Author: sbruno Date: Wed Jun 3 18:01:09 2015 New Revision: 283959 URL: https://svnweb.freebsd.org/changeset/base/283959
Log: Change EM_MULTIQUEUE to a real kernconf entry and enable support for up to 2 rx/tx queues for the 82574. Program the 82574 to enable 5 msix vectors, assign 1 to each rx queue, 1 to each tx queue and 1 to the link handler. Inspired by DragonFlyBSD, enable some RSS logic for handling tx queue handling/processing. Move multiqueue handler functions so that they line up better in a diff review to if_igb.c Always enqueue tx work to be done in em_mq_start, if unable to acquire the TX lock, then this will be processed in the background later by the taskqueue. Remove mbuf argument from em_start_mq_locked() as the work is always enqueued. (stolen from igb) Setup TARC, TXDCTL and RXDCTL registers for better performance and stability in multiqueue and singlequeue implementations. Handle Intel errata 3 and generic multiqueue behavior with the initialization of TARC(0) and TARC(1) Bind interrupt threads to cpus in order. (stolen from igb) Add 2 new DDB functions, one to display the queue(s) and their settings and one to reset the adapter. Primarily used for debugging. In the multiqueue configuration, bump RXD and TXD ring size to max for the adapter (4096). Setup an RDTR of 64 and an RADV of 128 in multiqueue configuration to cut down on the number of interrupts. RADV was arbitrarily set to 2x RDTR and can be adjusted as needed. Cleanup the display in top a bit to make it clearer where the taskqueue threads are running and what they should be doing. Ensure that both queues are processed by em_local_timer() by writing them both to the IMS register to generate soft interrupts. Ensure that an soft interrupt is generated when em_msix_link() is run so that any races between assertion of the link/status interrupt and a rx/tx interrupt are handled. Document existing tuneables: hw.em.eee_setting, hw.em.msix, hw.em.smart_pwr_down, hw.em.sbp Document use of hw.em.num_queues and the new kernel option EM_MULTIQUEUE Thanks to Intel for their continued support of FreeBSD. Reviewed by: erj jfv hiren gnn wblock Obtained from: Intel Corporation MFC after: 2 weeks Relnotes: Yes Sponsored by: Limelight Networks Differential Revision: https://reviews.freebsd.org/D1994 Modified: head/share/man/man4/em.4 head/sys/conf/NOTES head/sys/conf/options head/sys/dev/e1000/e1000_defines.h head/sys/dev/e1000/if_em.c head/sys/dev/e1000/if_em.h head/sys/dev/netmap/if_em_netmap.h head/sys/modules/em/Makefile Modified: head/share/man/man4/em.4 ============================================================================== --- head/share/man/man4/em.4 Wed Jun 3 17:45:45 2015 (r283958) +++ head/share/man/man4/em.4 Wed Jun 3 18:01:09 2015 (r283959) @@ -45,6 +45,14 @@ kernel configuration file: .Cd "device em" .Ed .Pp +Optional multiqueue support is available via the following kernel +compile options: +.Bd -ragged -offset indent +.Cd "options EM_MULTIQUEUE" +.Ed +.Pp +Note: Activating EM_MULTIQUEUE support is not supported by Intel. +.Pp Alternatively, to load the driver as a module at boot time, place the following line in .Xr loader.conf 5 : @@ -197,6 +205,18 @@ Tunables can be set at the prompt before booting the kernel or stored in .Xr loader.conf 5 . .Bl -tag -width indent +.It Va hw.em.eee_setting +Disable or enable Energy Efficient Ethernet. +Default 1 (disabled). +.It Va hw.em.msix +Enable or disable MSI-X style interrupts. +Default 1 (enabled). +.It Va hw.em.smart_pwr_down +Enable or disable smart power down features on newer adapters. +Default 0 (disabled). +.It Va hw.em.sbp +Show bad packets when in promiscuous mode. +Default 0 (off). .It Va hw.em.rxd Number of receive descriptors allocated by the driver. The default value is 1024 for adapters newer than 82547, @@ -228,6 +248,11 @@ If .Va hw.em.tx_int_delay is non-zero, this tunable limits the maximum delay in which a transmit interrupt is generated. +.It Va hw.em.num_queues +Number of hardware queues that will be configured on this adapter (maximum of 2) +Defaults to 1. +Only valid with kernel configuration +.Cd "options EM_MULTIQUEUE". .El .Sh FILES .Bl -tag -width /dev/led/em* @@ -287,3 +312,5 @@ You can enable it on an .Nm interface using .Xr ifconfig 8 . +.Pp +Activating EM_MULTIQUEUE support requires MSI-X features. Modified: head/sys/conf/NOTES ============================================================================== --- head/sys/conf/NOTES Wed Jun 3 17:45:45 2015 (r283958) +++ head/sys/conf/NOTES Wed Jun 3 18:01:09 2015 (r283959) @@ -2980,6 +2980,9 @@ options RANDOM_DEBUG # Debugging messag # Module to enable execution of application via emulators like QEMU options IMAGACT_BINMISC +# Intel em(4) driver +options EM_MULTIQUEUE # Activate multiqueue features/disable MSI-X + # zlib I/O stream support # This enables support for compressed core dumps. options GZIO Modified: head/sys/conf/options ============================================================================== --- head/sys/conf/options Wed Jun 3 17:45:45 2015 (r283958) +++ head/sys/conf/options Wed Jun 3 18:01:09 2015 (r283959) @@ -940,3 +940,6 @@ RCTL opt_global.h RANDOM_YARROW opt_random.h RANDOM_FORTUNA opt_random.h RANDOM_DEBUG opt_random.h + +# Intel em(4) driver +EM_MULTIQUEUE opt_em.h Modified: head/sys/dev/e1000/e1000_defines.h ============================================================================== --- head/sys/dev/e1000/e1000_defines.h Wed Jun 3 17:45:45 2015 (r283958) +++ head/sys/dev/e1000/e1000_defines.h Wed Jun 3 18:01:09 2015 (r283959) @@ -158,10 +158,12 @@ E1000_RXDEXT_STATERR_CXE | \ E1000_RXDEXT_STATERR_RXE) +#define E1000_MRQC_RSS_ENABLE_2Q 0x00000001 #define E1000_MRQC_RSS_FIELD_MASK 0xFFFF0000 #define E1000_MRQC_RSS_FIELD_IPV4_TCP 0x00010000 #define E1000_MRQC_RSS_FIELD_IPV4 0x00020000 #define E1000_MRQC_RSS_FIELD_IPV6_TCP_EX 0x00040000 +#define E1000_MRQC_RSS_FIELD_IPV6_EX 0x00080000 #define E1000_MRQC_RSS_FIELD_IPV6 0x00100000 #define E1000_MRQC_RSS_FIELD_IPV6_TCP 0x00200000 Modified: head/sys/dev/e1000/if_em.c ============================================================================== --- head/sys/dev/e1000/if_em.c Wed Jun 3 17:45:45 2015 (r283958) +++ head/sys/dev/e1000/if_em.c Wed Jun 3 18:01:09 2015 (r283959) @@ -32,6 +32,8 @@ ******************************************************************************/ /*$FreeBSD$*/ +#include "opt_em.h" +#include "opt_ddb.h" #include "opt_inet.h" #include "opt_inet6.h" @@ -41,6 +43,10 @@ #include <sys/param.h> #include <sys/systm.h> +#ifdef DDB +#include <sys/types.h> +#include <ddb/ddb.h> +#endif #if __FreeBSD_version >= 800000 #include <sys/buf_ring.h> #endif @@ -52,6 +58,7 @@ #include <sys/mbuf.h> #include <sys/module.h> #include <sys/rman.h> +#include <sys/smp.h> #include <sys/socket.h> #include <sys/sockio.h> #include <sys/sysctl.h> @@ -208,7 +215,7 @@ static int em_resume(device_t); #ifdef EM_MULTIQUEUE static int em_mq_start(if_t, struct mbuf *); static int em_mq_start_locked(if_t, - struct tx_ring *, struct mbuf *); + struct tx_ring *); static void em_qflush(if_t); #else static void em_start(if_t); @@ -299,6 +306,10 @@ static void em_handle_tx(void *context, static void em_handle_rx(void *context, int pending); static void em_handle_link(void *context, int pending); +#ifdef EM_MULTIQUEUE +static void em_enable_vectors_82574(struct adapter *); +#endif + static void em_set_sysctl_value(struct adapter *, const char *, const char *, int *, int); static int em_set_flowcntl(SYSCTL_HANDLER_ARGS); @@ -388,6 +399,19 @@ static int em_enable_msix = TRUE; SYSCTL_INT(_hw_em, OID_AUTO, enable_msix, CTLFLAG_RDTUN, &em_enable_msix, 0, "Enable MSI-X interrupts"); +#ifdef EM_MULTIQUEUE +static int em_num_queues = 1; +SYSCTL_INT(_hw_em, OID_AUTO, num_queues, CTLFLAG_RDTUN, &em_num_queues, 0, + "82574 only: Number of queues to configure, 0 indicates autoconfigure"); +#endif + +/* +** Global variable to store last used CPU when binding queues +** to CPUs in igb_allocate_msix. Starts at CPU_FIRST and increments when a +** queue is bound to a cpu. +*/ +static int em_last_bind_cpu = -1; + /* How many packets rxeof tries to clean at a time */ static int em_rx_process_limit = 100; SYSCTL_INT(_hw_em, OID_AUTO, rx_process_limit, CTLFLAG_RDTUN, @@ -420,10 +444,10 @@ static int em_probe(device_t dev) { char adapter_name[60]; - u16 pci_vendor_id = 0; - u16 pci_device_id = 0; - u16 pci_subvendor_id = 0; - u16 pci_subdevice_id = 0; + uint16_t pci_vendor_id = 0; + uint16_t pci_device_id = 0; + uint16_t pci_subvendor_id = 0; + uint16_t pci_subdevice_id = 0; em_vendor_info_t *ent; INIT_DEBUGOUT("em_probe: begin"); @@ -550,6 +574,11 @@ em_attach(device_t dev) goto err_pci; } + /* + * Setup MSI/X or MSI if PCI Express + */ + adapter->msix = em_setup_msix(adapter); + e1000_get_bus_info(hw); /* Set up some sysctls for the tunable interrupt delays */ @@ -880,7 +909,7 @@ em_resume(device_t dev) EM_TX_LOCK(txr); #ifdef EM_MULTIQUEUE if (!drbr_empty(ifp, txr->br)) - em_mq_start_locked(ifp, txr, NULL); + em_mq_start_locked(ifp, txr); #else if (!if_sendq_empty(ifp)) em_start_locked(ifp, txr); @@ -894,105 +923,7 @@ em_resume(device_t dev) } -#ifdef EM_MULTIQUEUE -/********************************************************************* - * Multiqueue Transmit routines - * - * em_mq_start is called by the stack to initiate a transmit. - * however, if busy the driver can queue the request rather - * than do an immediate send. It is this that is an advantage - * in this driver, rather than also having multiple tx queues. - **********************************************************************/ -static int -em_mq_start_locked(if_t ifp, struct tx_ring *txr, struct mbuf *m) -{ - struct adapter *adapter = txr->adapter; - struct mbuf *next; - int err = 0, enq = 0; - - if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != - IFF_DRV_RUNNING || adapter->link_active == 0) { - if (m != NULL) - err = drbr_enqueue(ifp, txr->br, m); - return (err); - } - - enq = 0; - if (m != NULL) { - err = drbr_enqueue(ifp, txr->br, m); - if (err) - return (err); - } - - /* Process the queue */ - while ((next = drbr_peek(ifp, txr->br)) != NULL) { - if ((err = em_xmit(txr, &next)) != 0) { - if (next == NULL) - drbr_advance(ifp, txr->br); - else - drbr_putback(ifp, txr->br, next); - break; - } - drbr_advance(ifp, txr->br); - enq++; - if_inc_counter(ifp, IFCOUNTER_OBYTES, next->m_pkthdr.len); - if (next->m_flags & M_MCAST) - if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1); - if_etherbpfmtap(ifp, next); - if ((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0) - break; - } - - /* Mark the queue as having work */ - if ((enq > 0) && (txr->busy == EM_TX_IDLE)) - txr->busy = EM_TX_BUSY; - - if (txr->tx_avail < EM_MAX_SCATTER) - em_txeof(txr); - if (txr->tx_avail < EM_MAX_SCATTER) - if_setdrvflagbits(ifp, IFF_DRV_OACTIVE,0); - return (err); -} - -/* -** Multiqueue capable stack interface -*/ -static int -em_mq_start(if_t ifp, struct mbuf *m) -{ - struct adapter *adapter = if_getsoftc(ifp); - struct tx_ring *txr = adapter->tx_rings; - int error; - - if (EM_TX_TRYLOCK(txr)) { - error = em_mq_start_locked(ifp, txr, m); - EM_TX_UNLOCK(txr); - } else - error = drbr_enqueue(ifp, txr->br, m); - - return (error); -} - -/* -** Flush all ring buffers -*/ -static void -em_qflush(if_t ifp) -{ - struct adapter *adapter = if_getsoftc(ifp); - struct tx_ring *txr = adapter->tx_rings; - struct mbuf *m; - - for (int i = 0; i < adapter->num_queues; i++, txr++) { - EM_TX_LOCK(txr); - while ((m = buf_ring_dequeue_sc(txr->br)) != NULL) - m_freem(m); - EM_TX_UNLOCK(txr); - } - if_qflush(ifp); -} -#else /* !EM_MULTIQUEUE */ - +#ifndef EM_MULTIQUEUE static void em_start_locked(if_t ifp, struct tx_ring *txr) { @@ -1035,7 +966,8 @@ em_start_locked(if_t ifp, struct tx_ring txr->busy = EM_TX_BUSY; /* Send a copy of the frame to the BPF listener */ - if_etherbpfmtap(ifp, m_head); + ETHER_BPF_MTAP(ifp, m_head); + } return; @@ -1054,6 +986,115 @@ em_start(if_t ifp) } return; } +#else /* EM_MULTIQUEUE */ +/********************************************************************* + * Multiqueue Transmit routines + * + * em_mq_start is called by the stack to initiate a transmit. + * however, if busy the driver can queue the request rather + * than do an immediate send. It is this that is an advantage + * in this driver, rather than also having multiple tx queues. + **********************************************************************/ +/* +** Multiqueue capable stack interface +*/ +static int +em_mq_start(if_t ifp, struct mbuf *m) +{ + struct adapter *adapter = if_getsoftc(ifp); + struct tx_ring *txr = adapter->tx_rings; + unsigned int i, error; + + if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) + i = m->m_pkthdr.flowid % adapter->num_queues; + else + i = curcpu % adapter->num_queues; + + txr = &adapter->tx_rings[i]; + + error = drbr_enqueue(ifp, txr->br, m); + if (error) + return (error); + + if (EM_TX_TRYLOCK(txr)) { + em_mq_start_locked(ifp, txr); + EM_TX_UNLOCK(txr); + } else + taskqueue_enqueue(txr->tq, &txr->tx_task); + + return (0); +} + +static int +em_mq_start_locked(if_t ifp, struct tx_ring *txr) +{ + struct adapter *adapter = txr->adapter; + struct mbuf *next; + int err = 0, enq = 0; + + EM_TX_LOCK_ASSERT(txr); + + if (((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0) || + adapter->link_active == 0) { + return (ENETDOWN); + } + + /* Process the queue */ + while ((next = drbr_peek(ifp, txr->br)) != NULL) { + if ((err = em_xmit(txr, &next)) != 0) { + if (next == NULL) { + /* It was freed, move forward */ + drbr_advance(ifp, txr->br); + } else { + /* + * Still have one left, it may not be + * the same since the transmit function + * may have changed it. + */ + drbr_putback(ifp, txr->br, next); + } + break; + } + drbr_advance(ifp, txr->br); + enq++; + if_inc_counter(ifp, IFCOUNTER_OBYTES, next->m_pkthdr.len); + if (next->m_flags & M_MCAST) + if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1); + ETHER_BPF_MTAP(ifp, next); + if ((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0) + break; + } + + /* Mark the queue as having work */ + if ((enq > 0) && (txr->busy == EM_TX_IDLE)) + txr->busy = EM_TX_BUSY; + + if (txr->tx_avail < EM_MAX_SCATTER) + em_txeof(txr); + if (txr->tx_avail < EM_MAX_SCATTER) { + if_setdrvflagbits(ifp, IFF_DRV_OACTIVE,0); + } + return (err); +} + +/* +** Flush all ring buffers +*/ +static void +em_qflush(if_t ifp) +{ + struct adapter *adapter = if_getsoftc(ifp); + struct tx_ring *txr = adapter->tx_rings; + struct mbuf *m; + + for (int i = 0; i < adapter->num_queues; i++, txr++) { + EM_TX_LOCK(txr); + while ((m = buf_ring_dequeue_sc(txr->br)) != NULL) + m_freem(m); + EM_TX_UNLOCK(txr); + } + if_qflush(ifp); +} #endif /* EM_MULTIQUEUE */ /********************************************************************* @@ -1449,7 +1490,7 @@ em_poll(if_t ifp, enum poll_cmd cmd, int em_txeof(txr); #ifdef EM_MULTIQUEUE if (!drbr_empty(ifp, txr->br)) - em_mq_start_locked(ifp, txr, NULL); + em_mq_start_locked(ifp, txr); #else if (!if_sendq_empty(ifp)) em_start_locked(ifp, txr); @@ -1516,14 +1557,14 @@ em_handle_que(void *context, int pending struct tx_ring *txr = adapter->tx_rings; struct rx_ring *rxr = adapter->rx_rings; - if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { bool more = em_rxeof(rxr, adapter->rx_process_limit, NULL); + EM_TX_LOCK(txr); em_txeof(txr); #ifdef EM_MULTIQUEUE if (!drbr_empty(ifp, txr->br)) - em_mq_start_locked(ifp, txr, NULL); + em_mq_start_locked(ifp, txr); #else if (!if_sendq_empty(ifp)) em_start_locked(ifp, txr); @@ -1557,11 +1598,12 @@ em_msix_tx(void *arg) em_txeof(txr); #ifdef EM_MULTIQUEUE if (!drbr_empty(ifp, txr->br)) - em_mq_start_locked(ifp, txr, NULL); + em_mq_start_locked(ifp, txr); #else if (!if_sendq_empty(ifp)) em_start_locked(ifp, txr); #endif + /* Reenable this interrupt */ E1000_WRITE_REG(&adapter->hw, E1000_IMS, txr->ims); EM_TX_UNLOCK(txr); @@ -1587,9 +1629,10 @@ em_msix_rx(void *arg) more = em_rxeof(rxr, adapter->rx_process_limit, NULL); if (more) taskqueue_enqueue(rxr->tq, &rxr->rx_task); - else + else { /* Reenable this interrupt */ E1000_WRITE_REG(&adapter->hw, E1000_IMS, rxr->ims); + } return; } @@ -1616,6 +1659,16 @@ em_msix_link(void *arg) } else E1000_WRITE_REG(&adapter->hw, E1000_IMS, EM_MSIX_LINK | E1000_IMS_LSC); + /* + ** Because we must read the ICR for this interrupt + ** it may clear other causes using autoclear, for + ** this reason we simply create a soft interrupt + ** for all these vectors. + */ + if (reg_icr) { + E1000_WRITE_REG(&adapter->hw, + E1000_ICS, adapter->ims); + } return; } @@ -1629,9 +1682,10 @@ em_handle_rx(void *context, int pending) more = em_rxeof(rxr, adapter->rx_process_limit, NULL); if (more) taskqueue_enqueue(rxr->tq, &rxr->rx_task); - else + else { /* Reenable this interrupt */ E1000_WRITE_REG(&adapter->hw, E1000_IMS, rxr->ims); + } } static void @@ -1645,7 +1699,7 @@ em_handle_tx(void *context, int pending) em_txeof(txr); #ifdef EM_MULTIQUEUE if (!drbr_empty(ifp, txr->br)) - em_mq_start_locked(ifp, txr, NULL); + em_mq_start_locked(ifp, txr); #else if (!if_sendq_empty(ifp)) em_start_locked(ifp, txr); @@ -1675,7 +1729,7 @@ em_handle_link(void *context, int pendin EM_TX_LOCK(txr); #ifdef EM_MULTIQUEUE if (!drbr_empty(ifp, txr->br)) - em_mq_start_locked(ifp, txr, NULL); + em_mq_start_locked(ifp, txr); #else if (if_sendq_empty(ifp)) em_start_locked(ifp, txr); @@ -2219,7 +2273,7 @@ em_local_timer(void *arg) if_t ifp = adapter->ifp; struct tx_ring *txr = adapter->tx_rings; struct rx_ring *rxr = adapter->rx_rings; - u32 trigger; + u32 trigger = 0; EM_CORE_LOCK_ASSERT(adapter); @@ -2232,9 +2286,11 @@ em_local_timer(void *arg) e1000_rar_set(&adapter->hw, adapter->hw.mac.addr, 0); /* Mask to use in the irq trigger */ - if (adapter->msix_mem) - trigger = rxr->ims; - else + if (adapter->msix_mem) { + for (int i = 0; i < adapter->num_queues; i++, rxr++) + trigger |= rxr->ims; + rxr = adapter->rx_rings; + } else trigger = E1000_ICS_RXDMT0; /* @@ -2243,7 +2299,6 @@ em_local_timer(void *arg) ** and the HUNG state will be static if set. */ for (int i = 0; i < adapter->num_queues; i++, txr++) { - /* Last cycle a queue was declared hung */ if (txr->busy == EM_TX_HUNG) goto hung; if (txr->busy >= EM_TX_MAXTRIES) @@ -2261,14 +2316,9 @@ em_local_timer(void *arg) return; hung: /* Looks like we're hung */ - device_printf(adapter->dev, "Watchdog timeout -- resetting\n"); - device_printf(adapter->dev, - "Queue(%d) tdh = %d, hw tdt = %d\n", txr->me, - E1000_READ_REG(&adapter->hw, E1000_TDH(txr->me)), - E1000_READ_REG(&adapter->hw, E1000_TDT(txr->me))); - device_printf(adapter->dev,"TX(%d) desc avail = %d," - "Next TX to Clean = %d\n", - txr->me, txr->tx_avail, txr->next_to_clean); + device_printf(adapter->dev, "Watchdog timeout Queue[%d]-- resetting\n", + txr->me); + em_print_debug_info(adapter); if_setdrvflagbits(ifp, 0, IFF_DRV_RUNNING); adapter->watchdog_events++; em_init_locked(adapter); @@ -2320,7 +2370,7 @@ em_update_link_status(struct adapter *ad (hw->mac.type == e1000_82572))) { int tarc0; tarc0 = E1000_READ_REG(hw, E1000_TARC(0)); - tarc0 &= ~SPEED_MODE_BIT; + tarc0 &= ~TARC_SPEED_MODE_BIT; E1000_WRITE_REG(hw, E1000_TARC(0), tarc0); } if (bootverbose) @@ -2436,14 +2486,6 @@ em_allocate_pci_resources(struct adapter rman_get_bushandle(adapter->memory); adapter->hw.hw_addr = (u8 *)&adapter->osdep.mem_bus_space_handle; - /* Default to a single queue */ - adapter->num_queues = 1; - - /* - * Setup MSI/X or MSI if PCI Express - */ - adapter->msix = em_setup_msix(adapter); - adapter->hw.back = &adapter->osdep; return (0); @@ -2518,13 +2560,14 @@ em_allocate_msix(struct adapter *adapter struct tx_ring *txr = adapter->tx_rings; struct rx_ring *rxr = adapter->rx_rings; int error, rid, vector = 0; + int cpu_id = 0; /* Make sure all interrupts are disabled */ E1000_WRITE_REG(&adapter->hw, E1000_IMC, 0xffffffff); /* First set up ring resources */ - for (int i = 0; i < adapter->num_queues; i++, txr++, rxr++) { + for (int i = 0; i < adapter->num_queues; i++, rxr++, vector++) { /* RX ring */ rid = vector + 1; @@ -2544,14 +2587,20 @@ em_allocate_msix(struct adapter *adapter return (error); } #if __FreeBSD_version >= 800504 - bus_describe_intr(dev, rxr->res, rxr->tag, "rx %d", i); + bus_describe_intr(dev, rxr->res, rxr->tag, "rx%d", i); #endif - rxr->msix = vector++; /* NOTE increment vector for TX */ + rxr->msix = vector; + + if (em_last_bind_cpu < 0) + em_last_bind_cpu = CPU_FIRST(); + cpu_id = em_last_bind_cpu; + bus_bind_intr(dev, rxr->res, cpu_id); + TASK_INIT(&rxr->rx_task, 0, em_handle_rx, rxr); rxr->tq = taskqueue_create_fast("em_rxq", M_NOWAIT, taskqueue_thread_enqueue, &rxr->tq); - taskqueue_start_threads(&rxr->tq, 1, PI_NET, "%s rxq", - device_get_nameunit(adapter->dev)); + taskqueue_start_threads(&rxr->tq, 1, PI_NET, "%s rxq (cpuid %d)", + device_get_nameunit(adapter->dev), cpu_id); /* ** Set the bit to enable interrupt ** in E1000_IMS -- bits 20 and 21 @@ -2559,8 +2608,13 @@ em_allocate_msix(struct adapter *adapter ** NOTHING to do with the MSIX vector */ rxr->ims = 1 << (20 + i); + adapter->ims |= rxr->ims; adapter->ivars |= (8 | rxr->msix) << (i * 4); + em_last_bind_cpu = CPU_NEXT(em_last_bind_cpu); + } + + for (int i = 0; i < adapter->num_queues; i++, txr++, vector++) { /* TX ring */ rid = vector + 1; txr->res = bus_alloc_resource_any(dev, @@ -2578,14 +2632,20 @@ em_allocate_msix(struct adapter *adapter return (error); } #if __FreeBSD_version >= 800504 - bus_describe_intr(dev, txr->res, txr->tag, "tx %d", i); + bus_describe_intr(dev, txr->res, txr->tag, "tx%d", i); #endif - txr->msix = vector++; /* Increment vector for next pass */ + txr->msix = vector; + + if (em_last_bind_cpu < 0) + em_last_bind_cpu = CPU_FIRST(); + cpu_id = em_last_bind_cpu; + bus_bind_intr(dev, txr->res, cpu_id); + TASK_INIT(&txr->tx_task, 0, em_handle_tx, txr); txr->tq = taskqueue_create_fast("em_txq", M_NOWAIT, taskqueue_thread_enqueue, &txr->tq); - taskqueue_start_threads(&txr->tq, 1, PI_NET, "%s txq", - device_get_nameunit(adapter->dev)); + taskqueue_start_threads(&txr->tq, 1, PI_NET, "%s txq (cpuid %d)", + device_get_nameunit(adapter->dev), cpu_id); /* ** Set the bit to enable interrupt ** in E1000_IMS -- bits 22 and 23 @@ -2593,13 +2653,16 @@ em_allocate_msix(struct adapter *adapter ** NOTHING to do with the MSIX vector */ txr->ims = 1 << (22 + i); + adapter->ims |= txr->ims; adapter->ivars |= (8 | txr->msix) << (8 + (i * 4)); + + em_last_bind_cpu = CPU_NEXT(em_last_bind_cpu); } /* Link interrupt */ - ++rid; + rid = vector + 1; adapter->res = bus_alloc_resource_any(dev, - SYS_RES_IRQ, &rid, RF_ACTIVE); + SYS_RES_IRQ, &rid, RF_SHAREABLE | RF_ACTIVE); if (!adapter->res) { device_printf(dev,"Unable to allocate " "bus resource: Link interrupt [%d]\n", rid); @@ -2615,7 +2678,7 @@ em_allocate_msix(struct adapter *adapter return (error); } #if __FreeBSD_version >= 800504 - bus_describe_intr(dev, adapter->res, adapter->tag, "link"); + bus_describe_intr(dev, adapter->res, adapter->tag, "link"); #endif adapter->linkvec = vector; adapter->ivars |= (8 | vector) << 16; @@ -2639,9 +2702,8 @@ em_free_pci_resources(struct adapter *ad */ for (int i = 0; i < adapter->num_queues; i++) { txr = &adapter->tx_rings[i]; - rxr = &adapter->rx_rings[i]; /* an early abort? */ - if ((txr == NULL) || (rxr == NULL)) + if (txr == NULL) break; rid = txr->msix +1; if (txr->tag != NULL) { @@ -2651,6 +2713,11 @@ em_free_pci_resources(struct adapter *ad if (txr->res != NULL) bus_release_resource(dev, SYS_RES_IRQ, rid, txr->res); + + rxr = &adapter->rx_rings[i]; + /* an early abort? */ + if (rxr == NULL) + break; rid = rxr->msix +1; if (rxr->tag != NULL) { bus_teardown_intr(dev, rxr->res, rxr->tag); @@ -2700,14 +2767,19 @@ em_setup_msix(struct adapter *adapter) device_t dev = adapter->dev; int val; + /* Nearly always going to use one queue */ + adapter->num_queues = 1; + /* - ** Setup MSI/X for Hartwell: tests have shown - ** use of two queues to be unstable, and to - ** provide no great gain anyway, so we simply - ** seperate the interrupts and use a single queue. + ** Try using MSI-X for Hartwell adapters */ if ((adapter->hw.mac.type == e1000_82574) && (em_enable_msix == TRUE)) { +#ifdef EM_MULTIQUEUE + adapter->num_queues = (em_num_queues == 1) ? 1 : 2; + if (adapter->num_queues > 1) + em_enable_vectors_82574(adapter); +#endif /* Map the MSIX BAR */ int rid = PCIR_BAR(EM_MSIX_BAR); adapter->msix_mem = bus_alloc_resource_any(dev, @@ -2719,16 +2791,34 @@ em_setup_msix(struct adapter *adapter) goto msi; } val = pci_msix_count(dev); - /* We only need/want 3 vectors */ - if (val >= 3) - val = 3; - else { - device_printf(adapter->dev, - "MSIX: insufficient vectors, using MSI\n"); - goto msi; + +#ifdef EM_MULTIQUEUE + /* We need 5 vectors in the multiqueue case */ + if (adapter->num_queues > 1 ) { + if (val >= 5) + val = 5; + else { + adapter->num_queues = 1; + device_printf(adapter->dev, + "Insufficient MSIX vectors for >1 queue, " + "using single queue...\n"); + goto msix_one; + } + } else { +msix_one: +#endif + if (val >= 3) + val = 3; + else { + device_printf(adapter->dev, + "Insufficient MSIX vectors, using MSI\n"); + goto msi; + } +#ifdef EM_MULTIQUEUE } +#endif - if ((pci_alloc_msix(dev, &val) == 0) && (val == 3)) { + if ((pci_alloc_msix(dev, &val) == 0)) { device_printf(adapter->dev, "Using MSIX interrupts " "with %d vectors\n", val); @@ -2749,7 +2839,7 @@ msi: } val = 1; if (pci_alloc_msi(dev, &val) == 0) { - device_printf(adapter->dev,"Using an MSI interrupt\n"); + device_printf(adapter->dev, "Using an MSI interrupt\n"); return (val); } /* Should only happen due to manual configuration */ @@ -3394,7 +3484,7 @@ em_initialize_transmit_unit(struct adapt { struct tx_ring *txr = adapter->tx_rings; struct e1000_hw *hw = &adapter->hw; - u32 tctl, tarc, tipg = 0; + u32 tctl, txdctl = 0, tarc, tipg = 0; INIT_DEBUGOUT("em_initialize_transmit_unit: begin"); @@ -3416,6 +3506,15 @@ em_initialize_transmit_unit(struct adapt E1000_READ_REG(&adapter->hw, E1000_TDLEN(i))); txr->busy = EM_TX_IDLE; + txdctl = 0; /* clear txdctl */ + txdctl |= 0x1f; /* PTHRESH */ + txdctl |= 1 << 8; /* HTHRESH */ + txdctl |= 1 << 16;/* WTHRESH */ + txdctl |= 1 << 22; /* Reserved bit 22 must always be 1 */ + txdctl |= E1000_TXDCTL_GRAN; + txdctl |= 1 << 25; /* LWTHRESH */ + + E1000_WRITE_REG(hw, E1000_TXDCTL(i), txdctl); } /* Set the default values for the Tx Inter Packet Gap timer */ @@ -3446,15 +3545,25 @@ em_initialize_transmit_unit(struct adapt if ((adapter->hw.mac.type == e1000_82571) || (adapter->hw.mac.type == e1000_82572)) { tarc = E1000_READ_REG(&adapter->hw, E1000_TARC(0)); - tarc |= SPEED_MODE_BIT; + tarc |= TARC_SPEED_MODE_BIT; E1000_WRITE_REG(&adapter->hw, E1000_TARC(0), tarc); } else if (adapter->hw.mac.type == e1000_80003es2lan) { + /* errata: program both queues to unweighted RR */ tarc = E1000_READ_REG(&adapter->hw, E1000_TARC(0)); tarc |= 1; E1000_WRITE_REG(&adapter->hw, E1000_TARC(0), tarc); tarc = E1000_READ_REG(&adapter->hw, E1000_TARC(1)); tarc |= 1; E1000_WRITE_REG(&adapter->hw, E1000_TARC(1), tarc); + } else if (adapter->hw.mac.type == e1000_82574) { + tarc = E1000_READ_REG(&adapter->hw, E1000_TARC(0)); + tarc |= TARC_ERRATA_BIT; + if ( adapter->num_queues > 1) { + tarc |= (TARC_COMPENSATION_MODE | TARC_MQ_FIX); + E1000_WRITE_REG(&adapter->hw, E1000_TARC(0), tarc); + E1000_WRITE_REG(&adapter->hw, E1000_TARC(1), tarc); + } else + E1000_WRITE_REG(&adapter->hw, E1000_TARC(0), tarc); } adapter->txd_cmd = E1000_TXD_CMD_IFCS; @@ -3885,8 +3994,9 @@ em_txeof(struct tx_ring *txr) * TX lock which, with a single queue, guarantees * sanity. */ - if (txr->tx_avail >= EM_MAX_SCATTER) + if (txr->tx_avail >= EM_MAX_SCATTER) { if_setdrvflagbits(ifp, 0, IFF_DRV_OACTIVE); + } /* Disable hang detection if all clean */ if (txr->tx_avail == adapter->num_tx_desc) @@ -4258,6 +4368,9 @@ em_initialize_receive_unit(struct adapte E1000_WRITE_REG(&adapter->hw, E1000_RADV, adapter->rx_abs_int_delay.value); + + E1000_WRITE_REG(&adapter->hw, E1000_RDTR, + adapter->rx_int_delay.value); /* * Set the interrupt throttling rate. Value is calculated * as DEFAULT_ITR = 1/(MAX_INTS_PER_SEC * 256ns) @@ -4269,20 +4382,65 @@ em_initialize_receive_unit(struct adapte ** using the EITR register (82574 only) */ if (hw->mac.type == e1000_82574) { + u32 rfctl; for (int i = 0; i < 4; i++) E1000_WRITE_REG(hw, E1000_EITR_82574(i), DEFAULT_ITR); /* Disable accelerated acknowledge */ - E1000_WRITE_REG(hw, E1000_RFCTL, E1000_RFCTL_ACK_DIS); + rfctl = E1000_READ_REG(hw, E1000_RFCTL); + rfctl |= E1000_RFCTL_ACK_DIS; + E1000_WRITE_REG(hw, E1000_RFCTL, rfctl); } rxcsum = E1000_READ_REG(hw, E1000_RXCSUM); - if (if_getcapenable(ifp) & IFCAP_RXCSUM) + if (if_getcapenable(ifp) & IFCAP_RXCSUM) { +#ifdef EM_MULTIQUEUE + rxcsum |= E1000_RXCSUM_TUOFL | + E1000_RXCSUM_IPOFL | + E1000_RXCSUM_PCSD; +#else rxcsum |= E1000_RXCSUM_TUOFL; - else +#endif + } else rxcsum &= ~E1000_RXCSUM_TUOFL; + E1000_WRITE_REG(hw, E1000_RXCSUM, rxcsum); +#ifdef EM_MULTIQUEUE + if (adapter->num_queues > 1) { + uint32_t rss_key[10]; + uint32_t reta; + int i; + + /* + * Configure RSS key + */ + arc4rand(rss_key, sizeof(rss_key), 0); + for (i = 0; i < 10; ++i) + E1000_WRITE_REG_ARRAY(hw,E1000_RSSRK(0), i, rss_key[i]); + + /* + * Configure RSS redirect table in following fashion: + * (hash & ring_cnt_mask) == rdr_table[(hash & rdr_table_mask)] + */ + reta = 0; + for (i = 0; i < 4; ++i) { + uint32_t q; + q = (i % adapter->num_queues) << 7; + reta |= q << (8 * i); + } + for (i = 0; i < 32; ++i) + E1000_WRITE_REG(hw, E1000_RETA(i), reta); + + E1000_WRITE_REG(hw, E1000_MRQC, E1000_MRQC_RSS_ENABLE_2Q | + E1000_MRQC_RSS_FIELD_IPV4_TCP | + E1000_MRQC_RSS_FIELD_IPV4 | + E1000_MRQC_RSS_FIELD_IPV6_TCP_EX | + E1000_MRQC_RSS_FIELD_IPV6_EX | + E1000_MRQC_RSS_FIELD_IPV6 | + E1000_MRQC_RSS_FIELD_IPV6_TCP); + } +#endif /* ** XXX TEMPORARY WORKAROUND: on some systems with 82573 ** long latencies are observed, like Lenovo X60. This @@ -4317,13 +4475,30 @@ em_initialize_receive_unit(struct adapte E1000_WRITE_REG(hw, E1000_RDT(i), rdt); } - /* Set PTHRESH for improved jumbo performance */ + /* + * Set PTHRESH for improved jumbo performance + * According to 10.2.5.11 of Intel 82574 Datasheet, + * RXDCTL(1) is written whenever RXDCTL(0) is written. + * Only write to RXDCTL(1) if there is a need for different + * settings. + */ if (((adapter->hw.mac.type == e1000_ich9lan) || (adapter->hw.mac.type == e1000_pch2lan) || (adapter->hw.mac.type == e1000_ich10lan)) && (if_getmtu(ifp) > ETHERMTU)) { u32 rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(0)); E1000_WRITE_REG(hw, E1000_RXDCTL(0), rxdctl | 3); + } else if ((adapter->hw.mac.type == e1000_82574) && + (if_getmtu(ifp) > ETHERMTU)) { + for (int i = 0; i < adapter->num_queues; i++) { + u32 rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(i)); + + rxdctl |= 0x20; /* PTHRESH */ + rxdctl |= 4 << 8; /* HTHRESH */ + rxdctl |= 4 << 16;/* WTHRESH */ + rxdctl |= 1 << 24; /* Switch to granularity */ + E1000_WRITE_REG(hw, E1000_RXDCTL(i), rxdctl); + } } if (adapter->hw.mac.type >= e1000_pch2lan) { @@ -4390,6 +4565,11 @@ em_rxeof(struct rx_ring *rxr, int count, EM_RX_LOCK(rxr); + /* Sync the ring */ + bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map, + BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); + + #ifdef DEV_NETMAP if (netmap_rx_irq(ifp, rxr->me, &processed)) { EM_RX_UNLOCK(rxr); @@ -4402,9 +4582,6 @@ em_rxeof(struct rx_ring *rxr, int count, if ((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0) break; - bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map, *** DIFF OUTPUT TRUNCATED AT 1000 LINES *** _______________________________________________ svn-src-head@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-head To unsubscribe, send any mail to "svn-src-head-unsubscr...@freebsd.org"
