This function is safe to call multiple
- * times when the NIC is in any state.
+/* If the interface is supposed to be running but is not, start
+ * the hardware and software data path, regular activity for the port
+ * (MAC statistics, link polling, etc.) and schedule the port to be
+ * reconfigured. Interrupts must already be enabled. This function
+ * is safe to call multiple times, so long as the NIC is not disabled.
+ * Requires the RTNL lock.
*/
static void efx_start_all(struct efx_nic *efx)
{
EFX_ASSERT_RESET_SERIALISED(efx);
+ BUG_ON(efx->state == STATE_DISABLED);
/* Check that it is appropriate to restart the interface. All
* of these flags are safe to read under just the rtnl lock */
- if (efx->port_enabled)
- return;
- if ((efx->state != STATE_RUNNING) && (efx->state != STATE_INIT))
- return;
- if (!netif_running(efx->net_dev))
+ if (efx->port_enabled || !netif_running(efx->net_dev))
return;
efx_start_port(efx);
@@ -1582,11 +1610,11 @@
cancel_work_sync(&efx->mac_work);
}
-/* Quiesce hardware and software without bringing the link down.
- * Safe to call multiple times, when the nic and interface is in any
- * state. The caller is guaranteed to subsequently be in a position
- * to modify any hardware and software state they see fit without
- * taking locks. */
+/* Quiesce the hardware and software data path, and regular activity
+ * for the port without bringing the link down. Safe to call multiple
+ * times with the NIC in almost any state, but interrupts should be
+ * enabled. Requires the RTNL lock.
+ */
static void efx_stop_all(struct efx_nic *efx)
{
EFX_ASSERT_RESET_SERIALISED(efx);
@@ -1739,8 +1767,6 @@
struct efx_nic *efx = netdev_priv(net_dev);
struct mii_ioctl_data *data = if_mii(ifr);
- EFX_ASSERT_RESET_SERIALISED(efx);
-
/* Convert phy_id from older PRTAD/DEVAD format */
if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
(data->phy_id & 0xfc00) == 0x0400)
@@ -1820,13 +1846,14 @@
static int efx_net_open(struct net_device *net_dev)
{
struct efx_nic *efx = netdev_priv(net_dev);
- EFX_ASSERT_RESET_SERIALISED(efx);
+ int rc;
netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
raw_smp_processor_id());
- if (efx->state == STATE_DISABLED)
- return -EIO;
+ rc = efx_check_disabled(efx);
+ if (rc)
+ return rc;
if (efx->phy_mode & PHY_MODE_SPECIAL)
return -EBUSY;
if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
@@ -1852,10 +1879,8 @@
netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
raw_smp_processor_id());
- if (efx->state != STATE_DISABLED) {
- /* Stop the device and flush all the channels */
- efx_stop_all(efx);
- }
+ /* Stop the device and flush all the channels */
+ efx_stop_all(efx);
return 0;
}
@@ -1915,9 +1940,11 @@
static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
{
struct efx_nic *efx = netdev_priv(net_dev);
+ int rc;
- EFX_ASSERT_RESET_SERIALISED(efx);
-
+ rc = efx_check_disabled(efx);
+ if (rc)
+ return rc;
if (new_mtu > EFX_MAX_MTU)
return -EINVAL;
@@ -1926,8 +1953,6 @@
netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
mutex_lock(&efx->mac_lock);
- /* Reconfigure the MAC before enabling the dma queues so that
- * the RX buffers don't overflow */
net_dev->mtu = new_mtu;
efx->type->reconfigure_mac(efx);
mutex_unlock(&efx->mac_lock);
@@ -1942,8 +1967,6 @@
struct sockaddr *addr = data;
char *new_addr = addr->sa_data;
- EFX_ASSERT_RESET_SERIALISED(efx);
-
if (!is_valid_ether_addr(new_addr)) {
netif_err(efx, drv, efx->net_dev,
"invalid ethernet MAC address requested: %pM\n",
@@ -2079,11 +2102,27 @@
rtnl_lock();
+ /* Enable resets to be scheduled and check whether any were
+ * already requested. If so, the NIC is probably hosed so we
+ * abort.
+ */
+ efx->state = STATE_READY;
+ smp_mb(); /* ensure we change state before checking reset_pending */
+ if (efx->reset_pending) {
+ netif_err(efx, probe, efx->net_dev,
+ "aborting probe due to scheduled reset\n");
+ rc = -EIO;
+ goto fail_locked;
+ }
+
rc = dev_alloc_name(net_dev, net_dev->name);
if (rc < 0)
goto fail_locked;
efx_update_name(efx);
+ /* Always start with carrier off; PHY events will detect the link */
+ netif_carrier_off(net_dev);
+
rc = register_netdevice(net_dev);
if (rc)
goto fail_locked;
@@ -2094,9 +2133,6 @@
efx_init_tx_queue_core_txq(tx_queue);
}
- /* Always start with carrier off; PHY events will detect the link */
- netif_carrier_off(net_dev);
-
rtnl_unlock();
rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
@@ -2108,14 +2144,14 @@
return 0;
+fail_registered:
+ rtnl_lock();
+ unregister_netdevice(net_dev);
fail_locked:
+ efx->state = STATE_UNINIT;
rtnl_unlock();
netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
return rc;
-
-fail_registered:
- unregister_netdev(net_dev);
- return rc;
}
static void efx_unregister_netdev(struct efx_nic *efx)
@@ -2138,7 +2174,11 @@
strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
- unregister_netdev(efx->net_dev);
+
+ rtnl_lock();
+ unregister_netdevice(efx->net_dev);
+ efx->state = STATE_UNINIT;
+ rtnl_unlock();
}
/**************************************************************************
@@ -2154,9 +2194,9 @@
EFX_ASSERT_RESET_SERIALISED(efx);
efx_stop_all(efx);
- mutex_lock(&efx->mac_lock);
-
efx_stop_interrupts(efx, false);
+
+ mutex_lock(&efx->mac_lock);
if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
efx->phy_op->fini(efx);
efx->type->fini(efx);
@@ -2276,16 +2316,15 @@
if (!pending)
return;
- /* If we're not RUNNING then don't reset. Leave the reset_pending
- * flags set so that efx_pci_probe_main will be retried */
- if (efx->state != STATE_RUNNING) {
- netif_info(efx, drv, efx->net_dev,
- "scheduled reset quenched. NIC not RUNNING\n");
- return;
- }
-
rtnl_lock();
- (void)efx_reset(efx, fls(pending) - 1);
+
+ /* We checked the state in efx_schedule_reset() but it may
+ * have changed by now. Now that we have the RTNL lock,
+ * it cannot change again.
+ */
+ if (efx->state == STATE_READY)
+ (void)efx_reset(efx, fls(pending) - 1);
+
rtnl_unlock();
}
@@ -2311,6 +2350,13 @@
}
set_bit(method, &efx->reset_pending);
+ smp_mb(); /* ensure we change reset_pending before checking state */
+
+ /* If we're not READY then just leave the flags set as the cue
+ * to abort probing or reschedule the reset later.
+ */
+ if (ACCESS_ONCE(efx->state) != STATE_READY)
+ return;
/* efx_process_channel() will no longer read events once a
* reset is scheduled. So switch back to poll'd MCDI completions. */
@@ -2376,13 +2422,12 @@
/* This zeroes out and then fills in the invariants in a struct
* efx_nic (including all sub-structures).
*/
-static int efx_init_struct(struct efx_nic *efx, const struct efx_nic_type *type,
+static int efx_init_struct(struct efx_nic *efx,
struct pci_dev *pci_dev, struct net_device *net_dev)
{
int i;
/* Initialise common structures */
- memset(efx, 0, sizeof(*efx));
spin_lock_init(&efx->biu_lock);
#ifdef CONFIG_SFC_MTD
INIT_LIST_HEAD(&efx->mtd_list);
@@ -2392,7 +2437,7 @@
INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
efx->pci_dev = pci_dev;
efx->msg_enable = debug;
- efx->state = STATE_INIT;
+ efx->state = STATE_UNINIT;
strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
efx->net_dev = net_dev;
@@ -2409,8 +2454,6 @@
goto fail;
}
- efx->type = type;
-
EFX_BUG_ON_PARANOID(efx->type->phys_addr_channels > EFX_MAX_CHANNELS);
/* Higher numbered interrupt modes are less capable! */
@@ -2455,6 +2498,12 @@
*/
static void efx_pci_remove_main(struct efx_nic *efx)
{
+ /* Flush reset_work. It can no longer be scheduled since we
+ * are not READY.
+ */
+ BUG_ON(efx->state == STATE_READY);
+ cancel_work_sync(&efx->reset_work);
+
#ifdef CONFIG_RFS_ACCEL
free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
efx->net_dev->rx_cpu_rmap = NULL;
@@ -2480,24 +2529,15 @@
/* Mark the NIC as fini, then stop the interface */
rtnl_lock();
- efx->state = STATE_FINI;
dev_close(efx->net_dev);
-
- /* Allow any queued efx_resets() to complete */
+ efx_stop_interrupts(efx, false);
rtnl_unlock();
- efx_stop_interrupts(efx, false);
efx_sriov_fini(efx);
efx_unregister_netdev(efx);
efx_mtd_remove(efx);
- /* Wait for any scheduled resets to complete. No more will be
- * scheduled from this point because efx_stop_all() has been
- * called, we are no longer registered with driverlink, and
- * the net_device's have been removed. */
- cancel_work_sync(&efx->reset_work);
-
efx_pci_remove_main(efx);
efx_fini_io(efx);
@@ -2617,7 +2657,6 @@
static int __devinit efx_pci_probe(struct pci_dev *pci_dev,
const struct pci_device_id *entry)
{
- const struct efx_nic_type *type = (const struct efx_nic_type *) entry->driver_data;
struct net_device *net_dev;
struct efx_nic *efx;
int rc;
@@ -2627,10 +2666,12 @@
EFX_MAX_RX_QUEUES);
if (!net_dev)
return -ENOMEM;
- net_dev->features |= (type->offload_features | NETIF_F_SG |
+ efx = netdev_priv(net_dev);
+ efx->type = (const struct efx_nic_type *) entry->driver_data;
+ net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
NETIF_F_HIGHDMA | NETIF_F_TSO |
NETIF_F_RXCSUM);
- if (type->offload_features & NETIF_F_V6_CSUM)
+ if (efx->type->offload_features & NETIF_F_V6_CSUM)
net_dev->features |= NETIF_F_TSO6;
/* Mask for features that also apply to VLAN devices */
net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
@@ -2638,10 +2679,9 @@
NETIF_F_RXCSUM);
/* All offloads can be toggled */
net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
- efx = netdev_priv(net_dev);
pci_set_drvdata(pci_dev, efx);
SET_NETDEV_DEV(net_dev, &pci_dev->dev);
- rc = efx_init_struct(efx, type, pci_dev, net_dev);
+ rc = efx_init_struct(efx, pci_dev, net_dev);
if (rc)
goto fail1;
@@ -2656,28 +2696,9 @@
goto fail2;
rc = efx_pci_probe_main(efx);
-
- /* Serialise against efx_reset(). No more resets will be
- * scheduled since efx_stop_all() has been called, and we have
- * not and never have been registered.
- */
- cancel_work_sync(&efx->reset_work);
-
if (rc)
goto fail3;
- /* If there was a scheduled reset during probe, the NIC is
- * probably hosed anyway.
- */
- if (efx->reset_pending) {
- rc = -EIO;
- goto fail4;
- }
-
- /* Switch to the running state before we expose the device to the OS,
- * so that dev_open()|efx_start_all() will actually start the device */
- efx->state = STATE_RUNNING;
-
rc = efx_register_netdev(efx);
if (rc)
goto fail4;
@@ -2717,12 +2738,18 @@
{
struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
- efx->state = STATE_FINI;
+ rtnl_lock();
- netif_device_detach(efx->net_dev);
+ if (efx->state != STATE_DISABLED) {
+ efx->state = STATE_UNINIT;
- efx_stop_all(efx);
- efx_stop_interrupts(efx, false);
+ netif_device_detach(efx->net_dev);
+
+ efx_stop_all(efx);
+ efx_stop_interrupts(efx, false);
+ }
+
+ rtnl_unlock();
return 0;
}
@@ -2731,21 +2758,25 @@
{
struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
- efx->state = STATE_INIT;
+ rtnl_lock();
- efx_start_interrupts(efx, false);
+ if (efx->state != STATE_DISABLED) {
+ efx_start_interrupts(efx, false);
- mutex_lock(&efx->mac_lock);
- efx->phy_op->reconfigure(efx);
- mutex_unlock(&efx->mac_lock);
+ mutex_lock(&efx->mac_lock);
+ efx->phy_op->reconfigure(efx);
+ mutex_unlock(&efx->mac_lock);
- efx_start_all(efx);
+ efx_start_all(efx);
- netif_device_attach(efx->net_dev);
+ netif_device_attach(efx->net_dev);
- efx->state = STATE_RUNNING;
+ efx->state = STATE_READY;
- efx->type->resume_wol(efx);
+ efx->type->resume_wol(efx);
+ }
+
+ rtnl_unlock();
/* Reschedule any quenched resets scheduled during efx_pm_freeze() */
queue_work(reset_workqueue, &efx->reset_work);
diff --git a/drivers/net/ethernet/sfc/ethtool.c b/drivers/net/ethernet/sfc/ethtool.c
index 8cba2df..2bd5c2d 100644
--- a/drivers/net/ethernet/sfc/ethtool.c
+++ b/drivers/net/ethernet/sfc/ethtool.c
@@ -529,9 +529,7 @@
if (!efx_tests)
goto fail;
-
- ASSERT_RTNL();
- if (efx->state != STATE_RUNNING) {
+ if (efx->state != STATE_READY) {
rc = -EIO;
goto fail1;
}
diff --git a/drivers/net/ethernet/sfc/falcon_boards.c b/drivers/net/ethernet/sfc/falcon_boards.c
index 8687a6c..ec1e99d 100644
--- a/drivers/net/ethernet/sfc/falcon_boards.c
+++ b/drivers/net/ethernet/sfc/falcon_boards.c
@@ -380,7 +380,7 @@
new_mode = PHY_MODE_SPECIAL;
if (!((old_mode ^ new_mode) & PHY_MODE_SPECIAL)) {
err = 0;
- } else if (efx->state != STATE_RUNNING || netif_running(efx->net_dev)) {
+ } else if (efx->state != STATE_READY || netif_running(efx->net_dev)) {
err = -EBUSY;
} else {
/* Reset the PHY, reconfigure the MAC and enable/disable
diff --git a/drivers/net/ethernet/sfc/net_driver.h b/drivers/net/ethernet/sfc/net_driver.h
index cd9c0a9..7ab1232 100644
--- a/drivers/net/ethernet/sfc/net_driver.h
+++ b/drivers/net/ethernet/sfc/net_driver.h
@@ -91,29 +91,31 @@
};
/**
- * struct efx_tx_buffer - An Efx TX buffer
- * @skb: The associated socket buffer.
- * Set only on the final fragment of a packet; %NULL for all other
- * fragments. When this fragment completes, then we can free this
- * skb.
- * @tsoh: The associated TSO header structure, or %NULL if this
- * buffer is not a TSO header.
+ * 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
+ * @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.
* This field is zero when the queue slot is empty.
- * @continuation: True if this fragment is not the end of a packet.
- * @unmap_single: True if dma_unmap_single should be used.
* @unmap_len: Length of this fragment to unmap
*/
struct efx_tx_buffer {
- const struct sk_buff *skb;
- struct efx_tso_header *tsoh;
+ union {
+ const struct sk_buff *skb;
+ void *heap_buf;
+ };
dma_addr_t dma_addr;
+ unsigned short flags;
unsigned short len;
- bool continuation;
- bool unmap_single;
unsigned short unmap_len;
};
+#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_HEAP 4 /* buffer was allocated with kmalloc() */
+#define EFX_TX_BUF_MAP_SINGLE 8 /* buffer was mapped with dma_map_single() */
/**
* struct efx_tx_queue - An Efx TX queue
@@ -133,6 +135,7 @@
* @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?
@@ -156,9 +159,6 @@
* 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
@@ -175,6 +175,7 @@
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;
@@ -187,7 +188,6 @@
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;
@@ -430,11 +430,9 @@
#define EFX_INT_MODE_USE_MSI(x) (((x)->interrupt_mode) <= EFX_INT_MODE_MSI)
enum nic_state {
- STATE_INIT = 0,
- STATE_RUNNING = 1,
- STATE_FINI = 2,
- STATE_DISABLED = 3,
- STATE_MAX,
+ STATE_UNINIT = 0, /* device being probed/removed or is frozen */
+ STATE_READY = 1, /* hardware ready and netdev registered */
+ STATE_DISABLED = 2, /* device disabled due to hardware errors */
};
/*
@@ -654,7 +652,7 @@
* @irq_rx_adaptive: Adaptive IRQ moderation enabled for RX event queues
* @irq_rx_moderation: IRQ moderation time for RX event queues
* @msg_enable: Log message enable flags
- * @state: Device state flag. Serialised by the rtnl_lock.
+ * @state: Device state number (%STATE_*). Serialised by the rtnl_lock.
* @reset_pending: Bitmask for pending resets
* @tx_queue: TX DMA queues
* @rx_queue: RX DMA queues
@@ -664,6 +662,8 @@
* should be allocated for this NIC
* @rxq_entries: Size of receive queues requested by user.
* @txq_entries: Size of transmit queues requested by user.
+ * @txq_stop_thresh: TX queue fill level at or above which we stop it.
+ * @txq_wake_thresh: TX queue fill level at or below which we wake it.
* @tx_dc_base: Base qword address in SRAM of TX queue descriptor caches
* @rx_dc_base: Base qword address in SRAM of RX queue descriptor caches
* @sram_lim_qw: Qword address limit of SRAM
@@ -774,6 +774,9 @@
unsigned rxq_entries;
unsigned txq_entries;
+ unsigned int txq_stop_thresh;
+ unsigned int txq_wake_thresh;
+
unsigned tx_dc_base;
unsigned rx_dc_base;
unsigned sram_lim_qw;
diff --git a/drivers/net/ethernet/sfc/nic.c b/drivers/net/ethernet/sfc/nic.c
index 326d799..cdff40b 100644
--- a/drivers/net/ethernet/sfc/nic.c
+++ b/drivers/net/ethernet/sfc/nic.c
@@ -298,7 +298,7 @@
/**************************************************************************
*
* Generic buffer handling
- * These buffers are used for interrupt status and MAC stats
+ * These buffers are used for interrupt status, MAC stats, etc.
*
**************************************************************************/
@@ -401,8 +401,10 @@
++tx_queue->write_count;
/* Create TX descriptor ring entry */
+ BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
EFX_POPULATE_QWORD_4(*txd,
- FSF_AZ_TX_KER_CONT, buffer->continuation,
+ FSF_AZ_TX_KER_CONT,
+ buffer->flags & EFX_TX_BUF_CONT,
FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
FSF_AZ_TX_KER_BUF_REGION, 0,
FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
diff --git a/drivers/net/ethernet/sfc/tx.c b/drivers/net/ethernet/sfc/tx.c
index 1871343..ebca75e 100644
--- a/drivers/net/ethernet/sfc/tx.c
+++ b/drivers/net/ethernet/sfc/tx.c
@@ -22,14 +22,6 @@
#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,
@@ -39,67 +31,32 @@
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);
+ } 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)
@@ -138,6 +95,56 @@
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
*
@@ -151,7 +158,7 @@
* 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)
@@ -160,12 +167,11 @@
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);
@@ -183,14 +189,11 @@
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 */
@@ -205,39 +208,10 @@
/* 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);
@@ -247,13 +221,14 @@
/* 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;
@@ -264,20 +239,22 @@
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:
@@ -289,7 +266,6 @@
/* 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;
@@ -297,12 +273,11 @@
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 */
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
@@ -310,7 +285,7 @@
DMA_TO_DEVICE);
}
- return rc;
+ return NETDEV_TX_OK;
}
/* Remove packets from the TX queue
@@ -340,8 +315,6 @@
}
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;
@@ -450,6 +423,7 @@
{
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);
@@ -457,15 +431,18 @@
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);
}
@@ -480,11 +457,26 @@
}
}
+/* 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);
@@ -500,17 +492,28 @@
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;
@@ -546,8 +549,6 @@
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;
}
@@ -568,13 +569,12 @@
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;
@@ -582,6 +582,14 @@
"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;
}
@@ -604,22 +612,7 @@
#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
@@ -631,10 +624,12 @@
* @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
*
* The state used during segmentation. It is put into this data structure
* just to make it easy to pass into inline functions.
@@ -651,11 +646,13 @@
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;
};
@@ -687,91 +684,43 @@
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;
+ EFX_BUG_ON_PARANOID(buffer->len);
+ EFX_BUG_ON_PARANOID(buffer->flags);
+ EFX_BUG_ON_PARANOID(buffer->unmap_len);
+
+ 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;
+
+ if (unlikely(!page_buf->addr) &&
+ efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE))
+ 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++;
+
+ 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;
}
- /* dma_alloc_coherent() allocates pages. */
- EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1u));
+ buffer->len = len;
- 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;
-
- 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;
- }
-
- dma_free_coherent(dma_dev, PAGE_SIZE, (void *)base_kva, base_dma);
-}
-
-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;
- }
-
- tsoh->unmap_len = header_len;
- return tsoh;
-}
-
-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;
}
/**
@@ -781,47 +730,19 @@
* @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;
@@ -830,12 +751,9 @@
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);
+ EFX_BUG_ON_PARANOID(buffer->flags);
buffer->dma_addr = dma_addr;
@@ -845,7 +763,8 @@
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;
}
@@ -853,7 +772,6 @@
EFX_BUG_ON_PARANOID(!len);
buffer->len = len;
*final_buffer = buffer;
- return 0;
}
@@ -864,54 +782,42 @@
* 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->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)
- 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->continuation = true;
+ efx_dequeue_buffer(tx_queue, buffer, NULL, NULL);
}
}
@@ -919,17 +825,16 @@
/* Parse the SKB header and initialise state. */
static void tso_start(struct tso_state *st, 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;
-
- 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;
+ st->header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u);
+ 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);
@@ -938,7 +843,7 @@
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,
@@ -947,7 +852,7 @@
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;
@@ -965,7 +870,7 @@
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;
@@ -982,20 +887,19 @@
* @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);
@@ -1006,25 +910,24 @@
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;
}
@@ -1035,36 +938,25 @@
* @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));
+ tsoh_th = (struct tcphdr *)(header + st->tcp_off);
/* Copy and update the headers. */
memcpy(header, skb->data, st->header_len);
@@ -1073,19 +965,19 @@
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);
+ st->packet_space = skb_shinfo(skb)->gso_size;
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;
+ st->packet_space = st->out_len;
tsoh_th->fin = tcp_hdr(skb)->fin;
tsoh_th->psh = tcp_hdr(skb)->psh;
}
+ ip_length = st->ip_base_len + st->packet_space;
if (st->protocol == htons(ETH_P_IP)) {
- struct iphdr *tsoh_iph =
- (struct iphdr *)(header + SKB_IPV4_OFF(skb));
+ struct iphdr *tsoh_iph = (struct iphdr *)(header + st->ip_off);
tsoh_iph->tot_len = htons(ip_length);
@@ -1094,16 +986,16 @@
st->ipv4_id++;
} else {
struct ipv6hdr *tsoh_iph =
- (struct ipv6hdr *)(header + SKB_IPV6_OFF(skb));
+ (struct ipv6hdr *)(header + st->ip_off);
- tsoh_iph->payload_len = htons(ip_length - sizeof(*tsoh_iph));
+ tsoh_iph->payload_len = htons(ip_length);
}
- st->packet_space = skb_shinfo(skb)->gso_size;
- ++tx_queue->tso_packets;
+ rc = efx_tso_put_header(tx_queue, buffer, header);
+ if (unlikely(rc))
+ return rc;
- /* Form a descriptor for this header. */
- efx_tso_put_header(tx_queue, tsoh, st->header_len);
+ ++tx_queue->tso_packets;
return 0;
}
@@ -1118,13 +1010,13 @@
*
* 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 */
@@ -1156,11 +1048,7 @@
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) {
@@ -1184,6 +1072,8 @@
/* 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;
@@ -1192,10 +1082,9 @@
"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
@@ -1204,25 +1093,5 @@
}
efx_enqueue_unwind(tx_queue);
- return rc2;
-}
-
-
-/*
- * 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);
+ return NETDEV_TX_OK;
}