diff options
Diffstat (limited to 'arch/x86/platform/uv/tlb_uv.c')
-rw-r--r-- | arch/x86/platform/uv/tlb_uv.c | 1484 |
1 files changed, 816 insertions, 668 deletions
diff --git a/arch/x86/platform/uv/tlb_uv.c b/arch/x86/platform/uv/tlb_uv.c index c58e0ea39ef5..68e467f69fec 100644 --- a/arch/x86/platform/uv/tlb_uv.c +++ b/arch/x86/platform/uv/tlb_uv.c @@ -1,7 +1,7 @@ /* * SGI UltraViolet TLB flush routines. * - * (c) 2008-2010 Cliff Wickman <cpw@sgi.com>, SGI. + * (c) 2008-2011 Cliff Wickman <cpw@sgi.com>, SGI. * * This code is released under the GNU General Public License version 2 or * later. @@ -35,6 +35,7 @@ static int timeout_base_ns[] = { 5242880, 167772160 }; + static int timeout_us; static int nobau; static int baudisabled; @@ -42,20 +43,70 @@ static spinlock_t disable_lock; static cycles_t congested_cycles; /* tunables: */ -static int max_bau_concurrent = MAX_BAU_CONCURRENT; -static int max_bau_concurrent_constant = MAX_BAU_CONCURRENT; -static int plugged_delay = PLUGGED_DELAY; -static int plugsb4reset = PLUGSB4RESET; -static int timeoutsb4reset = TIMEOUTSB4RESET; -static int ipi_reset_limit = IPI_RESET_LIMIT; -static int complete_threshold = COMPLETE_THRESHOLD; -static int congested_response_us = CONGESTED_RESPONSE_US; -static int congested_reps = CONGESTED_REPS; -static int congested_period = CONGESTED_PERIOD; +static int max_concurr = MAX_BAU_CONCURRENT; +static int max_concurr_const = MAX_BAU_CONCURRENT; +static int plugged_delay = PLUGGED_DELAY; +static int plugsb4reset = PLUGSB4RESET; +static int timeoutsb4reset = TIMEOUTSB4RESET; +static int ipi_reset_limit = IPI_RESET_LIMIT; +static int complete_threshold = COMPLETE_THRESHOLD; +static int congested_respns_us = CONGESTED_RESPONSE_US; +static int congested_reps = CONGESTED_REPS; +static int congested_period = CONGESTED_PERIOD; + +static struct tunables tunables[] = { + {&max_concurr, MAX_BAU_CONCURRENT}, /* must be [0] */ + {&plugged_delay, PLUGGED_DELAY}, + {&plugsb4reset, PLUGSB4RESET}, + {&timeoutsb4reset, TIMEOUTSB4RESET}, + {&ipi_reset_limit, IPI_RESET_LIMIT}, + {&complete_threshold, COMPLETE_THRESHOLD}, + {&congested_respns_us, CONGESTED_RESPONSE_US}, + {&congested_reps, CONGESTED_REPS}, + {&congested_period, CONGESTED_PERIOD} +}; + static struct dentry *tunables_dir; static struct dentry *tunables_file; -static int __init setup_nobau(char *arg) +/* these correspond to the statistics printed by ptc_seq_show() */ +static char *stat_description[] = { + "sent: number of shootdown messages sent", + "stime: time spent sending messages", + "numuvhubs: number of hubs targeted with shootdown", + "numuvhubs16: number times 16 or more hubs targeted", + "numuvhubs8: number times 8 or more hubs targeted", + "numuvhubs4: number times 4 or more hubs targeted", + "numuvhubs2: number times 2 or more hubs targeted", + "numuvhubs1: number times 1 hub targeted", + "numcpus: number of cpus targeted with shootdown", + "dto: number of destination timeouts", + "retries: destination timeout retries sent", + "rok: : destination timeouts successfully retried", + "resetp: ipi-style resource resets for plugs", + "resett: ipi-style resource resets for timeouts", + "giveup: fall-backs to ipi-style shootdowns", + "sto: number of source timeouts", + "bz: number of stay-busy's", + "throt: number times spun in throttle", + "swack: image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE", + "recv: shootdown messages received", + "rtime: time spent processing messages", + "all: shootdown all-tlb messages", + "one: shootdown one-tlb messages", + "mult: interrupts that found multiple messages", + "none: interrupts that found no messages", + "retry: number of retry messages processed", + "canc: number messages canceled by retries", + "nocan: number retries that found nothing to cancel", + "reset: number of ipi-style reset requests processed", + "rcan: number messages canceled by reset requests", + "disable: number times use of the BAU was disabled", + "enable: number times use of the BAU was re-enabled" +}; + +static int __init +setup_nobau(char *arg) { nobau = 1; return 0; @@ -63,7 +114,7 @@ static int __init setup_nobau(char *arg) early_param("nobau", setup_nobau); /* base pnode in this partition */ -static int uv_partition_base_pnode __read_mostly; +static int uv_base_pnode __read_mostly; /* position of pnode (which is nasid>>1): */ static int uv_nshift __read_mostly; static unsigned long uv_mmask __read_mostly; @@ -109,60 +160,52 @@ static int __init uvhub_to_first_apicid(int uvhub) * clear of the Timeout bit (as well) will free the resource. No reply will * be sent (the hardware will only do one reply per message). */ -static inline void uv_reply_to_message(struct msg_desc *mdp, - struct bau_control *bcp) +static void reply_to_message(struct msg_desc *mdp, struct bau_control *bcp) { unsigned long dw; - struct bau_payload_queue_entry *msg; + struct bau_pq_entry *msg; msg = mdp->msg; if (!msg->canceled) { - dw = (msg->sw_ack_vector << UV_SW_ACK_NPENDING) | - msg->sw_ack_vector; - uv_write_local_mmr( - UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, dw); + dw = (msg->swack_vec << UV_SW_ACK_NPENDING) | msg->swack_vec; + write_mmr_sw_ack(dw); } msg->replied_to = 1; - msg->sw_ack_vector = 0; + msg->swack_vec = 0; } /* * Process the receipt of a RETRY message */ -static inline void uv_bau_process_retry_msg(struct msg_desc *mdp, - struct bau_control *bcp) +static void bau_process_retry_msg(struct msg_desc *mdp, + struct bau_control *bcp) { int i; int cancel_count = 0; - int slot2; unsigned long msg_res; unsigned long mmr = 0; - struct bau_payload_queue_entry *msg; - struct bau_payload_queue_entry *msg2; - struct ptc_stats *stat; + struct bau_pq_entry *msg = mdp->msg; + struct bau_pq_entry *msg2; + struct ptc_stats *stat = bcp->statp; - msg = mdp->msg; - stat = bcp->statp; stat->d_retries++; /* * cancel any message from msg+1 to the retry itself */ for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) { - if (msg2 > mdp->va_queue_last) - msg2 = mdp->va_queue_first; + if (msg2 > mdp->queue_last) + msg2 = mdp->queue_first; if (msg2 == msg) break; - /* same conditions for cancellation as uv_do_reset */ + /* same conditions for cancellation as do_reset */ if ((msg2->replied_to == 0) && (msg2->canceled == 0) && - (msg2->sw_ack_vector) && ((msg2->sw_ack_vector & - msg->sw_ack_vector) == 0) && + (msg2->swack_vec) && ((msg2->swack_vec & + msg->swack_vec) == 0) && (msg2->sending_cpu == msg->sending_cpu) && (msg2->msg_type != MSG_NOOP)) { - slot2 = msg2 - mdp->va_queue_first; - mmr = uv_read_local_mmr - (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE); - msg_res = msg2->sw_ack_vector; + mmr = read_mmr_sw_ack(); + msg_res = msg2->swack_vec; /* * This is a message retry; clear the resources held * by the previous message only if they timed out. @@ -170,6 +213,7 @@ static inline void uv_bau_process_retry_msg(struct msg_desc *mdp, * situation to report. */ if (mmr & (msg_res << UV_SW_ACK_NPENDING)) { + unsigned long mr; /* * is the resource timed out? * make everyone ignore the cancelled message. @@ -177,10 +221,8 @@ static inline void uv_bau_process_retry_msg(struct msg_desc *mdp, msg2->canceled = 1; stat->d_canceled++; cancel_count++; - uv_write_local_mmr( - UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, - (msg_res << UV_SW_ACK_NPENDING) | - msg_res); + mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res; + write_mmr_sw_ack(mr); } } } @@ -192,20 +234,19 @@ static inline void uv_bau_process_retry_msg(struct msg_desc *mdp, * Do all the things a cpu should do for a TLB shootdown message. * Other cpu's may come here at the same time for this message. */ -static void uv_bau_process_message(struct msg_desc *mdp, - struct bau_control *bcp) +static void bau_process_message(struct msg_desc *mdp, + struct bau_control *bcp) { - int msg_ack_count; short socket_ack_count = 0; - struct ptc_stats *stat; - struct bau_payload_queue_entry *msg; + short *sp; + struct atomic_short *asp; + struct ptc_stats *stat = bcp->statp; + struct bau_pq_entry *msg = mdp->msg; struct bau_control *smaster = bcp->socket_master; /* * This must be a normal message, or retry of a normal message */ - msg = mdp->msg; - stat = bcp->statp; if (msg->address == TLB_FLUSH_ALL) { local_flush_tlb(); stat->d_alltlb++; @@ -222,30 +263,32 @@ static void uv_bau_process_message(struct msg_desc *mdp, * cpu number. */ if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master) - uv_bau_process_retry_msg(mdp, bcp); + bau_process_retry_msg(mdp, bcp); /* - * This is a sw_ack message, so we have to reply to it. + * This is a swack message, so we have to reply to it. * Count each responding cpu on the socket. This avoids * pinging the count's cache line back and forth between * the sockets. */ - socket_ack_count = atomic_add_short_return(1, (struct atomic_short *) - &smaster->socket_acknowledge_count[mdp->msg_slot]); + sp = &smaster->socket_acknowledge_count[mdp->msg_slot]; + asp = (struct atomic_short *)sp; + socket_ack_count = atom_asr(1, asp); if (socket_ack_count == bcp->cpus_in_socket) { + int msg_ack_count; /* * Both sockets dump their completed count total into * the message's count. */ smaster->socket_acknowledge_count[mdp->msg_slot] = 0; - msg_ack_count = atomic_add_short_return(socket_ack_count, - (struct atomic_short *)&msg->acknowledge_count); + asp = (struct atomic_short *)&msg->acknowledge_count; + msg_ack_count = atom_asr(socket_ack_count, asp); if (msg_ack_count == bcp->cpus_in_uvhub) { /* * All cpus in uvhub saw it; reply */ - uv_reply_to_message(mdp, bcp); + reply_to_message(mdp, bcp); } } @@ -268,62 +311,51 @@ static int uvhub_to_first_cpu(int uvhub) * Last resort when we get a large number of destination timeouts is * to clear resources held by a given cpu. * Do this with IPI so that all messages in the BAU message queue - * can be identified by their nonzero sw_ack_vector field. + * can be identified by their nonzero swack_vec field. * * This is entered for a single cpu on the uvhub. * The sender want's this uvhub to free a specific message's - * sw_ack resources. + * swack resources. */ -static void -uv_do_reset(void *ptr) +static void do_reset(void *ptr) { int i; - int slot; - int count = 0; - unsigned long mmr; - unsigned long msg_res; - struct bau_control *bcp; - struct reset_args *rap; - struct bau_payload_queue_entry *msg; - struct ptc_stats *stat; + struct bau_control *bcp = &per_cpu(bau_control, smp_processor_id()); + struct reset_args *rap = (struct reset_args *)ptr; + struct bau_pq_entry *msg; + struct ptc_stats *stat = bcp->statp; - bcp = &per_cpu(bau_control, smp_processor_id()); - rap = (struct reset_args *)ptr; - stat = bcp->statp; stat->d_resets++; - /* * We're looking for the given sender, and - * will free its sw_ack resource. + * will free its swack resource. * If all cpu's finally responded after the timeout, its * message 'replied_to' was set. */ - for (msg = bcp->va_queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) { - /* uv_do_reset: same conditions for cancellation as - uv_bau_process_retry_msg() */ + for (msg = bcp->queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) { + unsigned long msg_res; + /* do_reset: same conditions for cancellation as + bau_process_retry_msg() */ if ((msg->replied_to == 0) && (msg->canceled == 0) && (msg->sending_cpu == rap->sender) && - (msg->sw_ack_vector) && + (msg->swack_vec) && (msg->msg_type != MSG_NOOP)) { + unsigned long mmr; + unsigned long mr; /* * make everyone else ignore this message */ msg->canceled = 1; - slot = msg - bcp->va_queue_first; - count++; /* * only reset the resource if it is still pending */ - mmr = uv_read_local_mmr - (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE); - msg_res = msg->sw_ack_vector; + mmr = read_mmr_sw_ack(); + msg_res = msg->swack_vec; + mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res; if (mmr & msg_res) { stat->d_rcanceled++; - uv_write_local_mmr( - UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, - (msg_res << UV_SW_ACK_NPENDING) | - msg_res); + write_mmr_sw_ack(mr); } } } @@ -334,39 +366,38 @@ uv_do_reset(void *ptr) * Use IPI to get all target uvhubs to release resources held by * a given sending cpu number. */ -static void uv_reset_with_ipi(struct bau_target_uvhubmask *distribution, - int sender) +static void reset_with_ipi(struct bau_targ_hubmask *distribution, int sender) { int uvhub; - int cpu; + int maskbits; cpumask_t mask; struct reset_args reset_args; reset_args.sender = sender; - cpus_clear(mask); /* find a single cpu for each uvhub in this distribution mask */ - for (uvhub = 0; - uvhub < sizeof(struct bau_target_uvhubmask) * BITSPERBYTE; - uvhub++) { + maskbits = sizeof(struct bau_targ_hubmask) * BITSPERBYTE; + for (uvhub = 0; uvhub < maskbits; uvhub++) { + int cpu; if (!bau_uvhub_isset(uvhub, distribution)) continue; /* find a cpu for this uvhub */ cpu = uvhub_to_first_cpu(uvhub); cpu_set(cpu, mask); } - /* IPI all cpus; Preemption is already disabled */ - smp_call_function_many(&mask, uv_do_reset, (void *)&reset_args, 1); + + /* IPI all cpus; preemption is already disabled */ + smp_call_function_many(&mask, do_reset, (void *)&reset_args, 1); return; } -static inline unsigned long -cycles_2_us(unsigned long long cyc) +static inline unsigned long cycles_2_us(unsigned long long cyc) { unsigned long long ns; unsigned long us; - ns = (cyc * per_cpu(cyc2ns, smp_processor_id())) - >> CYC2NS_SCALE_FACTOR; + int cpu = smp_processor_id(); + + ns = (cyc * per_cpu(cyc2ns, cpu)) >> CYC2NS_SCALE_FACTOR; us = ns / 1000; return us; } @@ -376,56 +407,56 @@ cycles_2_us(unsigned long long cyc) * leaves uvhub_quiesce set so that no new broadcasts are started by * bau_flush_send_and_wait() */ -static inline void -quiesce_local_uvhub(struct bau_control *hmaster) +static inline void quiesce_local_uvhub(struct bau_control *hmaster) { - atomic_add_short_return(1, (struct atomic_short *) - &hmaster->uvhub_quiesce); + atom_asr(1, (struct atomic_short *)&hmaster->uvhub_quiesce); } /* * mark this quiet-requestor as done */ -static inline void -end_uvhub_quiesce(struct bau_control *hmaster) +static inline void end_uvhub_quiesce(struct bau_control *hmaster) { - atomic_add_short_return(-1, (struct atomic_short *) - &hmaster->uvhub_quiesce); + atom_asr(-1, (struct atomic_short *)&hmaster->uvhub_quiesce); +} + +static unsigned long uv1_read_status(unsigned long mmr_offset, int right_shift) +{ + unsigned long descriptor_status; + + descriptor_status = uv_read_local_mmr(mmr_offset); + descriptor_status >>= right_shift; + descriptor_status &= UV_ACT_STATUS_MASK; + return descriptor_status; } /* * Wait for completion of a broadcast software ack message * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP */ -static int uv_wait_completion(struct bau_desc *bau_desc, - unsigned long mmr_offset, int right_shift, int this_cpu, - struct bau_control *bcp, struct bau_control *smaster, long try) +static int uv1_wait_completion(struct bau_desc *bau_desc, + unsigned long mmr_offset, int right_shift, + struct bau_control *bcp, long try) { unsigned long descriptor_status; - cycles_t ttime; + cycles_t ttm; struct ptc_stats *stat = bcp->statp; - struct bau_control *hmaster; - - hmaster = bcp->uvhub_master; + descriptor_status = uv1_read_status(mmr_offset, right_shift); /* spin on the status MMR, waiting for it to go idle */ - while ((descriptor_status = (((unsigned long) - uv_read_local_mmr(mmr_offset) >> - right_shift) & UV_ACT_STATUS_MASK)) != - DESC_STATUS_IDLE) { + while ((descriptor_status != DS_IDLE)) { /* - * Our software ack messages may be blocked because there are - * no swack resources available. As long as none of them - * has timed out hardware will NACK our message and its - * state will stay IDLE. + * Our software ack messages may be blocked because + * there are no swack resources available. As long + * as none of them has timed out hardware will NACK + * our message and its state will stay IDLE. */ - if (descriptor_status == DESC_STATUS_SOURCE_TIMEOUT) { + if (descriptor_status == DS_SOURCE_TIMEOUT) { stat->s_stimeout++; return FLUSH_GIVEUP; - } else if (descriptor_status == - DESC_STATUS_DESTINATION_TIMEOUT) { + } else if (descriptor_status == DS_DESTINATION_TIMEOUT) { stat->s_dtimeout++; - ttime = get_cycles(); + ttm = get_cycles(); /* * Our retries may be blocked by all destination @@ -433,8 +464,7 @@ static int uv_wait_completion(struct bau_desc *bau_desc, * pending. In that case hardware returns the * ERROR that looks like a destination timeout. */ - if (cycles_2_us(ttime - bcp->send_message) < - timeout_us) { + if (cycles_2_us(ttm - bcp->send_message) < timeout_us) { bcp->conseccompletes = 0; return FLUSH_RETRY_PLUGGED; } @@ -447,80 +477,160 @@ static int uv_wait_completion(struct bau_desc *bau_desc, */ cpu_relax(); } + descriptor_status = uv1_read_status(mmr_offset, right_shift); } bcp->conseccompletes++; return FLUSH_COMPLETE; } -static inline cycles_t -sec_2_cycles(unsigned long sec) +/* + * UV2 has an extra bit of status in the ACTIVATION_STATUS_2 register. + */ +static unsigned long uv2_read_status(unsigned long offset, int rshft, int cpu) { - unsigned long ns; - cycles_t cyc; + unsigned long descriptor_status; + unsigned long descriptor_status2; - ns = sec * 1000000000; - cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id())); - return cyc; + descriptor_status = ((read_lmmr(offset) >> rshft) & UV_ACT_STATUS_MASK); + descriptor_status2 = (read_mmr_uv2_status() >> cpu) & 0x1UL; + descriptor_status = (descriptor_status << 1) | descriptor_status2; + return descriptor_status; +} + +static int uv2_wait_completion(struct bau_desc *bau_desc, + unsigned long mmr_offset, int right_shift, + struct bau_control *bcp, long try) +{ + unsigned long descriptor_stat; + cycles_t ttm; + int cpu = bcp->uvhub_cpu; + struct ptc_stats *stat = bcp->statp; + + descriptor_stat = uv2_read_status(mmr_offset, right_shift, cpu); + + /* spin on the status MMR, waiting for it to go idle */ + while (descriptor_stat != UV2H_DESC_IDLE) { + /* + * Our software ack messages may be blocked because + * there are no swack resources available. As long + * as none of them has timed out hardware will NACK + * our message and its state will stay IDLE. + */ + if ((descriptor_stat == UV2H_DESC_SOURCE_TIMEOUT) || + (descriptor_stat == UV2H_DESC_DEST_STRONG_NACK) || + (descriptor_stat == UV2H_DESC_DEST_PUT_ERR)) { + stat->s_stimeout++; + return FLUSH_GIVEUP; + } else if (descriptor_stat == UV2H_DESC_DEST_TIMEOUT) { + stat->s_dtimeout++; + ttm = get_cycles(); + /* + * Our retries may be blocked by all destination + * swack resources being consumed, and a timeout + * pending. In that case hardware returns the + * ERROR that looks like a destination timeout. + */ + if (cycles_2_us(ttm - bcp->send_message) < timeout_us) { + bcp->conseccompletes = 0; + return FLUSH_RETRY_PLUGGED; + } + bcp->conseccompletes = 0; + return FLUSH_RETRY_TIMEOUT; + } else { + /* + * descriptor_stat is still BUSY + */ + cpu_relax(); + } + descriptor_stat = uv2_read_status(mmr_offset, right_shift, cpu); + } + bcp->conseccompletes++; + return FLUSH_COMPLETE; } /* - * conditionally add 1 to *v, unless *v is >= u - * return 0 if we cannot add 1 to *v because it is >= u - * return 1 if we can add 1 to *v because it is < u - * the add is atomic - * - * This is close to atomic_add_unless(), but this allows the 'u' value - * to be lowered below the current 'v'. atomic_add_unless can only stop - * on equal. + * There are 2 status registers; each and array[32] of 2 bits. Set up for + * which register to read and position in that register based on cpu in + * current hub. */ -static inline int atomic_inc_unless_ge(spinlock_t *lock, atomic_t *v, int u) +static int wait_completion(struct bau_desc *bau_desc, + struct bau_control *bcp, long try) { - spin_lock(lock); - if (atomic_read(v) >= u) { - spin_unlock(lock); - return 0; + int right_shift; + unsigned long mmr_offset; + int cpu = bcp->uvhub_cpu; + + if (cpu < UV_CPUS_PER_AS) { + mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0; + right_shift = cpu * UV_ACT_STATUS_SIZE; + } else { + mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1; + right_shift = ((cpu - UV_CPUS_PER_AS) * UV_ACT_STATUS_SIZE); } - atomic_inc(v); - spin_unlock(lock); - return 1; + + if (is_uv1_hub()) + return uv1_wait_completion(bau_desc, mmr_offset, right_shift, + bcp, try); + else + return uv2_wait_completion(bau_desc, mmr_offset, right_shift, + bcp, try); +} + +static inline cycles_t sec_2_cycles(unsigned long sec) +{ + unsigned long ns; + cycles_t cyc; + + ns = sec * 1000000000; + cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id())); + return cyc; } /* - * Our retries are blocked by all destination swack resources being + * Our retries are blocked by all destination sw ack resources being * in use, and a timeout is pending. In that case hardware immediately * returns the ERROR that looks like a destination timeout. */ -static void -destination_plugged(struct bau_desc *bau_desc, struct bau_control *bcp, +static void destination_plugged(struct bau_desc *bau_desc, + struct bau_control *bcp, struct bau_control *hmaster, struct ptc_stats *stat) { udelay(bcp->plugged_delay); bcp->plugged_tries++; + if (bcp->plugged_tries >= bcp->plugsb4reset) { bcp->plugged_tries = 0; + quiesce_local_uvhub(hmaster); + spin_lock(&hmaster->queue_lock); - uv_reset_with_ipi(&bau_desc->distribution, bcp->cpu); + reset_with_ipi(&bau_desc->distribution, bcp->cpu); spin_unlock(&hmaster->queue_lock); + end_uvhub_quiesce(hmaster); + bcp->ipi_attempts++; stat->s_resets_plug++; } } -static void -destination_timeout(struct bau_desc *bau_desc, struct bau_control *bcp, - struct bau_control *hmaster, struct ptc_stats *stat) +static void destination_timeout(struct bau_desc *bau_desc, + struct bau_control *bcp, struct bau_control *hmaster, + struct ptc_stats *stat) { - hmaster->max_bau_concurrent = 1; + hmaster->max_concurr = 1; bcp->timeout_tries++; if (bcp->timeout_tries >= bcp->timeoutsb4reset) { bcp->timeout_tries = 0; + quiesce_local_uvhub(hmaster); + spin_lock(&hmaster->queue_lock); - uv_reset_with_ipi(&bau_desc->distribution, bcp->cpu); + reset_with_ipi(&bau_desc->distribution, bcp->cpu); spin_unlock(&hmaster->queue_lock); + end_uvhub_quiesce(hmaster); + bcp->ipi_attempts++; stat->s_resets_timeout++; } @@ -530,34 +640,104 @@ destination_timeout(struct bau_desc *bau_desc, struct bau_control *bcp, * Completions are taking a very long time due to a congested numalink * network. */ -static void -disable_for_congestion(struct bau_control *bcp, struct ptc_stats *stat) +static void disable_for_congestion(struct bau_control *bcp, + struct ptc_stats *stat) { - int tcpu; - struct bau_control *tbcp; - /* let only one cpu do this disabling */ spin_lock(&disable_lock); + if (!baudisabled && bcp->period_requests && ((bcp->period_time / bcp->period_requests) > congested_cycles)) { + int tcpu; + struct bau_control *tbcp; /* it becomes this cpu's job to turn on the use of the BAU again */ baudisabled = 1; bcp->set_bau_off = 1; - bcp->set_bau_on_time = get_cycles() + - sec_2_cycles(bcp->congested_period); + bcp->set_bau_on_time = get_cycles(); + bcp->set_bau_on_time += sec_2_cycles(bcp->cong_period); stat->s_bau_disabled++; for_each_present_cpu(tcpu) { tbcp = &per_cpu(bau_control, tcpu); - tbcp->baudisabled = 1; + tbcp->baudisabled = 1; } } + spin_unlock(&disable_lock); } -/** - * uv_flush_send_and_wait - * +static void count_max_concurr(int stat, struct bau_control *bcp, + struct bau_control *hmaster) +{ + bcp->plugged_tries = 0; + bcp->timeout_tries = 0; + if (stat != FLUSH_COMPLETE) + return; + if (bcp->conseccompletes <= bcp->complete_threshold) + return; + if (hmaster->max_concurr >= hmaster->max_concurr_const) + return; + hmaster->max_concurr++; +} + +static void record_send_stats(cycles_t time1, cycles_t time2, + struct bau_control *bcp, struct ptc_stats *stat, + int completion_status, int try) +{ + cycles_t elapsed; + + if (time2 > time1) { + elapsed = time2 - time1; + stat->s_time += elapsed; + + if ((completion_status == FLUSH_COMPLETE) && (try == 1)) { + bcp->period_requests++; + bcp->period_time += elapsed; + if ((elapsed > congested_cycles) && + (bcp->period_requests > bcp->cong_reps)) + disable_for_congestion(bcp, stat); + } + } else + stat->s_requestor--; + + if (completion_status == FLUSH_COMPLETE && try > 1) + stat->s_retriesok++; + else if (completion_status == FLUSH_GIVEUP) + stat->s_giveup++; +} + +/* + * Because of a uv1 hardware bug only a limited number of concurrent + * requests can be made. + */ +static void uv1_throttle(struct bau_control *hmaster, struct ptc_stats *stat) +{ + spinlock_t *lock = &hmaster->uvhub_lock; + atomic_t *v; + + v = &hmaster->active_descriptor_count; + if (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)) { + stat->s_throttles++; + do { + cpu_relax(); + } while (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)); + } +} + +/* + * Handle the completion status of a message send. + */ +static void handle_cmplt(int completion_status, struct bau_desc *bau_desc, + struct bau_control *bcp, struct bau_control *hmaster, + struct ptc_stats *stat) +{ + if (completion_status == FLUSH_RETRY_PLUGGED) + destination_plugged(bau_desc, bcp, hmaster, stat); + else if (completion_status == FLUSH_RETRY_TIMEOUT) + destination_timeout(bau_desc, bcp, hmaster, stat); +} + +/* * Send a broadcast and wait for it to complete. * * The flush_mask contains the cpus the broadcast is to be sent to including @@ -568,44 +748,23 @@ disable_for_congestion(struct bau_control *bcp, struct ptc_stats *stat) * returned to the kernel. */ int uv_flush_send_and_wait(struct bau_desc *bau_desc, - struct cpumask *flush_mask, struct bau_control *bcp) + struct cpumask *flush_mask, struct bau_control *bcp) { - int right_shift; - int completion_status = 0; int seq_number = 0; + int completion_stat = 0; long try = 0; - int cpu = bcp->uvhub_cpu; - int this_cpu = bcp->cpu; - unsigned long mmr_offset; unsigned long index; cycles_t time1; cycles_t time2; - cycles_t elapsed; struct ptc_stats *stat = bcp->statp; - struct bau_control *smaster = bcp->socket_master; struct bau_control *hmaster = bcp->uvhub_master; - if (!atomic_inc_unless_ge(&hmaster->uvhub_lock, - &hmaster->active_descriptor_count, - hmaster->max_bau_concurrent)) { - stat->s_throttles++; - do { - cpu_relax(); - } while (!atomic_inc_unless_ge(&hmaster->uvhub_lock, - &hmaster->active_descriptor_count, - hmaster->max_bau_concurrent)); - } + if (is_uv1_hub()) + uv1_throttle(hmaster, stat); + while (hmaster->uvhub_quiesce) cpu_relax(); - if (cpu < UV_CPUS_PER_ACT_STATUS) { - mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0; - right_shift = cpu * UV_ACT_STATUS_SIZE; - } else { - mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1; - right_shift = - ((cpu - UV_CPUS_PER_ACT_STATUS) * UV_ACT_STATUS_SIZE); - } time1 = get_cycles(); do { if (try == 0) { @@ -615,64 +774,134 @@ int uv_flush_send_and_wait(struct bau_desc *bau_desc, bau_desc->header.msg_type = MSG_RETRY; stat->s_retry_messages++; } + bau_desc->header.sequence = seq_number; - index = (1UL << UVH_LB_BAU_SB_ACTIVATION_CONTROL_PUSH_SHFT) | - bcp->uvhub_cpu; + index = (1UL << AS_PUSH_SHIFT) | bcp->uvhub_cpu; bcp->send_message = get_cycles(); - uv_write_local_mmr(UVH_LB_BAU_SB_ACTIVATION_CONTROL, index); + + write_mmr_activation(index); + try++; - completion_status = uv_wait_completion(bau_desc, mmr_offset, - right_shift, this_cpu, bcp, smaster, try); + completion_stat = wait_completion(bau_desc, bcp, try); + + handle_cmplt(completion_stat, bau_desc, bcp, hmaster, stat); - if (completion_status == FLUSH_RETRY_PLUGGED) { - destination_plugged(bau_desc, bcp, hmaster, stat); - } else if (completion_status == FLUSH_RETRY_TIMEOUT) { - destination_timeout(bau_desc, bcp, hmaster, stat); - } if (bcp->ipi_attempts >= bcp->ipi_reset_limit) { bcp->ipi_attempts = 0; - completion_status = FLUSH_GIVEUP; + completion_stat = FLUSH_GIVEUP; break; } cpu_relax(); - } while ((completion_status == FLUSH_RETRY_PLUGGED) || - (completion_status == FLUSH_RETRY_TIMEOUT)); + } while ((completion_stat == FLUSH_RETRY_PLUGGED) || + (completion_stat == FLUSH_RETRY_TIMEOUT)); + time2 = get_cycles(); - bcp->plugged_tries = 0; - bcp->timeout_tries = 0; - if ((completion_status == FLUSH_COMPLETE) && - (bcp->conseccompletes > bcp->complete_threshold) && - (hmaster->max_bau_concurrent < - hmaster->max_bau_concurrent_constant)) - hmaster->max_bau_concurrent++; + + count_max_concurr(completion_stat, bcp, hmaster); + while (hmaster->uvhub_quiesce) cpu_relax(); + atomic_dec(&hmaster->active_descriptor_count); - if (time2 > time1) { - elapsed = time2 - time1; - stat->s_time += elapsed; - if ((completion_status == FLUSH_COMPLETE) && (try == 1)) { - bcp->period_requests++; - bcp->period_time += elapsed; - if ((elapsed > congested_cycles) && - (bcp->period_requests > bcp->congested_reps)) { - disable_for_congestion(bcp, stat); + + record_send_stats(time1, time2, bcp, stat, completion_stat, try); + + if (completion_stat == FLUSH_GIVEUP) + return 1; + return 0; +} + +/* + * The BAU is disabled. When the disabled time period has expired, the cpu + * that disabled it must re-enable it. + * Return 0 if it is re-enabled for all cpus. + */ +static int check_enable(struct bau_control *bcp, struct ptc_stats *stat) +{ + int tcpu; + struct bau_control *tbcp; + + if (bcp->set_bau_off) { + if (get_cycles() >= bcp->set_bau_on_time) { + stat->s_bau_reenabled++; + baudisabled = 0; + for_each_present_cpu(tcpu) { + tbcp = &per_cpu(bau_control, tcpu); + tbcp->baudisabled = 0; + tbcp->period_requests = 0; + tbcp->period_time = 0; } + return 0; } + } + return -1; +} + +static void record_send_statistics(struct ptc_stats *stat, int locals, int hubs, + int remotes, struct bau_desc *bau_desc) +{ + stat->s_requestor++; + stat->s_ntargcpu += remotes + locals; + stat->s_ntargremotes += remotes; + stat->s_ntarglocals += locals; + + /* uvhub statistics */ + hubs = bau_uvhub_weight(&bau_desc->distribution); + if (locals) { + stat->s_ntarglocaluvhub++; + stat->s_ntargremoteuvhub += (hubs - 1); } else - stat->s_requestor--; - if (completion_status == FLUSH_COMPLETE && try > 1) - stat->s_retriesok++; - else if (completion_status == FLUSH_GIVEUP) { - stat->s_giveup++; - return 1; + stat->s_ntargremoteuvhub += hubs; + + stat->s_ntarguvhub += hubs; + + if (hubs >= 16) + stat->s_ntarguvhub16++; + else if (hubs >= 8) + stat->s_ntarguvhub8++; + else if (hubs >= 4) + stat->s_ntarguvhub4++; + else if (hubs >= 2) + stat->s_ntarguvhub2++; + else + stat->s_ntarguvhub1++; +} + +/* + * Translate a cpu mask to the uvhub distribution mask in the BAU + * activation descriptor. + */ +static int set_distrib_bits(struct cpumask *flush_mask, struct bau_control *bcp, + struct bau_desc *bau_desc, int *localsp, int *remotesp) +{ + int cpu; + int pnode; + int cnt = 0; + struct hub_and_pnode *hpp; + + for_each_cpu(cpu, flush_mask) { + /* + * The distribution vector is a bit map of pnodes, relative + * to the partition base pnode (and the partition base nasid + * in the header). + * Translate cpu to pnode and hub using a local memory array. + */ + hpp = &bcp->socket_master->thp[cpu]; + pnode = hpp->pnode - bcp->partition_base_pnode; + bau_uvhub_set(pnode, &bau_desc->distribution); + cnt++; + if (hpp->uvhub == bcp->uvhub) + (*localsp)++; + else + (*remotesp)++; } + if (!cnt) + return 1; return 0; } -/** - * uv_flush_tlb_others - globally purge translation cache of a virtual - * address or all TLB's +/* + * globally purge translation cache of a virtual address or all TLB's * @cpumask: mask of all cpu's in which the address is to be removed * @mm: mm_struct containing virtual address range * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu) @@ -696,20 +925,16 @@ int uv_flush_send_and_wait(struct bau_desc *bau_desc, * done. The returned pointer is valid till preemption is re-enabled. */ const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask, - struct mm_struct *mm, - unsigned long va, unsigned int cpu) + struct mm_struct *mm, unsigned long va, + unsigned int cpu) { int locals = 0; int remotes = 0; int hubs = 0; - int tcpu; - int tpnode; struct bau_desc *bau_desc; struct cpumask *flush_mask; struct ptc_stats *stat; struct bau_control *bcp; - struct bau_control *tbcp; - struct hub_and_pnode *hpp; /* kernel was booted 'nobau' */ if (nobau) @@ -720,20 +945,8 @@ const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask, /* bau was disabled due to slow response */ if (bcp->baudisabled) { - /* the cpu that disabled it must re-enable it */ - if (bcp->set_bau_off) { - if (get_cycles() >= bcp->set_bau_on_time) { - stat->s_bau_reenabled++; - baudisabled = 0; - for_each_present_cpu(tcpu) { - tbcp = &per_cpu(bau_control, tcpu); - tbcp->baudisabled = 0; - tbcp->period_requests = 0; - tbcp->period_time = 0; - } - } - } - return cpumask; + if (check_enable(bcp, stat)) + return cpumask; } /* @@ -744,59 +957,20 @@ const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask, flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu); /* don't actually do a shootdown of the local cpu */ cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu)); + if (cpu_isset(cpu, *cpumask)) stat->s_ntargself++; |