Menu
▾
▴
[Lse-tech] virtual address to node mappings
|
From: Per E. <pe...@pd...> - 2004-10-11 15:14:20
|
This is a patch that adds a per-process /proc entry that maps virtual
memory address ranges to node numbers. It may be useful to people
debugging NUMA performance anomalies. I thought I'd float it by this
list for comments.
# cat /proc/self/nodemem
0000000000504000-0000000000525fff node 0
0000002a95556000-0000002a95556fff node 0
...
0000002a9576a000-0000002a9576afff node 3
Process running on node 0
Process has 2392064 bytes (584 pages) on node 0
Process has 122880 bytes (30 pages) on node 1
Process has 122880 bytes (30 pages) on node 2
Process has 90112 bytes (22 pages) on node 3
Process has 0 bytes (0 pages) on node 4
Process has 0 bytes (0 pages) on node 5
Process has 0 bytes (0 pages) on node 6
Process has 0 bytes (0 pages) on node 7
Process has 0 bytes (0 pages) not in core
/Per
---
diff -uprN linux-2.6.8.1/fs/proc/Makefile linux-2.6.8.1-patched/fs/proc/Makefile
--- linux-2.6.8.1/fs/proc/Makefile 2004-08-14 12:54:48.000000000 +0200
+++ linux-2.6.8.1-patched/fs/proc/Makefile 2004-10-11 15:46:32.000000000 +0200
@@ -8,7 +8,7 @@ proc-y := task_nommu.o
proc-$(CONFIG_MMU) := task_mmu.o
proc-y += inode.o root.o base.o generic.o array.o \
- kmsg.o proc_tty.o proc_misc.o
+ kmsg.o proc_tty.o proc_misc.o nodemem.o
proc-$(CONFIG_PROC_KCORE) += kcore.o
proc-$(CONFIG_PROC_DEVICETREE) += proc_devtree.o
diff -uprN linux-2.6.8.1/fs/proc/base.c linux-2.6.8.1-patched/fs/proc/base.c
--- linux-2.6.8.1/fs/proc/base.c 2004-08-14 12:55:35.000000000 +0200
+++ linux-2.6.8.1-patched/fs/proc/base.c 2004-10-11 15:46:46.000000000 +0200
@@ -48,6 +48,9 @@ enum pid_directory_inos {
PROC_TGID_TASK,
PROC_TGID_STATUS,
PROC_TGID_MEM,
+#ifdef CONFIG_NUMA
+ PROC_TGID_NODEMEM,
+#endif
PROC_TGID_CWD,
PROC_TGID_ROOT,
PROC_TGID_EXE,
@@ -71,6 +74,9 @@ enum pid_directory_inos {
PROC_TID_INO,
PROC_TID_STATUS,
PROC_TID_MEM,
+#ifdef CONFIG_NUMA
+ PROC_TID_NODEMEM,
+#endif
PROC_TID_CWD,
PROC_TID_ROOT,
PROC_TID_EXE,
@@ -113,6 +119,9 @@ static struct pid_entry tgid_base_stuff[
E(PROC_TGID_STATM, "statm", S_IFREG|S_IRUGO),
E(PROC_TGID_MAPS, "maps", S_IFREG|S_IRUGO),
E(PROC_TGID_MEM, "mem", S_IFREG|S_IRUSR|S_IWUSR),
+#ifdef CONFIG_NUMA
+ E(PROC_TGID_NODEMEM, "nodemem", S_IFREG|S_IRUGO),
+#endif
E(PROC_TGID_CWD, "cwd", S_IFLNK|S_IRWXUGO),
E(PROC_TGID_ROOT, "root", S_IFLNK|S_IRWXUGO),
E(PROC_TGID_EXE, "exe", S_IFLNK|S_IRWXUGO),
@@ -135,6 +144,9 @@ static struct pid_entry tid_base_stuff[]
E(PROC_TID_STATM, "statm", S_IFREG|S_IRUGO),
E(PROC_TID_MAPS, "maps", S_IFREG|S_IRUGO),
E(PROC_TID_MEM, "mem", S_IFREG|S_IRUSR|S_IWUSR),
+#ifdef CONFIG_NUMA
+ E(PROC_TID_NODEMEM, "nodemem", S_IFREG|S_IRUGO),
+#endif
E(PROC_TID_CWD, "cwd", S_IFLNK|S_IRWXUGO),
E(PROC_TID_ROOT, "root", S_IFLNK|S_IRWXUGO),
E(PROC_TID_EXE, "exe", S_IFLNK|S_IRWXUGO),
@@ -689,6 +701,31 @@ static struct inode_operations proc_mem_
.permission = proc_permission,
};
+
+#ifdef CONFIG_NUMA
+
+
+extern int nodemem_show(struct seq_file *s, void *v);
+
+static int nodemem_open(struct inode *inode, struct file *file) {
+ struct task_struct *task = proc_task(inode);
+ int ret = single_open(file, nodemem_show, NULL);
+ if (!ret) {
+ struct seq_file *m = file->private_data;
+ m->private = task;
+ }
+ return ret;
+}
+
+static struct file_operations proc_nodemem_operations = {
+ .open = nodemem_open,
+ .read = seq_read,
+ .llseek = seq_lseek,
+ .release = single_release,
+};
+
+#endif /* CONFIG_NUMA */
+
static int proc_pid_follow_link(struct dentry *dentry, struct nameidata *nd)
{
struct inode *inode = dentry->d_inode;
@@ -1342,6 +1379,12 @@ static struct dentry *proc_pident_lookup
inode->i_op = &proc_mem_inode_operations;
inode->i_fop = &proc_mem_operations;
break;
+#ifdef CONFIG_NUMA
+ case PROC_TID_NODEMEM:
+ case PROC_TGID_NODEMEM:
+ inode->i_fop = &proc_nodemem_operations;
+ break;
+#endif
case PROC_TID_MOUNTS:
case PROC_TGID_MOUNTS:
inode->i_fop = &proc_mounts_operations;
diff -uprN linux-2.6.8.1/fs/proc/nodemem.c linux-2.6.8.1-patched/fs/proc/nodemem.c
--- linux-2.6.8.1/fs/proc/nodemem.c 1970-01-01 01:00:00.000000000 +0100
+++ linux-2.6.8.1-patched/fs/proc/nodemem.c 2004-10-11 16:22:48.863258816 +0200
@@ -0,0 +1,221 @@
+/*
+ * linux/fs/proc/nodemem.c
+ *
+ * Copyright (C) 2004 Per Ekman (pe...@pd...)
+ * Development supported by AMD.
+ *
+ * This file is released under the GPL.
+ *
+ * Per-process proc device to map virtual memory to nodes.
+ *
+ */
+
+#include <linux/errno.h>
+#include <linux/proc_fs.h>
+#include <linux/kernel.h>
+#include <linux/mmzone.h>
+#include <linux/mm.h>
+#include <linux/seq_file.h>
+#include <linux/sched.h>
+#include <linux/threads.h>
+
+/* Offsets into the range list array. If the range_list array is
+ * extended, make sure NOT_PRESENT remains the size of it. */
+#define NOT_PRESENT NR_CPUS
+
+#define RP_ENTRY_SIZE (2*(sizeof(unsigned long)))
+#define RP_MAX_ENTRIES ((PAGE_SIZE / RP_ENTRY_SIZE) - 1)
+
+#define RP_START(p, i) (p->v[i*2])
+#define RP_END(p, i) (p->v[i*2 + 1])
+#define RP_ISEMPTY(p, i) ((p->v[i*2] == 0) && (p->v[i*2 + 1] == 0))
+
+struct range_page {
+ unsigned long count;
+ unsigned long v[2 * RP_MAX_ENTRIES] __attribute__ ((packed));
+ struct range_page *next __attribute__ ((packed));
+};
+
+struct range_list {
+ struct range_page *rp;
+ unsigned long *cache;
+};
+
+static inline int mergeable_range(unsigned long a, unsigned long b,
+ unsigned long x, unsigned long y)
+{
+ if (a < x) {
+ if (b >= x - 1)
+ return 1;
+ } else if (a <= y + 1)
+ return 1;
+
+ return 0;
+}
+
+static inline void join_pair(unsigned long *v, unsigned long start,
+ unsigned long end)
+{
+ if (start < v[0])
+ v[0] = start;
+ if (end > v[1])
+ v[1] = end;
+}
+
+static inline struct range_page *grow_rp(struct range_page *rp)
+{
+ struct range_page *p;
+
+ if (rp == NULL)
+ p = rp;
+ else
+ p = rp->next;
+ if ((p = (struct range_page *)__get_free_page(GFP_KERNEL)) == NULL)
+ return NULL;
+ memset(p, 0, PAGE_SIZE);
+
+ return p;
+}
+
+static void free_rl(struct range_list *rlp)
+{
+ struct range_page *p, *rp = rlp->rp;
+
+ while (rp) {
+ p = rp;
+ rp = rp->next;
+ free_page((unsigned long)p);
+ }
+}
+
+static void insert_rl(struct range_list *rlp, unsigned long start,
+ unsigned long end)
+{
+ struct range_page *rp, *p = NULL;
+ unsigned long *empty = NULL;
+ int i;
+
+ if (rlp->cache &&
+ mergeable_range(start, end, rlp->cache[0], rlp->cache[1])) {
+ join_pair(rlp->cache, start, end);
+ return;
+ }
+ rp = rlp->rp;
+ while (rp) {
+ p = rp;
+ for (i = 0; i < p->count; i++) {
+ if (RP_ISEMPTY(p, i)) {
+ if (empty == NULL)
+ empty = &RP_START(p, i);
+ } else if (mergeable_range(start, end, RP_START(p, i),
+ RP_END(p, i))) {
+ join_pair(&RP_START(p, i), start, end);
+ rlp->cache = &RP_START(p, i);
+ return;
+ }
+ }
+ rp = rp->next;
+ }
+
+ /* No hit */
+ if (empty) {
+ empty[0] = start;
+ empty[1] = end;
+ return;
+ }
+ if (!p || p->count == RP_MAX_ENTRIES)
+ p = grow_rp(p);
+ if (!rlp->rp)
+ rlp->rp = p;
+ RP_START(p, p->count) = start;
+ RP_END(p, p->count) = end;
+ rlp->cache = &RP_START(p, p->count);
+ p->count++;
+}
+
+static inline int page2node(struct page *page)
+{
+ struct zone *zone = page_zone(page);
+ return zone->zone_pgdat->node_id;
+}
+
+int nodemem_show(struct seq_file *s, void *v)
+{
+ struct range_list *rlp;
+ struct range_page *rp;
+ struct task_struct *task = s->private;
+ struct vm_area_struct *vma;
+ pgd_t *pgd;
+ pmd_t *pmd;
+ pte_t *ptep, pte;
+ unsigned long pga;
+ int n, i;
+
+ long node_pgc[NOT_PRESENT + 1];
+
+ rlp = kmalloc((sizeof *rlp) * (NOT_PRESENT + 1), GFP_KERNEL);
+ if (!rlp)
+ return 1;
+ memset(rlp, 0, (sizeof *rlp) * (NOT_PRESENT + 1));
+
+ for (n = 0; n <= NOT_PRESENT; n++)
+ node_pgc[n] = 0;
+ down_read(&task->active_mm->mmap_sem);
+ for (vma = task->active_mm->mmap; vma; vma = vma->vm_next) {
+ for (pga = PAGE_ALIGN(vma->vm_start); pga < vma->vm_end;
+ pga += PAGE_SIZE) {
+ n = NOT_PRESENT;
+ pgd = pgd_offset(task->active_mm, pga);
+ if (pgd_none(*pgd) || pgd_bad(*pgd))
+ goto out;
+ pmd = pmd_offset(pgd, pga);
+ if (pmd_none(*pmd) || pmd_bad(*pmd))
+ goto out;
+ ptep = pte_offset_map(pmd, pga);
+ if (!ptep)
+ goto out;
+ pte = *ptep;
+ pte_unmap(ptep);
+ n = page2node(pte_page(pte));
+ out:
+ node_pgc[n]++;
+ insert_rl(&rlp[n], pga, pga + PAGE_SIZE - 1);
+ }
+ }
+ up_read(&task->active_mm->mmap_sem);
+ for (n = 0; n <= NOT_PRESENT; n++) {
+ for (rp = rlp[n].rp; rp; rp = rp->next) {
+ for (i = 0; i < rp->count; i++) {
+ if (RP_ISEMPTY(rp, i))
+ continue;
+ if (n == NOT_PRESENT)
+ seq_printf(s,
+ "%016lx-%016lx not "
+ "present\n",
+ RP_START(rp, i),
+ RP_END(rp, i));
+ else
+ seq_printf(s, "%016lx-%016lx node %d\n",
+ RP_START(rp, i),
+ RP_END(rp, i), n);
+
+ }
+ }
+ }
+ seq_printf(s, "Process running on node %d\n", task->thread_info->cpu);
+ for (n = 0; n <= NOT_PRESENT; n++) {
+ if (n == NOT_PRESENT)
+ seq_printf(s, "Process has %ld bytes (%ld pages) "
+ "not in core\n",
+ node_pgc[n] * PAGE_SIZE, node_pgc[n]);
+ else
+ seq_printf(s, "Process has %ld bytes (%ld pages)"
+ " on node %d\n",
+ node_pgc[n] * PAGE_SIZE, node_pgc[n], n);
+ }
+
+ for (n = 0; n <= NOT_PRESENT; n++)
+ free_rl(&rlp[n]);
+
+ return 0;
+}
|