From: Ben Pfaff <blp@cs.stanford.edu>
Date: Thu, 7 Apr 2005 18:00:06 +0000 (+0000)
Subject: In idle(), "sti; hlt" must be atomic or we can waste time waiting for
X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?p=pintos-anon;a=commitdiff_plain;h=7fd56df9056d6442d52f23d3fb356010b988668d

In idle(), "sti; hlt" must be atomic or we can waste time waiting for
the next clock tick.
Fix inspired by Linus Torvalds in message
<Pine.LNX.4.58.0504071000450.28951@ppc970.osdl.org>.
---

diff --git a/src/threads/thread.c b/src/threads/thread.c
index 6eda280..02f3c7b 100644
--- a/src/threads/thread.c
+++ b/src/threads/thread.c
@@ -299,11 +299,19 @@ idle (void *aux UNUSED)
       /* Let someone else run. */
       intr_disable ();
       thread_block ();
-      intr_enable ();
 
-      /* Use CPU `hlt' instruction to wait for interrupt.
-         See [IA32-v2a] "HLT" and [IA32-v3] 7.7. */
-      asm ("hlt");
+      /* Re-enable interrupts and wait for the next one.
+
+         The `sti' instruction disables interrupts until the
+         completion of the next instruction, so these two
+         instructions are executed atomically.  This atomicity is
+         important; otherwise, an interrupt could be handled
+         between re-enabling interrupts and waiting for the next
+         one to occur, wasting as much as one clock tick worth of
+         time.
+
+         See [IA32-v2a] "HLT", [IA32-v2b] "STI", and [IA32-v3] 7.7. */
+      asm ("sti; hlt");
     }
 }