From c29a7baf091fc6b2c9e40561030f8c62e6145a19 Mon Sep 17 00:00:00 2001 From: Michael Holzheu Date: Thu, 6 Mar 2014 18:47:21 +0100 Subject: [PATCH] s390/numa: add emulation support NUMA emulation (aka fake NUMA) distributes the available memory to nodes without using real topology information about the physical memory of the machine. Splitting the system memory into nodes replicates the memory management structures for each node. Particularly each node has its own "mm locks" and its own "kswapd" task. For large systems, under certain conditions, this results in improved system performance and/or latency based on reduced pressure on the mm locks and the kswapd tasks. NUMA emulation distributes CPUs to nodes while respecting the original machine topology information. This is done by trying to avoid to separate CPUs which reside on the same book or even on the same MC. Because the current Linux scheduler code requires a stable cpu to node mapping, cores are pinned to nodes when the first CPU thread is set online. This patch is based on the initial implementation from Philipp Hachtmann. Signed-off-by: Michael Holzheu Signed-off-by: Martin Schwidefsky --- arch/s390/Kconfig | 37 +++ arch/s390/include/asm/numa.h | 4 + arch/s390/numa/Makefile | 1 + arch/s390/numa/mode_emu.c | 511 +++++++++++++++++++++++++++++++++++ arch/s390/numa/numa.c | 4 + arch/s390/numa/numa_mode.h | 1 + drivers/s390/char/sclp_cmd.c | 18 +- 7 files changed, 569 insertions(+), 7 deletions(-) create mode 100644 arch/s390/numa/mode_emu.c diff --git a/arch/s390/Kconfig b/arch/s390/Kconfig index 25510adb07d3..cb418dcc2d45 100644 --- a/arch/s390/Kconfig +++ b/arch/s390/Kconfig @@ -423,6 +423,43 @@ config NODES_SHIFT Specify the maximum number of NUMA nodes available on the target system. Increases memory reserved to accommodate various tables. +menu "Select NUMA modes" + depends on NUMA + +config NUMA_EMU + bool "NUMA emulation" + default y + help + Numa emulation mode will split the available system memory into + equal chunks which then are distributed over the configured number + of nodes in a round-robin manner. + + The number of fake nodes is limited by the number of available memory + chunks (i.e. memory size / fake size) and the number of supported + nodes in the kernel. + + The CPUs are assigned to the nodes in a way that partially respects + the original machine topology (if supported by the machine). + Fair distribution of the CPUs is not guaranteed. + +config EMU_SIZE + hex "NUMA emulation memory chunk size" + default 0x10000000 + range 0x400000 0x100000000 + depends on NUMA_EMU + help + Select the default size by which the memory is chopped and then + assigned to emulated NUMA nodes. + + This can be overridden by specifying + + emu_size= + + on the kernel command line where also suffixes K, M, G, and T are + supported. + +endmenu + config SCHED_MC def_bool n diff --git a/arch/s390/include/asm/numa.h b/arch/s390/include/asm/numa.h index ea4edbfba9f6..2a0efc63b9e5 100644 --- a/arch/s390/include/asm/numa.h +++ b/arch/s390/include/asm/numa.h @@ -26,6 +26,10 @@ extern int numa_debug_enabled; static inline void numa_setup(void) { } static inline void numa_update_cpu_topology(void) { } +static inline int numa_pfn_to_nid(unsigned long pfn) +{ + return 0; +} #endif /* CONFIG_NUMA */ #endif /* _ASM_S390_NUMA_H */ diff --git a/arch/s390/numa/Makefile b/arch/s390/numa/Makefile index 31372293b62e..f94ecaffa71b 100644 --- a/arch/s390/numa/Makefile +++ b/arch/s390/numa/Makefile @@ -1,2 +1,3 @@ obj-y += numa.o obj-y += toptree.o +obj-$(CONFIG_NUMA_EMU) += mode_emu.o diff --git a/arch/s390/numa/mode_emu.c b/arch/s390/numa/mode_emu.c new file mode 100644 index 000000000000..9d4e1e15a6f0 --- /dev/null +++ b/arch/s390/numa/mode_emu.c @@ -0,0 +1,511 @@ +/* + * NUMA support for s390 + * + * NUMA emulation (aka fake NUMA) distributes the available memory to nodes + * without using real topology information about the physical memory of the + * machine. + * + * It distributes the available CPUs to nodes while respecting the original + * machine topology information. This is done by trying to avoid to separate + * CPUs which reside on the same book or even on the same MC. + * + * Because the current Linux scheduler code requires a stable cpu to node + * mapping, cores are pinned to nodes when the first CPU thread is set online. + * + * Copyright IBM Corp. 2015 + */ + +#define KMSG_COMPONENT "numa_emu" +#define pr_fmt(fmt) KMSG_COMPONENT ": " fmt + +#include +#include +#include +#include +#include +#include +#include +#include "numa_mode.h" +#include "toptree.h" + +/* Distances between the different system components */ +#define DIST_EMPTY 0 +#define DIST_CORE 1 +#define DIST_MC 2 +#define DIST_BOOK 3 +#define DIST_MAX 4 + +/* Node distance reported to common code */ +#define EMU_NODE_DIST 10 + +/* Node ID for free (not yet pinned) cores */ +#define NODE_ID_FREE -1 + +/* Different levels of toptree */ +enum toptree_level {CORE, MC, BOOK, NODE, TOPOLOGY}; + +/* The two toptree IDs */ +enum {TOPTREE_ID_PHYS, TOPTREE_ID_NUMA}; + +/* Number of NUMA nodes */ +static int emu_nodes = 1; +/* NUMA stripe size */ +static unsigned long emu_size; +/* Pinned core to node mapping */ +static int cores_to_node_id[CONFIG_NR_CPUS]; +/* Total number of pinned cores */ +static int cores_total; +/* Number of cores per node without extra cores */ +static int cores_per_node_target; +/* Number of cores pinned to node */ +static int cores_per_node[MAX_NUMNODES]; + +/* + * Pin a core to a node + */ +static void pin_core_to_node(int core_id, int node_id) +{ + if (cores_to_node_id[core_id] == NODE_ID_FREE) { + cores_per_node[node_id]++; + cores_to_node_id[core_id] = node_id; + cores_total++; + } else { + WARN_ON(cores_to_node_id[core_id] != node_id); + } +} + +/* + * Number of pinned cores of a node + */ +static int cores_pinned(struct toptree *node) +{ + return cores_per_node[node->id]; +} + +/* + * ID of the node where the core is pinned (or NODE_ID_FREE) + */ +static int core_pinned_to_node_id(struct toptree *core) +{ + return cores_to_node_id[core->id]; +} + +/* + * Number of cores in the tree that are not yet pinned + */ +static int cores_free(struct toptree *tree) +{ + struct toptree *core; + int count = 0; + + toptree_for_each(core, tree, CORE) { + if (core_pinned_to_node_id(core) == NODE_ID_FREE) + count++; + } + return count; +} + +/* + * Return node of core + */ +static struct toptree *core_node(struct toptree *core) +{ + return core->parent->parent->parent; +} + +/* + * Return book of core + */ +static struct toptree *core_book(struct toptree *core) +{ + return core->parent->parent; +} + +/* + * Return mc of core + */ +static struct toptree *core_mc(struct toptree *core) +{ + return core->parent; +} + +/* + * Distance between two cores + */ +static int dist_core_to_core(struct toptree *core1, struct toptree *core2) +{ + if (core_book(core1)->id != core_book(core2)->id) + return DIST_BOOK; + if (core_mc(core1)->id != core_mc(core2)->id) + return DIST_MC; + /* Same core or sibling on same MC */ + return DIST_CORE; +} + +/* + * Distance of a node to a core + */ +static int dist_node_to_core(struct toptree *node, struct toptree *core) +{ + struct toptree *core_node; + int dist_min = DIST_MAX; + + toptree_for_each(core_node, node, CORE) + dist_min = min(dist_min, dist_core_to_core(core_node, core)); + return dist_min == DIST_MAX ? DIST_EMPTY : dist_min; +} + +/* + * Unify will delete empty nodes, therefore recreate nodes. + */ +static void toptree_unify_tree(struct toptree *tree) +{ + int nid; + + toptree_unify(tree); + for (nid = 0; nid < emu_nodes; nid++) + toptree_get_child(tree, nid); +} + +/* + * Find the best/nearest node for a given core and ensure that no node + * gets more than "cores_per_node_target + extra" cores. + */ +static struct toptree *node_for_core(struct toptree *numa, struct toptree *core, + int extra) +{ + struct toptree *node, *node_best = NULL; + int dist_cur, dist_best; + + dist_best = DIST_MAX; + node_best = NULL; + toptree_for_each(node, numa, NODE) { + /* Already pinned cores must use their nodes */ + if (core_pinned_to_node_id(core) == node->id) { + node_best = node; + break; + } + /* Skip nodes that already have enough cores */ + if (cores_pinned(node) >= cores_per_node_target + extra) + continue; + dist_cur = dist_node_to_core(node, core); + if (dist_cur < dist_best) { + dist_best = dist_cur; + node_best = node; + } + } + return node_best; +} + +/* + * Find the best node for each core with respect to "extra" core count + */ +static void toptree_to_numa_single(struct toptree *numa, struct toptree *phys, + int extra) +{ + struct toptree *node, *core, *tmp; + + toptree_for_each_safe(core, tmp, phys, CORE) { + node = node_for_core(numa, core, extra); + if (!node) + return; + toptree_move(core, node); + pin_core_to_node(core->id, node->id); + } +} + +/* + * Move structures of given level to specified NUMA node + */ +static void move_level_to_numa_node(struct toptree *node, struct toptree *phys, + enum toptree_level level, bool perfect) +{ + struct toptree *cur, *tmp; + int cores_free; + + toptree_for_each_safe(cur, tmp, phys, level) { + cores_free = cores_per_node_target - toptree_count(node, CORE); + if (perfect) { + if (cores_free == toptree_count(cur, CORE)) + toptree_move(cur, node); + } else { + if (cores_free >= toptree_count(cur, CORE)) + toptree_move(cur, node); + } + } +} + +/* + * Move structures of a given level to NUMA nodes. If "perfect" is specified + * move only perfectly fitting structures. Otherwise move also smaller + * than needed structures. + */ +static void move_level_to_numa(struct toptree *numa, struct toptree *phys, + enum toptree_level level, bool perfect) +{ + struct toptree *node; + + toptree_for_each(node, numa, NODE) + move_level_to_numa_node(node, phys, level, perfect); +} + +/* + * For the first run try to move the big structures + */ +static void toptree_to_numa_first(struct toptree *numa, struct toptree *phys) +{ + struct toptree *core; + + /* Always try to move perfectly fitting structures first */ + move_level_to_numa(numa, phys, BOOK, true); + move_level_to_numa(numa, phys, BOOK, false); + move_level_to_numa(numa, phys, MC, true); + move_level_to_numa(numa, phys, MC, false); + /* Now pin all the moved cores */ + toptree_for_each(core, numa, CORE) + pin_core_to_node(core->id, core_node(core)->id); +} + +/* + * Allocate new topology and create required nodes + */ +static struct toptree *toptree_new(int id, int nodes) +{ + struct toptree *tree; + int nid; + + tree = toptree_alloc(TOPOLOGY, id); + if (!tree) + goto fail; + for (nid = 0; nid < nodes; nid++) { + if (!toptree_get_child(tree, nid)) + goto fail; + } + return tree; +fail: + panic("NUMA emulation could not allocate topology"); +} + +/* + * Move cores from physical topology into NUMA target topology + * and try to keep as much of the physical topology as possible. + */ +static struct toptree *toptree_to_numa(struct toptree *phys) +{ + static int first = 1; + struct toptree *numa; + + cores_per_node_target = (cores_total + cores_free(phys)) / emu_nodes; + numa = toptree_new(TOPTREE_ID_NUMA, emu_nodes); + if (first) { + toptree_to_numa_first(numa, phys); + first = 0; + } + toptree_to_numa_single(numa, phys, 0); + toptree_to_numa_single(numa, phys, 1); + toptree_unify_tree(numa); + + WARN_ON(cpumask_weight(&phys->mask)); + return numa; +} + +/* + * Create a toptree out of the physical topology that we got from the hypervisor + */ +static struct toptree *toptree_from_topology(void) +{ + struct toptree *phys, *node, *book, *mc, *core; + struct cpu_topology_s390 *top; + int cpu; + + phys = toptree_new(TOPTREE_ID_PHYS, 1); + + for_each_online_cpu(cpu) { + top = &per_cpu(cpu_topology, cpu); + node = toptree_get_child(phys, 0); + book = toptree_get_child(node, top->book_id); + mc = toptree_get_child(book, top->socket_id); + core = toptree_get_child(mc, top->core_id); + if (!book || !mc || !core) + panic("NUMA emulation could not allocate memory"); + cpumask_set_cpu(cpu, &core->mask); + toptree_update_mask(mc); + } + return phys; +} + +/* + * Add toptree core to topology and create correct CPU masks + */ +static void topology_add_core(struct toptree *core) +{ + struct cpu_topology_s390 *top; + int cpu; + + for_each_cpu(cpu, &core->mask) { + top = &per_cpu(cpu_topology, cpu); + cpumask_copy(&top->thread_mask, &core->mask); + cpumask_copy(&top->core_mask, &core_mc(core)->mask); + cpumask_copy(&top->book_mask, &core_book(core)->mask); + cpumask_set_cpu(cpu, node_to_cpumask_map[core_node(core)->id]); + top->node_id = core_node(core)->id; + } +} + +/* + * Apply toptree to topology and create CPU masks + */ +static void toptree_to_topology(struct toptree *numa) +{ + struct toptree *core; + int i; + + /* Clear all node masks */ + for (i = 0; i < MAX_NUMNODES; i++) + cpumask_clear(node_to_cpumask_map[i]); + + /* Rebuild all masks */ + toptree_for_each(core, numa, CORE) + topology_add_core(core); +} + +/* + * Show the node to core mapping + */ +static void print_node_to_core_map(void) +{ + int nid, cid; + + if (!numa_debug_enabled) + return; + printk(KERN_DEBUG "NUMA node to core mapping\n"); + for (nid = 0; nid < emu_nodes; nid++) { + printk(KERN_DEBUG " node %3d: ", nid); + for (cid = 0; cid < ARRAY_SIZE(cores_to_node_id); cid++) { + if (cores_to_node_id[cid] == nid) + printk(KERN_CONT "%d ", cid); + } + printk(KERN_CONT "\n"); + } +} + +/* + * Transfer physical topology into a NUMA topology and modify CPU masks + * according to the NUMA topology. + * + * This function is called under the CPU hotplug lock. + */ +static void emu_update_cpu_topology(void) +{ + struct toptree *phys, *numa; + + phys = toptree_from_topology(); + numa = toptree_to_numa(phys); + toptree_free(phys); + toptree_to_topology(numa); + toptree_free(numa); + print_node_to_core_map(); +} + +/* + * If emu_size is not set, use CONFIG_EMU_SIZE. Then round to minimum + * alignment (needed for memory hotplug). + */ +static unsigned long emu_setup_size_adjust(unsigned long size) +{ + size = size ? : CONFIG_EMU_SIZE; + size = roundup(size, memory_block_size_bytes()); + return size; +} + +/* + * If we have not enough memory for the specified nodes, reduce the node count. + */ +static int emu_setup_nodes_adjust(int nodes) +{ + int nodes_max; + + nodes_max = memblock.memory.total_size / emu_size; + nodes_max = max(nodes_max, 1); + if (nodes_max >= nodes) + return nodes; + pr_warn("Not enough memory for %d nodes, reducing node count\n", nodes); + return nodes_max; +} + +/* + * Early emu setup + */ +static void emu_setup(void) +{ + int i; + + emu_size = emu_setup_size_adjust(emu_size); + emu_nodes = emu_setup_nodes_adjust(emu_nodes); + for (i = 0; i < ARRAY_SIZE(cores_to_node_id); i++) + cores_to_node_id[i] = NODE_ID_FREE; + pr_info("Creating %d nodes with memory stripe size %ld MB\n", + emu_nodes, emu_size >> 20); +} + +/* + * Return node id for given page number + */ +static int emu_pfn_to_nid(unsigned long pfn) +{ + return (pfn / (emu_size >> PAGE_SHIFT)) % emu_nodes; +} + +/* + * Return stripe size + */ +static unsigned long emu_align(void) +{ + return emu_size; +} + +/* + * Return distance between two nodes + */ +static int emu_distance(int node1, int node2) +{ + return (node1 != node2) * EMU_NODE_DIST; +} + +/* + * Define callbacks for generic s390 NUMA infrastructure + */ +const struct numa_mode numa_mode_emu = { + .name = "emu", + .setup = emu_setup, + .update_cpu_topology = emu_update_cpu_topology, + .__pfn_to_nid = emu_pfn_to_nid, + .align = emu_align, + .distance = emu_distance, +}; + +/* + * Kernel parameter: emu_nodes= + */ +static int __init early_parse_emu_nodes(char *p) +{ + int count; + + if (kstrtoint(p, 0, &count) != 0 || count <= 0) + return 0; + if (count <= 0) + return 0; + emu_nodes = min(count, MAX_NUMNODES); + return 0; +} +early_param("emu_nodes", early_parse_emu_nodes); + +/* + * Kernel parameter: emu_size=[[k|M|G|T]] + */ +static int __init early_parse_emu_size(char *p) +{ + emu_size = memparse(p, NULL); + return 0; +} +early_param("emu_size", early_parse_emu_size); diff --git a/arch/s390/numa/numa.c b/arch/s390/numa/numa.c index 0416a3671e33..09b1d2355bd9 100644 --- a/arch/s390/numa/numa.c +++ b/arch/s390/numa/numa.c @@ -175,6 +175,10 @@ static int __init parse_numa(char *parm) { if (strcmp(parm, numa_mode_plain.name) == 0) mode = &numa_mode_plain; +#ifdef CONFIG_NUMA_EMU + if (strcmp(parm, numa_mode_emu.name) == 0) + mode = &numa_mode_emu; +#endif return 0; } early_param("numa", parse_numa); diff --git a/arch/s390/numa/numa_mode.h b/arch/s390/numa/numa_mode.h index 775659848011..08953b0b1c7f 100644 --- a/arch/s390/numa/numa_mode.h +++ b/arch/s390/numa/numa_mode.h @@ -19,5 +19,6 @@ struct numa_mode { }; extern const struct numa_mode numa_mode_plain; +extern const struct numa_mode numa_mode_emu; #endif /* __S390_NUMA_MODE_H */ diff --git a/drivers/s390/char/sclp_cmd.c b/drivers/s390/char/sclp_cmd.c index e9485fbbb373..806239c2cf2f 100644 --- a/drivers/s390/char/sclp_cmd.c +++ b/drivers/s390/char/sclp_cmd.c @@ -25,6 +25,7 @@ #include #include #include +#include #include "sclp.h" @@ -388,11 +389,11 @@ static struct notifier_block sclp_mem_nb = { }; static void __init align_to_block_size(unsigned long long *start, - unsigned long long *size) + unsigned long long *size, + unsigned long long alignment) { - unsigned long long start_align, size_align, alignment; + unsigned long long start_align, size_align; - alignment = memory_block_size_bytes(); start_align = roundup(*start, alignment); size_align = rounddown(*start + *size, alignment) - start_align; @@ -404,8 +405,8 @@ static void __init align_to_block_size(unsigned long long *start, static void __init add_memory_merged(u16 rn) { + unsigned long long start, size, addr, block_size; static u16 first_rn, num; - unsigned long long start, size; if (rn && first_rn && (first_rn + num == rn)) { num++; @@ -423,9 +424,12 @@ static void __init add_memory_merged(u16 rn) goto skip_add; if (memory_end_set && (start + size > memory_end)) size = memory_end - start; - align_to_block_size(&start, &size); - if (size) - add_memory(0, start, size); + block_size = memory_block_size_bytes(); + align_to_block_size(&start, &size, block_size); + if (!size) + goto skip_add; + for (addr = start; addr < start + size; addr += block_size) + add_memory(numa_pfn_to_nid(PFN_DOWN(addr)), addr, block_size); skip_add: first_rn = rn; num = 1; -- 2.20.1