This patch uses find_vma() to improve the read response of /proc/pid/maps.
It attempts to make the liner scan instead of quadratic walk and utilise
rb tree. Reading the file was doing sequential scan from the begining to
file position all the time, and taking a quite long time.
The improvements came from f_version/m_version and resulting in mmap_cache match.
Even if mmap_cache does not match, rb tree walk should be faster than sequential
walk. First attempt was to put the state across read system calls into private
data. Later got inspiration from wli's pid patch using f_version in readdir of /proc.
Other advantage is, f_version will be cleared automatically by lseek.
The test program creates 32K maps and splits them into two(limited by max_map_count
sysctl) using mprotect(0). After the patch, the read time improves from many
seconds to milliseconds, and does not grow superlinearly with number of read calls.
Help taken from Peter Swain in idea and testing.
After the patch:
Reading /proc/self/maps:65528 time: 0 secs and 780728 usecs buf:4096 bytes:3811362
Reading /proc/self/maps:65528 time: 1 secs and 117573 usecs buf:1024 bytes:3866627
Reading /proc/self/maps:65528 time: 0 secs and 473459 usecs buf: 256 bytes:3866627
Reading /proc/self/maps:65528 time: 0 secs and 901288 usecs buf: 64 bytes:3866627
Reading /proc/self/maps:65528 time: 1 secs and 480185 usecs buf: 16 bytes:3866627
Reading /proc/self/maps:65528 time: 1 secs and 636268 usecs buf: 4 bytes:3866627
Reading /proc/self/maps:65528 time: 4 secs and 118327 usecs buf: 1 bytes:3866627
Before the patch:
Reading /proc/self/maps:65528 time: 4 secs and 359556 usecs buf:4096 bytes:3866647
Reading /proc/self/maps:65528 time:16 secs and 218584 usecs buf:1024 bytes:3866688
Reading /proc/self/maps:65528 time:67 secs and 870200 usecs buf: 256 bytes:3866688
Reading /proc/self/maps:65528 time:255 secs and 186934 usecs buf: 64 bytes:3866688
Small reads never completed.
Signed-off-by: Prasanna Meda <[email protected]>
--- a/include/linux/seq_file.h Tue Dec 28 03:54:29 2004
+++ b/include/linux/seq_file.h Tue Dec 28 03:55:10 2004
@@ -18,6 +18,7 @@
size_t from;
size_t count;
loff_t index;
+ loff_t version;
struct semaphore sem;
struct seq_operations *op;
void *private;
--- a/fs/seq_file.c Tue Dec 28 03:50:12 2004
+++ b/fs/seq_file.c Tue Dec 28 21:54:12 2004
@@ -35,6 +35,7 @@
sema_init(&p->sem, 1);
p->op = op;
file->private_data = p;
+ file->f_version = 0;
/* SEQ files support lseek, but not pread/pwrite */
file->f_mode &= ~(FMODE_PREAD | FMODE_PWRITE);
@@ -58,6 +59,7 @@
int err = 0;
down(&m->sem);
+ m->version = file->f_version;
/* grab buffer if we didn't have one */
if (!m->buf) {
m->buf = kmalloc(m->size = PAGE_SIZE, GFP_KERNEL);
@@ -136,6 +138,7 @@
copied = err;
else
*ppos += copied;
+ file->f_version = m->version;
up(&m->sem);
return copied;
Enomem:
--- a/fs/proc/task_mmu.c Tue Dec 28 02:51:19 2004
+++ b/fs/proc/task_mmu.c Tue Dec 28 21:34:00 2004
@@ -81,24 +81,52 @@
static void *m_start(struct seq_file *m, loff_t *pos)
{
struct task_struct *task = m->private;
- struct mm_struct *mm = get_task_mm(task);
- struct vm_area_struct * map;
+ unsigned long last_addr = m->version;
+ struct mm_struct *mm;
+ struct vm_area_struct *map, *tail_map;
loff_t l = *pos;
- if (!mm)
+ if (last_addr == -1UL)
return NULL;
+ mm = get_task_mm(task);
+ if (!mm) {
+ m->version = -1UL;
+ return NULL;
+ }
+
+ tail_map = get_gate_vma(task);
down_read(&mm->mmap_sem);
- map = mm->mmap;
- while (l-- && map)
+
+ /* Start with last addr hint */
+ map = find_vma(mm, last_addr);
+ if (map) {
map = map->vm_next;
- if (!map) {
- up_read(&mm->mmap_sem);
- mmput(mm);
- if (l == -1)
- map = get_gate_vma(task);
+ goto out;
}
- return map;
+
+ /* Check the map index is within the range */
+ if ((unsigned long)l < mm->map_count) {
+ map = mm->mmap;
+ while (l-- && map)
+ map = map->vm_next;
+ goto out;
+ }
+
+ if (l != mm->map_count)
+ tail_map = NULL; /* After gate map */
+
+out:
+ if (map && map != tail_map) {
+ m->version = map->vm_start;
+ return map;
+ }
+
+ /* End of maps has reached */
+ m->version = -1UL;
+ up_read(&mm->mmap_sem);
+ mmput(mm);
+ return tail_map;
}
static void m_stop(struct seq_file *m, void *v)
@@ -117,9 +145,12 @@
struct task_struct *task = m->private;
struct vm_area_struct *map = v;
(*pos)++;
- if (map->vm_next)
+ if (map && map->vm_next) {
+ m->version = map->vm_next->vm_start;
return map->vm_next;
+ }
m_stop(m, v);
+ m->version = -1UL;
if (map != get_gate_vma(task))
return get_gate_vma(task);
return NULL;
On Wed, Jan 05, 2005 at 12:55:37PM -0800, [email protected] wrote:
> This patch uses find_vma() to improve the read response of /proc/pid/maps.
> It attempts to make the liner scan instead of quadratic walk and utilise
> rb tree. Reading the file was doing sequential scan from the begining to
> file position all the time, and taking a quite long time.
> The improvements came from f_version/m_version and resulting in
> mmap_cache match. Even if mmap_cache does not match, rb tree walk
> should be faster than sequential walk. First attempt was to put the
> state across read system calls into private data. Later got
> inspiration from wli's pid patch using f_version in readdir of /proc.
> Other advantage is, f_version will be cleared automatically by lseek.
> The test program creates 32K maps and splits them into two(limited by
> max_map_count sysctl) using mprotect(0). After the patch, the read
> time improves from many seconds to milliseconds, and does not grow
> superlinearly with number of read calls.
> Help taken from Peter Swain in idea and testing.
This is an excellent improvement which makes kernel instrumentation
interfere less with the instrumented workload. I wholly endorse it.
-- wli