writeback: fix BDI_WRITTEN accounting disturbing bdi->completions

Peter,

This is an amazing bug. I'm not sure how the accounting goes wrong in some tricky way. But you can compare the exact bdi proportions pattern before/after patch. The gray "bdi setpoint" lines are vastly different.

---
When increasing BDI_WRITTEN together with bdi->completions inside the same local irq disabling block, bdi_thresh is found to go wild in the 1 disk + 1 usb stick writeback test case. Fix it by moving BDI_WRITTEN accounting out.

Signed-off-by Wu Fengguang <[email protected]>

---
 mm/page-writeback.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

mm/page-writeback.c +2 -1

@@ -230,7 +230,6 @@ int dirty_bytes_handler(struct ctl_table
  */
 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
 {
-	__inc_bdi_stat(bdi, BDI_WRITTEN);
 	__prop_inc_percpu_max(&vm_completions, &bdi->completions,
 			      bdi->max_prop_frac);
 }

@@ -239,6 +238,7 @@ void bdi_writeout_inc(struct backing_dev
 {
 	unsigned long flags;
 
+	inc_bdi_stat(bdi, BDI_WRITTEN);
 	local_irq_save(flags);
 	__bdi_writeout_inc(bdi);
 	local_irq_restore(flags);

@@ -1577,6 +1577,7 @@ int test_clear_page_writeback(struct pag
 						PAGECACHE_TAG_WRITEBACK);
 			if (bdi_cap_account_writeback(bdi)) {
 				__dec_bdi_stat(bdi, BDI_WRITEBACK);
+				__inc_bdi_stat(bdi, BDI_WRITTEN);
 				__bdi_writeout_inc(bdi);
 			}
 		}

Quoted 11 lines · Wu Fengguang

Peter,

This is an amazing bug. I'm not sure how the accounting goes wrong in some tricky way. But you can compare the exact bdi proportions pattern before/after patch. The gray "bdi setpoint" lines are vastly different.

---
When increasing BDI_WRITTEN together with bdi->completions inside the same local irq disabling block, bdi_thresh is found to go wild in the 1 disk + 1 usb stick writeback test case. Fix it by moving BDI_WRITTEN accounting out.

I don't understand this - the patch is just NOP. The change in
test_clear_page_writeback() does absolutely nothing and the change in bdi_writeout_inc() just changes:
local_irq_save(flags);
__inc_bdi_stat(bdi); __prop_inc_percpu_max(&vm_completions, &bdi->completions,
bdi->max_prop_frac);
local_irq_restore(flags);

to:
local_irq_save(flags);
__inc_bdi_stat(bdi);
local_irq_restore(flags); local_irq_save(flags); __prop_inc_percpu_max(&vm_completions, &bdi->completions,
bdi->max_prop_frac);
local_irq_restore(flags);

So the difference must be in something else...