acpi/hmat: Register processor domain to its memory
authorKeith Busch <keith.busch@intel.com>
Mon, 11 Mar 2019 20:56:03 +0000 (14:56 -0600)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 4 Apr 2019 16:41:21 +0000 (18:41 +0200)
If the HMAT Subsystem Address Range provides a valid processor proximity
domain for a memory domain, or a processor domain matches the performance
access of the valid processor proximity domain, register the memory
target with that initiator so this relationship will be visible under
the node's sysfs directory.

Since HMAT requires valid address ranges have an equivalent SRAT entry,
verify each memory target satisfies this requirement.

Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Brice Goglin <Brice.Goglin@inria.fr>
Tested-by: Brice Goglin <Brice.Goglin@inria.fr>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/acpi/hmat/Kconfig
drivers/acpi/hmat/hmat.c

index 2f7111b7af6231c7685fadacc8544d510724dca0..13cddd612a529686f7d2c2b8471ee76101681d1f 100644 (file)
@@ -4,4 +4,5 @@ config ACPI_HMAT
        depends on ACPI_NUMA
        help
         If set, this option has the kernel parse and report the
-        platform's ACPI HMAT (Heterogeneous Memory Attributes Table).
+        platform's ACPI HMAT (Heterogeneous Memory Attributes Table),
+        and register memory initiators with their targets.
index 4758beb3b2c104602ef9b98cda63c6c27f1dde05..01a6eddac6f7ab23404f749a1fbce6480bb4b25d 100644 (file)
 #include <linux/device.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/list_sort.h>
 #include <linux/node.h>
 #include <linux/sysfs.h>
 
 static __initdata u8 hmat_revision;
 
+static __initdata LIST_HEAD(targets);
+static __initdata LIST_HEAD(initiators);
+static __initdata LIST_HEAD(localities);
+
+/*
+ * The defined enum order is used to prioritize attributes to break ties when
+ * selecting the best performing node.
+ */
+enum locality_types {
+       WRITE_LATENCY,
+       READ_LATENCY,
+       WRITE_BANDWIDTH,
+       READ_BANDWIDTH,
+};
+
+static struct memory_locality *localities_types[4];
+
+struct memory_target {
+       struct list_head node;
+       unsigned int memory_pxm;
+       unsigned int processor_pxm;
+       struct node_hmem_attrs hmem_attrs;
+};
+
+struct memory_initiator {
+       struct list_head node;
+       unsigned int processor_pxm;
+};
+
+struct memory_locality {
+       struct list_head node;
+       struct acpi_hmat_locality *hmat_loc;
+};
+
+static __init struct memory_initiator *find_mem_initiator(unsigned int cpu_pxm)
+{
+       struct memory_initiator *initiator;
+
+       list_for_each_entry(initiator, &initiators, node)
+               if (initiator->processor_pxm == cpu_pxm)
+                       return initiator;
+       return NULL;
+}
+
+static __init struct memory_target *find_mem_target(unsigned int mem_pxm)
+{
+       struct memory_target *target;
+
+       list_for_each_entry(target, &targets, node)
+               if (target->memory_pxm == mem_pxm)
+                       return target;
+       return NULL;
+}
+
+static __init void alloc_memory_initiator(unsigned int cpu_pxm)
+{
+       struct memory_initiator *initiator;
+
+       if (pxm_to_node(cpu_pxm) == NUMA_NO_NODE)
+               return;
+
+       initiator = find_mem_initiator(cpu_pxm);
+       if (initiator)
+               return;
+
+       initiator = kzalloc(sizeof(*initiator), GFP_KERNEL);
+       if (!initiator)
+               return;
+
+       initiator->processor_pxm = cpu_pxm;
+       list_add_tail(&initiator->node, &initiators);
+}
+
+static __init void alloc_memory_target(unsigned int mem_pxm)
+{
+       struct memory_target *target;
+
+       if (pxm_to_node(mem_pxm) == NUMA_NO_NODE)
+               return;
+
+       target = find_mem_target(mem_pxm);
+       if (target)
+               return;
+
+       target = kzalloc(sizeof(*target), GFP_KERNEL);
+       if (!target)
+               return;
+
+       target->memory_pxm = mem_pxm;
+       target->processor_pxm = PXM_INVAL;
+       list_add_tail(&target->node, &targets);
+}
+
 static __init const char *hmat_data_type(u8 type)
 {
        switch (type) {
@@ -89,14 +183,83 @@ static __init u32 hmat_normalize(u16 entry, u64 base, u8 type)
        return value;
 }
 
+static __init void hmat_update_target_access(struct memory_target *target,
+                                            u8 type, u32 value)
+{
+       switch (type) {
+       case ACPI_HMAT_ACCESS_LATENCY:
+               target->hmem_attrs.read_latency = value;
+               target->hmem_attrs.write_latency = value;
+               break;
+       case ACPI_HMAT_READ_LATENCY:
+               target->hmem_attrs.read_latency = value;
+               break;
+       case ACPI_HMAT_WRITE_LATENCY:
+               target->hmem_attrs.write_latency = value;
+               break;
+       case ACPI_HMAT_ACCESS_BANDWIDTH:
+               target->hmem_attrs.read_bandwidth = value;
+               target->hmem_attrs.write_bandwidth = value;
+               break;
+       case ACPI_HMAT_READ_BANDWIDTH:
+               target->hmem_attrs.read_bandwidth = value;
+               break;
+       case ACPI_HMAT_WRITE_BANDWIDTH:
+               target->hmem_attrs.write_bandwidth = value;
+               break;
+       default:
+               break;
+       }
+}
+
+static __init void hmat_add_locality(struct acpi_hmat_locality *hmat_loc)
+{
+       struct memory_locality *loc;
+
+       loc = kzalloc(sizeof(*loc), GFP_KERNEL);
+       if (!loc) {
+               pr_notice_once("Failed to allocate HMAT locality\n");
+               return;
+       }
+
+       loc->hmat_loc = hmat_loc;
+       list_add_tail(&loc->node, &localities);
+
+       switch (hmat_loc->data_type) {
+       case ACPI_HMAT_ACCESS_LATENCY:
+               localities_types[READ_LATENCY] = loc;
+               localities_types[WRITE_LATENCY] = loc;
+               break;
+       case ACPI_HMAT_READ_LATENCY:
+               localities_types[READ_LATENCY] = loc;
+               break;
+       case ACPI_HMAT_WRITE_LATENCY:
+               localities_types[WRITE_LATENCY] = loc;
+               break;
+       case ACPI_HMAT_ACCESS_BANDWIDTH:
+               localities_types[READ_BANDWIDTH] = loc;
+               localities_types[WRITE_BANDWIDTH] = loc;
+               break;
+       case ACPI_HMAT_READ_BANDWIDTH:
+               localities_types[READ_BANDWIDTH] = loc;
+               break;
+       case ACPI_HMAT_WRITE_BANDWIDTH:
+               localities_types[WRITE_BANDWIDTH] = loc;
+               break;
+       default:
+               break;
+       }
+}
+
 static __init int hmat_parse_locality(union acpi_subtable_headers *header,
                                      const unsigned long end)
 {
        struct acpi_hmat_locality *hmat_loc = (void *)header;
+       struct memory_target *target;
        unsigned int init, targ, total_size, ipds, tpds;
        u32 *inits, *targs, value;
        u16 *entries;
-       u8 type;
+       u8 type, mem_hier;
 
        if (hmat_loc->header.length < sizeof(*hmat_loc)) {
                pr_notice("HMAT: Unexpected locality header length: %d\n",
@@ -105,6 +268,7 @@ static __init int hmat_parse_locality(union acpi_subtable_headers *header,
        }
 
        type = hmat_loc->data_type;
+       mem_hier = hmat_loc->flags & ACPI_HMAT_MEMORY_HIERARCHY;
        ipds = hmat_loc->number_of_initiator_Pds;
        tpds = hmat_loc->number_of_target_Pds;
        total_size = sizeof(*hmat_loc) + sizeof(*entries) * ipds * tpds +
@@ -123,6 +287,7 @@ static __init int hmat_parse_locality(union acpi_subtable_headers *header,
        targs = inits + ipds;
        entries = (u16 *)(targs + tpds);
        for (init = 0; init < ipds; init++) {
+               alloc_memory_initiator(inits[init]);
                for (targ = 0; targ < tpds; targ++) {
                        value = hmat_normalize(entries[init * tpds + targ],
                                               hmat_loc->entry_base_unit,
@@ -130,9 +295,18 @@ static __init int hmat_parse_locality(union acpi_subtable_headers *header,
                        pr_info("  Initiator-Target[%d-%d]:%d%s\n",
                                inits[init], targs[targ], value,
                                hmat_data_type_suffix(type));
+
+                       if (mem_hier == ACPI_HMAT_MEMORY) {
+                               target = find_mem_target(targs[targ]);
+                               if (target && target->processor_pxm == inits[init])
+                                       hmat_update_target_access(target, type, value);
+                       }
                }
        }
 
+       if (mem_hier == ACPI_HMAT_MEMORY)
+               hmat_add_locality(hmat_loc);
+
        return 0;
 }
 
@@ -160,6 +334,7 @@ static int __init hmat_parse_proximity_domain(union acpi_subtable_headers *heade
                                              const unsigned long end)
 {
        struct acpi_hmat_proximity_domain *p = (void *)header;
+       struct memory_target *target;
 
        if (p->header.length != sizeof(*p)) {
                pr_notice("HMAT: Unexpected address range header length: %d\n",
@@ -175,6 +350,23 @@ static int __init hmat_parse_proximity_domain(union acpi_subtable_headers *heade
                pr_info("HMAT: Memory Flags:%04x Processor Domain:%d Memory Domain:%d\n",
                        p->flags, p->processor_PD, p->memory_PD);
 
+       if (p->flags & ACPI_HMAT_MEMORY_PD_VALID) {
+               target = find_mem_target(p->memory_PD);
+               if (!target) {
+                       pr_debug("HMAT: Memory Domain missing from SRAT\n");
+                       return -EINVAL;
+               }
+       }
+       if (target && p->flags & ACPI_HMAT_PROCESSOR_PD_VALID) {
+               int p_node = pxm_to_node(p->processor_PD);
+
+               if (p_node == NUMA_NO_NODE) {
+                       pr_debug("HMAT: Invalid Processor Domain\n");
+                       return -EINVAL;
+               }
+               target->processor_pxm = p_node;
+       }
+
        return 0;
 }
 
@@ -198,6 +390,191 @@ static int __init hmat_parse_subtable(union acpi_subtable_headers *header,
        }
 }
 
+static __init int srat_parse_mem_affinity(union acpi_subtable_headers *header,
+                                         const unsigned long end)
+{
+       struct acpi_srat_mem_affinity *ma = (void *)header;
+
+       if (!ma)
+               return -EINVAL;
+       if (!(ma->flags & ACPI_SRAT_MEM_ENABLED))
+               return 0;
+       alloc_memory_target(ma->proximity_domain);
+       return 0;
+}
+
+static __init u32 hmat_initiator_perf(struct memory_target *target,
+                              struct memory_initiator *initiator,
+                              struct acpi_hmat_locality *hmat_loc)
+{
+       unsigned int ipds, tpds, i, idx = 0, tdx = 0;
+       u32 *inits, *targs;
+       u16 *entries;
+
+       ipds = hmat_loc->number_of_initiator_Pds;
+       tpds = hmat_loc->number_of_target_Pds;
+       inits = (u32 *)(hmat_loc + 1);
+       targs = inits + ipds;
+       entries = (u16 *)(targs + tpds);
+
+       for (i = 0; i < ipds; i++) {
+               if (inits[i] == initiator->processor_pxm) {
+                       idx = i;
+                       break;
+               }
+       }
+
+       if (i == ipds)
+               return 0;
+
+       for (i = 0; i < tpds; i++) {
+               if (targs[i] == target->memory_pxm) {
+                       tdx = i;
+                       break;
+               }
+       }
+       if (i == tpds)
+               return 0;
+
+       return hmat_normalize(entries[idx * tpds + tdx],
+                             hmat_loc->entry_base_unit,
+                             hmat_loc->data_type);
+}
+
+static __init bool hmat_update_best(u8 type, u32 value, u32 *best)
+{
+       bool updated = false;
+
+       if (!value)
+               return false;
+
+       switch (type) {
+       case ACPI_HMAT_ACCESS_LATENCY:
+       case ACPI_HMAT_READ_LATENCY:
+       case ACPI_HMAT_WRITE_LATENCY:
+               if (!*best || *best > value) {
+                       *best = value;
+                       updated = true;
+               }
+               break;
+       case ACPI_HMAT_ACCESS_BANDWIDTH:
+       case ACPI_HMAT_READ_BANDWIDTH:
+       case ACPI_HMAT_WRITE_BANDWIDTH:
+               if (!*best || *best < value) {
+                       *best = value;
+                       updated = true;
+               }
+               break;
+       }
+
+       return updated;
+}
+
+static int initiator_cmp(void *priv, struct list_head *a, struct list_head *b)
+{
+       struct memory_initiator *ia;
+       struct memory_initiator *ib;
+       unsigned long *p_nodes = priv;
+
+       ia = list_entry(a, struct memory_initiator, node);
+       ib = list_entry(b, struct memory_initiator, node);
+
+       set_bit(ia->processor_pxm, p_nodes);
+       set_bit(ib->processor_pxm, p_nodes);
+
+       return ia->processor_pxm - ib->processor_pxm;
+}
+
+static __init void hmat_register_target_initiators(struct memory_target *target)
+{
+       static DECLARE_BITMAP(p_nodes, MAX_NUMNODES);
+       struct memory_initiator *initiator;
+       unsigned int mem_nid, cpu_nid;
+       struct memory_locality *loc = NULL;
+       u32 best = 0;
+       int i;
+
+       mem_nid = pxm_to_node(target->memory_pxm);
+       /*
+        * If the Address Range Structure provides a local processor pxm, link
+        * only that one. Otherwise, find the best performance attributes and
+        * register all initiators that match.
+        */
+       if (target->processor_pxm != PXM_INVAL) {
+               cpu_nid = pxm_to_node(target->processor_pxm);
+               register_memory_node_under_compute_node(mem_nid, cpu_nid, 0);
+               return;
+       }
+
+       if (list_empty(&localities))
+               return;
+
+       /*
+        * We need the initiator list sorted so we can use bitmap_clear for
+        * previously set initiators when we find a better memory accessor.
+        * We'll also use the sorting to prime the candidate nodes with known
+        * initiators.
+        */
+       bitmap_zero(p_nodes, MAX_NUMNODES);
+       list_sort(p_nodes, &initiators, initiator_cmp);
+       for (i = WRITE_LATENCY; i <= READ_BANDWIDTH; i++) {
+               loc = localities_types[i];
+               if (!loc)
+                       continue;
+
+               best = 0;
+               list_for_each_entry(initiator, &initiators, node) {
+                       u32 value;
+
+                       if (!test_bit(initiator->processor_pxm, p_nodes))
+                               continue;
+
+                       value = hmat_initiator_perf(target, initiator, loc->hmat_loc);
+                       if (hmat_update_best(loc->hmat_loc->data_type, value, &best))
+                               bitmap_clear(p_nodes, 0, initiator->processor_pxm);
+                       if (value != best)
+                               clear_bit(initiator->processor_pxm, p_nodes);
+               }
+               if (best)
+                       hmat_update_target_access(target, loc->hmat_loc->data_type, best);
+       }
+
+       for_each_set_bit(i, p_nodes, MAX_NUMNODES) {
+               cpu_nid = pxm_to_node(i);
+               register_memory_node_under_compute_node(mem_nid, cpu_nid, 0);
+       }
+}
+
+static __init void hmat_register_targets(void)
+{
+       struct memory_target *target;
+
+       list_for_each_entry(target, &targets, node)
+               hmat_register_target_initiators(target);
+}
+
+static __init void hmat_free_structures(void)
+{
+       struct memory_target *target, *tnext;
+       struct memory_locality *loc, *lnext;
+       struct memory_initiator *initiator, *inext;
+
+       list_for_each_entry_safe(target, tnext, &targets, node) {
+               list_del(&target->node);
+               kfree(target);
+       }
+
+       list_for_each_entry_safe(initiator, inext, &initiators, node) {
+               list_del(&initiator->node);
+               kfree(initiator);
+       }
+
+       list_for_each_entry_safe(loc, lnext, &localities, node) {
+               list_del(&loc->node);
+               kfree(loc);
+       }
+}
+
 static __init int hmat_init(void)
 {
        struct acpi_table_header *tbl;
@@ -207,6 +584,17 @@ static __init int hmat_init(void)
        if (srat_disabled())
                return 0;
 
+       status = acpi_get_table(ACPI_SIG_SRAT, 0, &tbl);
+       if (ACPI_FAILURE(status))
+               return 0;
+
+       if (acpi_table_parse_entries(ACPI_SIG_SRAT,
+                               sizeof(struct acpi_table_srat),
+                               ACPI_SRAT_TYPE_MEMORY_AFFINITY,
+                               srat_parse_mem_affinity, 0) < 0)
+               goto out_put;
+       acpi_put_table(tbl);
+
        status = acpi_get_table(ACPI_SIG_HMAT, 0, &tbl);
        if (ACPI_FAILURE(status))
                return 0;
@@ -229,7 +617,9 @@ static __init int hmat_init(void)
                        goto out_put;
                }
        }
+       hmat_register_targets();
 out_put:
+       hmat_free_structures();
        acpi_put_table(tbl);
        return 0;
 }