unsigned long mmr_offset, int right_shift, int this_cpu,
struct bau_control *bcp, struct bau_control *smaster, long try)
{
- int relaxes = 0;
unsigned long descriptor_status;
- unsigned long mmr;
- unsigned long mask;
cycles_t ttime;
struct ptc_stats *stat = bcp->statp;
struct bau_control *hmaster;
return 1;
}
+/*
+ * Our retries are blocked by all destination swack 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,
+ 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);
+ 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)
+{
+ hmaster->max_bau_concurrent = 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);
+ spin_unlock(&hmaster->queue_lock);
+ end_uvhub_quiesce(hmaster);
+ bcp->ipi_attempts++;
+ stat->s_resets_timeout++;
+ }
+}
+
/*
* Completions are taking a very long time due to a congested numalink
* network.
*
* Send a broadcast and wait for it to complete.
*
- * The flush_mask contains the cpus the broadcast is to be sent to, plus
+ * The flush_mask contains the cpus the broadcast is to be sent to including
* cpus that are on the local uvhub.
*
- * Returns NULL if all flushing represented in the mask was done. The mask
- * is zeroed.
- * Returns @flush_mask if some remote flushing remains to be done. The
- * mask will have some bits still set, representing any cpus on the local
- * uvhub (not current cpu) and any on remote uvhubs if the broadcast failed.
+ * Returns 0 if all flushing represented in the mask was done.
+ * Returns 1 if it gives up entirely and the original cpu mask is to be
+ * returned to the kernel.
*/
-const struct cpumask *uv_flush_send_and_wait(struct bau_desc *bau_desc,
- struct cpumask *flush_mask,
- struct bau_control *bcp)
+int uv_flush_send_and_wait(struct bau_desc *bau_desc,
+ struct cpumask *flush_mask, struct bau_control *bcp)
{
int right_shift;
- int uvhub;
- int bit;
int completion_status = 0;
int seq_number = 0;
long try = 0;
int cpu = bcp->uvhub_cpu;
int this_cpu = bcp->cpu;
- int this_uvhub = bcp->uvhub;
unsigned long mmr_offset;
unsigned long index;
cycles_t time1;
struct bau_control *smaster = bcp->socket_master;
struct bau_control *hmaster = bcp->uvhub_master;
- /*
- * Spin here while there are hmaster->max_bau_concurrent or more active
- * descriptors. This is the per-uvhub 'throttle'.
- */
if (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
&hmaster->active_descriptor_count,
hmaster->max_bau_concurrent)) {
&hmaster->active_descriptor_count,
hmaster->max_bau_concurrent));
}
-
while (hmaster->uvhub_quiesce)
cpu_relax();
index = (1UL << UVH_LB_BAU_SB_ACTIVATION_CONTROL_PUSH_SHFT) |
bcp->uvhub_cpu;
bcp->send_message = get_cycles();
-
uv_write_local_mmr(UVH_LB_BAU_SB_ACTIVATION_CONTROL, index);
-
try++;
completion_status = uv_wait_completion(bau_desc, mmr_offset,
right_shift, this_cpu, bcp, smaster, try);
if (completion_status == FLUSH_RETRY_PLUGGED) {
- /*
- * Our retries may be blocked by all destination swack
- * resources being consumed, and a timeout pending. In
- * that case hardware immediately returns the ERROR
- * that looks like a destination timeout.
- */
- 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,
- this_cpu);
- spin_unlock(&hmaster->queue_lock);
- end_uvhub_quiesce(hmaster);
- bcp->ipi_attempts++;
- stat->s_resets_plug++;
- }
+ destination_plugged(bau_desc, bcp, hmaster, stat);
} else if (completion_status == FLUSH_RETRY_TIMEOUT) {
- hmaster->max_bau_concurrent = 1;
- bcp->timeout_tries++;
- udelay(TIMEOUT_DELAY);
- 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,
- this_cpu);
- spin_unlock(&hmaster->queue_lock);
- end_uvhub_quiesce(hmaster);
- bcp->ipi_attempts++;
- stat->s_resets_timeout++;
- }
+ destination_timeout(bau_desc, bcp, hmaster, stat);
}
if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
bcp->ipi_attempts = 0;
} while ((completion_status == FLUSH_RETRY_PLUGGED) ||
(completion_status == 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++;
-
- /*
- * hold any cpu not timing out here; no other cpu currently held by
- * the 'throttle' should enter the activation code
- */
while (hmaster->uvhub_quiesce)
cpu_relax();
atomic_dec(&hmaster->active_descriptor_count);
-
- /* guard against cycles wrap */
if (time2 > time1) {
elapsed = time2 - time1;
stat->s_time += elapsed;
}
}
} else
- stat->s_requestor--; /* don't count this one */
+ stat->s_requestor--;
if (completion_status == FLUSH_COMPLETE && try > 1)
stat->s_retriesok++;
else if (completion_status == FLUSH_GIVEUP) {
- /*
- * Cause the caller to do an IPI-style TLB shootdown on
- * the target cpu's, all of which are still in the mask.
- */
stat->s_giveup++;
- return flush_mask;
- }
-
- /*
- * Success, so clear the remote cpu's from the mask so we don't
- * use the IPI method of shootdown on them.
- */
- for_each_cpu(bit, flush_mask) {
- uvhub = uv_cpu_to_blade_id(bit);
- if (uvhub == this_uvhub)
- continue;
- cpumask_clear_cpu(bit, flush_mask);
+ return 1;
}
- if (!cpumask_empty(flush_mask))
- return flush_mask;
-
- return NULL;
+ return 0;
}
/**
struct mm_struct *mm,
unsigned long va, unsigned int cpu)
{
- int remotes;
int tcpu;
int uvhub;
int locals = 0;
+ int remotes = 0;
+ int hubs = 0;
struct bau_desc *bau_desc;
struct cpumask *flush_mask;
struct ptc_stats *stat;
/*
* Each sending cpu has a per-cpu mask which it fills from the caller's
- * cpu mask. Only remote cpus are converted to uvhubs and copied.
+ * cpu mask. All cpus are converted to uvhubs and copied to the
+ * activation descriptor.
*/
flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
- /*
- * copy cpumask to flush_mask, removing current cpu
- * (current cpu should already have been flushed by the caller and
- * should never be returned if we return flush_mask)
- */
+ /* don't actually do a shootdown of the local cpu */
cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
if (cpu_isset(cpu, *cpumask))
- locals++; /* current cpu was targeted */
+ stat->s_ntargself++;
bau_desc = bcp->descriptor_base;
bau_desc += UV_ITEMS_PER_DESCRIPTOR * bcp->uvhub_cpu;
-
bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
- remotes = 0;
+
+ /* cpu statistics */
for_each_cpu(tcpu, flush_mask) {
uvhub = uv_cpu_to_blade_id(tcpu);
- if (uvhub == bcp->uvhub) {
- locals++;
- continue;
- }
bau_uvhub_set(uvhub, &bau_desc->distribution);
- remotes++;
- }
- if (remotes == 0) {
- /*
- * No off_hub flushing; return status for local hub.
- * Return the caller's mask if all were local (the current
- * cpu may be in that mask).
- */
- if (locals)
- return cpumask;
+ if (uvhub == bcp->uvhub)
+ locals++;
else
- return NULL;
+ remotes++;
}
+ if ((locals + remotes) == 0)
+ return NULL;
stat->s_requestor++;
- stat->s_ntargcpu += remotes;
+ stat->s_ntargcpu += remotes + locals;
+ stat->s_ntargremotes += remotes;
+ stat->s_ntarglocals += locals;
remotes = bau_uvhub_weight(&bau_desc->distribution);
- stat->s_ntarguvhub += remotes;
- if (remotes >= 16)
+
+ /* uvhub statistics */
+ hubs = bau_uvhub_weight(&bau_desc->distribution);
+ if (locals) {
+ stat->s_ntarglocaluvhub++;
+ stat->s_ntargremoteuvhub += (hubs - 1);
+ } else
+ stat->s_ntargremoteuvhub += hubs;
+ stat->s_ntarguvhub += hubs;
+ if (hubs >= 16)
stat->s_ntarguvhub16++;
- else if (remotes >= 8)
+ else if (hubs >= 8)
stat->s_ntarguvhub8++;
- else if (remotes >= 4)
+ else if (hubs >= 4)
stat->s_ntarguvhub4++;
- else if (remotes >= 2)
+ else if (hubs >= 2)
stat->s_ntarguvhub2++;
else
stat->s_ntarguvhub1++;
bau_desc->payload.sending_cpu = cpu;
/*
- * uv_flush_send_and_wait returns null if all cpu's were messaged, or
- * the adjusted flush_mask if any cpu's were not messaged.
+ * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
+ * or 1 if it gave up and the original cpumask should be returned.
*/
- return uv_flush_send_and_wait(bau_desc, flush_mask, bcp);
+ if (!uv_flush_send_and_wait(bau_desc, flush_mask, bcp))
+ return NULL;
+ else
+ return cpumask;
}
/*
if (!cpu) {
seq_printf(file,
- "# cpu sent stime numuvhubs numuvhubs16 numuvhubs8 ");
+ "# cpu sent stime self locals remotes ncpus localhub ");
seq_printf(file,
- "numuvhubs4 numuvhubs2 numuvhubs1 numcpus dto ");
+ "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
+ seq_printf(file,
+ "numuvhubs4 numuvhubs2 numuvhubs1 dto ");
seq_printf(file,
"retries rok resetp resett giveup sto bz throt ");
seq_printf(file,
seq_printf(file,
"cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
cpu, stat->s_requestor, cycles_2_us(stat->s_time),
- stat->s_ntarguvhub, stat->s_ntarguvhub16,
+ stat->s_ntargself, stat->s_ntarglocals,
+ stat->s_ntargremotes, stat->s_ntargcpu,
+ stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
+ stat->s_ntarguvhub, stat->s_ntarguvhub16);
+ seq_printf(file, "%ld %ld %ld %ld %ld ",
stat->s_ntarguvhub8, stat->s_ntarguvhub4,
stat->s_ntarguvhub2, stat->s_ntarguvhub1,
- stat->s_ntargcpu, stat->s_dtimeout);
+ stat->s_dtimeout);
seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
stat->s_retry_messages, stat->s_retriesok,
stat->s_resets_plug, stat->s_resets_timeout,
/*
* initialize the bau_control structure for each cpu
*/
-static void uv_init_per_cpu(int nuvhubs)
+static void __init uv_init_per_cpu(int nuvhubs)
{
int i;
int cpu;
int pnode;
int uvhub;
+ int have_hmaster;
short socket = 0;
unsigned short socket_mask;
- unsigned int uvhub_mask;
+ unsigned char *uvhub_mask;
struct bau_control *bcp;
struct uvhub_desc *bdp;
struct socket_desc *sdp;
uvhub_descs = (struct uvhub_desc *)
kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
+ uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
for_each_present_cpu(cpu) {
bcp = &per_cpu(bau_control, cpu);
memset(bcp, 0, sizeof(struct bau_control));
pnode = uv_cpu_hub_info(cpu)->pnode;
uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
- uvhub_mask |= (1 << uvhub);
+ *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
bdp = &uvhub_descs[uvhub];
bdp->num_cpus++;
bdp->uvhub = uvhub;
bdp->pnode = pnode;
/* kludge: 'assuming' one node per socket, and assuming that
disabling a socket just leaves a gap in node numbers */
- socket = (cpu_to_node(cpu) & 1);;
+ socket = (cpu_to_node(cpu) & 1);
bdp->socket_mask |= (1 << socket);
sdp = &bdp->socket[socket];
sdp->cpu_number[sdp->num_cpus] = cpu;
sdp->num_cpus++;
}
- uvhub = 0;
- while (uvhub_mask) {
- if (!(uvhub_mask & 1))
- goto nexthub;
+ for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
+ if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
+ continue;
+ have_hmaster = 0;
bdp = &uvhub_descs[uvhub];
socket_mask = bdp->socket_mask;
socket = 0;
bcp->cpu = cpu;
if (i == 0) {
smaster = bcp;
- if (socket == 0)
+ if (!have_hmaster) {
+ have_hmaster++;
hmaster = bcp;
+ }
}
bcp->cpus_in_uvhub = bdp->num_cpus;
bcp->cpus_in_socket = sdp->num_cpus;
socket++;
socket_mask = (socket_mask >> 1);
}
-nexthub:
- uvhub++;
- uvhub_mask = (uvhub_mask >> 1);
}
kfree(uvhub_descs);
+ kfree(uvhub_mask);
for_each_present_cpu(cpu) {
bcp = &per_cpu(bau_control, cpu);
bcp->baudisabled = 0;
alloc_intr_gate(vector, uv_bau_message_intr1);
for_each_possible_blade(uvhub) {
- pnode = uv_blade_to_pnode(uvhub);
- /* INIT the bau */
- uv_write_global_mmr64(pnode, UVH_LB_BAU_SB_ACTIVATION_CONTROL,
- ((unsigned long)1 << 63));
- mmr = 1; /* should be 1 to broadcast to both sockets */
- uv_write_global_mmr64(pnode, UVH_BAU_DATA_BROADCAST, mmr);
+ if (uv_blade_nr_possible_cpus(uvhub)) {
+ pnode = uv_blade_to_pnode(uvhub);
+ /* INIT the bau */
+ uv_write_global_mmr64(pnode,
+ UVH_LB_BAU_SB_ACTIVATION_CONTROL,
+ ((unsigned long)1 << 63));
+ mmr = 1; /* should be 1 to broadcast to both sockets */
+ uv_write_global_mmr64(pnode, UVH_BAU_DATA_BROADCAST,
+ mmr);
+ }
}
return 0;