svn commit: r322171 - head/sys/amd64/amd64

[Konstantin Belousov]

Author: kib
Date: Mon Aug  7 17:23:10 2017
New Revision: 322171
URL: https://svnweb.freebsd.org/changeset/base/322171

Log:
  Explain why delayed invalidation is not required in pmap_protect() and
  pmap_remove_pages().
  
  Submitted by:	alc
  MFC after:	1 week

Modified:
  head/sys/amd64/amd64/pmap.c

Modified: head/sys/amd64/amd64/pmap.c
==============================================================================
--- head/sys/amd64/amd64/pmap.c	Mon Aug  7 16:23:53 2017	(r322170)
+++ head/sys/amd64/amd64/pmap.c	Mon Aug  7 17:23:10 2017	(r322171)
@@ -4085,6 +4085,26 @@ pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t
 	PG_RW = pmap_rw_bit(pmap);
 	anychanged = FALSE;
 
+	/*
+	 * Although this function delays and batches the invalidation
+	 * of stale TLB entries, it does not need to call
+	 * pmap_delayed_invl_started() and
+	 * pmap_delayed_invl_finished(), because it does not
+	 * ordinarily destroy mappings.  Stale TLB entries from
+	 * protection-only changes need only be invalidated before the
+	 * pmap lock is released, because protection-only changes do
+	 * not destroy PV entries.  Even operations that iterate over
+	 * a physical page's PV list of mappings, like
+	 * pmap_remove_write(), acquire the pmap lock for each
+	 * mapping.  Consequently, for protection-only changes, the
+	 * pmap lock suffices to synchronize both page table and TLB
+	 * updates.
+	 *
+	 * This function only destroys a mapping if pmap_demote_pde()
+	 * fails.  In that case, stale TLB entries are immediately
+	 * invalidated.
+	 */
+	
 	PMAP_LOCK(pmap);
 	for (; sva < eva; sva = va_next) {
 
@@ -5469,6 +5489,15 @@ pmap_page_is_mapped(vm_page_t m)
  * no processor is currently accessing the user address space.  In
  * particular, a page table entry's dirty bit won't change state once
  * this function starts.
+ *
+ * Although this function destroys all of the pmap's managed,
+ * non-wired mappings, it can delay and batch the invalidation of TLB
+ * entries without calling pmap_delayed_invl_started() and
+ * pmap_delayed_invl_finished().  Because the pmap is not active on
+ * any other processor, none of these TLB entries will ever be used
+ * before their eventual invalidation.  Consequently, there is no need
+ * for either pmap_remove_all() or pmap_remove_write() to wait for
+ * that eventual TLB invalidation.
  */
 void
 pmap_remove_pages(pmap_t pmap)