Re: [PATCH] f2fs: do decrease_sleep_time() if any of the victims have been selected

From: Yuwei Guan
Date: Sun Dec 11 2022 - 07:30:26 EST




在 2022/12/11 10:52, Chao Yu 写道:
On 2022/12/9 19:28, Yuwei Guan wrote:
In non-foreground gc mode, if no victim is selected, the gc process
will wait for no_gc_sleep_time before waking up again. In this
subsequent time, even though a victim will be selected, the gc process
still waits for no_gc_sleep_time before waking up. The configuration
of wait_ms is not reasonable.

After any of the victims have been selected, we need to do
decrease_sleep_time() to reduce wait_ms.

If it is GC_URGENT_HIGH or GC_URGENT_MID gc mode,
wait_ms will keep urgent_sleep_time after executing decrease_sleep_time().

In decrease_sleep_time() wait_time will be reduced to max_sleep_time
from no_gc_sleep_time, if *wait is no_gc_sleep_time. And then it goes
down in the next step.

Signed-off-by: Yuwei Guan <Yuwei.Guan@xxxxxxxxxxxxx>
---
  fs/f2fs/gc.c | 2 ++
  fs/f2fs/gc.h | 7 ++++++-
  2 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
index f0c6506d8975..c023ffeb9268 100644
--- a/fs/f2fs/gc.c
+++ b/fs/f2fs/gc.c
@@ -141,6 +141,8 @@ static int gc_thread_func(void *data)
              /* don't bother wait_ms by foreground gc */
              if (!foreground)
                  wait_ms = gc_th->no_gc_sleep_time;
+        } else {
+            decrease_sleep_time(gc_th, &wait_ms);

Once BGGC selects valid victim, it will go faster and fater?
> How about:

    } else {
        /* reset wait_ms to default sleep time */
        if (wait_ms == gc_th->no_gc_sleep_time)
            wait_ms = gc_th->min_sleep_time;
    }
Indeed. it will go faster and fater, until wait_ms reduces to min_sleep_time. But your suggestion seems more reasonable, I will send it in v2 patch.

Thanks,

          }
          if (foreground)
diff --git a/fs/f2fs/gc.h b/fs/f2fs/gc.h
index 19b956c2d697..6402584dcd72 100644
--- a/fs/f2fs/gc.h
+++ b/fs/f2fs/gc.h
@@ -150,8 +150,13 @@ static inline void decrease_sleep_time(struct f2fs_gc_kthread *gc_th,
  {
      unsigned int min_time = gc_th->min_sleep_time;
-    if (*wait == gc_th->no_gc_sleep_time)
+    if (*wait == gc_th->urgent_sleep_time)
+        return;
+
+    if (*wait == gc_th->no_gc_sleep_time) {
          *wait = gc_th->max_sleep_time;
+        return;
+    }
      if ((long long)*wait - (long long)min_time < (long long)min_time)
          *wait = min_time;