accounting per process swapped out pages

Hello everyone.

Here is patch to count per process swapped out pages. This patch is created against 2.6.16.1. So far, I had tested by forcing certain task to swap out (tail -f /dev/zero) and wait until top/vmstat/free reported that swap is occupied.

Comments and feedbacks are greatly appreciated. Please keep me CC'ed since I am not subscribed to linux-kernel mailing list.

regards,

Mulyadi

# diffstat ./swap-acct.patch
 fs/proc/task_mmu.c |   14 ++++++++++++--
 include/linux/mm.h |    1 +
 mm/memory.c        |    3 +++
 mm/rmap.c          |    1 +
 mm/swapfile.c      |    1 +
5 files changed, 18 insertions(+), 2 deletions(-)

@@ -15,6 +15,8 @@ char *task_mem(struct mm_struct *mm, cha
 {
 	unsigned long data, text, lib;
 	unsigned long hiwater_vm, total_vm, hiwater_rss, total_rss;
+	struct vm_area_struct * this_vma;
+	unsigned long swp_out_pages=0;
 
 	/*
 	 * Note: to minimize their overhead, mm maintains hiwater_vm and

@@ -33,6 +35,12 @@ char *task_mem(struct mm_struct *mm, cha
 	data = mm->total_vm - mm->shared_vm - mm->stack_vm;
 	text = (PAGE_ALIGN(mm->end_code) - (mm->start_code & PAGE_MASK)) >> 10;
 	lib = (mm->exec_vm << (PAGE_SHIFT-10)) - text;
+
+	for(this_vma = mm->mmap;this_vma;this_vma = this_vma->vm_next)
+	{
+		swp_out_pages += atomic_read(&(this_vma)->swapped_out);
+	}
+
 	buffer += sprintf(buffer,
 		"VmPeak:\t%8lu kB\n"
 		"VmSize:\t%8lu kB\n"

@@ -43,7 +51,8 @@ char *task_mem(struct mm_struct *mm, cha
 		"VmStk:\t%8lu kB\n"
 		"VmExe:\t%8lu kB\n"
 		"VmLib:\t%8lu kB\n"
-		"VmPTE:\t%8lu kB\n",
+		"VmPTE:\t%8lu kB\n"
+		"VmSwp:\t%8lu kB\n",
 		hiwater_vm << (PAGE_SHIFT-10),
 		(total_vm - mm->reserved_vm) << (PAGE_SHIFT-10),
 		mm->locked_vm << (PAGE_SHIFT-10),

@@ -51,7 +60,8 @@ char *task_mem(struct mm_struct *mm, cha
 		total_rss << (PAGE_SHIFT-10),
 		data << (PAGE_SHIFT-10),
 		mm->stack_vm << (PAGE_SHIFT-10), text, lib,
-		(PTRS_PER_PTE*sizeof(pte_t)*mm->nr_ptes) >> 10);
+		(PTRS_PER_PTE*sizeof(pte_t)*mm->nr_ptes) >> 10,
+		swp_out_pages << (PAGE_SHIFT-10) );
 	return buffer;
 }
 

@@ -111,6 +111,7 @@ struct vm_area_struct {
 #ifdef CONFIG_NUMA
 	struct mempolicy *vm_policy;	/* NUMA policy for the VMA */
 #endif
+	atomic_t  swapped_out;	/* swapped out pages in this VMA */
 };
 
 /*

@@ -635,6 +635,7 @@ static int try_to_unmap_one(struct page 
 		set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
 		BUG_ON(pte_file(*pte));
 		dec_mm_counter(mm, anon_rss);
+		atomic_inc(&(vma)->swapped_out);
 	} else
 		dec_mm_counter(mm, file_rss);
 

@@ -683,6 +683,7 @@ static unsigned long zap_pte_range(struc
 	} while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
 
 	add_mm_rss(mm, file_rss, anon_rss);
+	atomic_sub(anon_rss,&(vma)->swapped_out);
 	pte_unmap_unlock(pte - 1, ptl);
 
 	return addr;

@@ -1928,6 +1929,8 @@ again:
 	/* The page isn't present yet, go ahead with the fault. */
 
 	inc_mm_counter(mm, anon_rss);
+	printk("Reducing swapped_out variable...\n");
+	atomic_dec(&(vma)->swapped_out);
 	pte = mk_pte(page, vma->vm_page_prot);
 	if (write_access && can_share_swap_page(page)) {
 		pte = maybe_mkwrite(pte_mkdirty(pte), vma);

@@ -426,6 +426,7 @@ static void unuse_pte(struct vm_area_str
 		unsigned long addr, swp_entry_t entry, struct page *page)
 {
 	inc_mm_counter(vma->vm_mm, anon_rss);
+	atomic_dec(&(vma)->swapped_out);
 	get_page(page);
 	set_pte_at(vma->vm_mm, addr, pte,
 		   pte_mkold(mk_pte(page, vma->vm_page_prot)));

To be honest, I don't think there's much interest in this particular VmSwp statistic; and if there's little interest in it, we'd rather not spend the time and space on collecting it. But I could be wrong:
let's see who speaks up for it.

A few comments on the mechanics of your patch.

You waste space in every vm_area_struct for your swapped_out count, then /proc/<pid>/status has to loop over the vmas adding them up. Much better to make it an mm_counter like anon_rss, then you only use space in mm_struct, and don't have to add them up at the end, and avoid dirtying (vma) cachelines unnecessarily, and (in some cases) avoid the atomic operations.

While you've caught the main places where you'd need to adjust swapped_out, you've missed a couple (maybe I've missed more):
copy_pte_range (fork) needs to increment the count, zap_pte_range (munmap or truncate or exit) needs to decrement it. Check wherever anon_rss is adjusted, some not all would need swapped_out adjusted.

Oh, you are doing something in zap_pte_range, but I'm sorry to say what you do there is nonsense: the number you're subtracting has nothing to do with the number of swapped out pages.

And you probably wouldn't want that printk in your final patch!

Hugh

Hi ...

I got the same impression too, but I do think this statistic is important if someone wants to know which program suffers most when (heavy) swapping happens. Some programs like "top" would also get benefit since its "swapped" accounting is just based on virtual-rss, which is wrong (just showing parts that haven't been brought to RAM)

Let's see if I can adapt this idea. Actually, in the first place, I was thinking this way but sometimes it is easier to get pointer to vm_area_struct rather than to mm_struct. I'll take a deeper look...

Avoiding atomic operation is also my another primary target....

Yes, I use anon_rss a the "guidance" on where I probably put the swapped_out accounting and decide whether it actually has anything to do with swapping or not. Hmmmm, looks like I still miss a lot...

:( OK, re-checking...

Perfect closing, i forgot to clean that printk(). Thanks for reminding me...

Refined patch will follow ASAP..

regards,

Mulyadi