Skip to contentSkip to conversation

percpu: flush tlb after pcpu_depopulate_chunk()

PATCH Guenter Roeck 2 participants 3 messages Jul 3, 2021 Memory management

lore
Share & tools
lore

Guenter Roeck

author Tested-by
#0 lore raw
Quoted 9 lines · Dennis Zhou

Prior to "percpu: implement partial chunk depopulation", pcpu_depopulate_chunk() was called only on the destruction path. This meant the virtual address range was on its way back to vmalloc which will handle flushing the tlbs for us.

However, now that we call pcpu_depopulate_chunk() during the active lifecycle of a chunk, we need to flush the tlb as well otherwise we can end up accessing the wrong page through an invalid tlb mapping.

Quoted 7 lines

This was reported in [1].

[1] https://lore.kernel.org/lkml/[email protected]/

Fixes f183324133ea ("percpu: implement partial chunk depopulation")

Thanks!
Guenter

Quoted 91 lines
---
 mm/percpu-km.c |  3 ++-
 mm/percpu-vm.c | 11 +++++++++--
 mm/percpu.c    |  7 ++++---
 3 files changed, 15 insertions(+), 6 deletions(-)
mm/percpu-km.c +2 -1
index c9d529dc7651..6875fc3b2ed7 100644
@@ -39,7 +39,8 @@ static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
 }
 
 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
-				  int page_start, int page_end)
+				  int page_start, int page_end,
+				  bool flush_tlb)
 {
 	/* nada */
 }
mm/percpu-vm.c +9 -2
index ee5d89fcd66f..6353cda1718e 100644
@@ -299,6 +299,7 @@ static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
  * @chunk: chunk to depopulate
  * @page_start: the start page
  * @page_end: the end page
+ * @flush_tlb: if should we flush the tlb
  *
  * For each cpu, depopulate and unmap pages [@page_start,@page_end)
  * from @chunk.
@@ -307,7 +308,8 @@ static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
  * pcpu_alloc_mutex.
  */
 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
-				  int page_start, int page_end)
+				  int page_start, int page_end,
+				  bool flush_tlb)
 {
 	struct page **pages;
 
@@ -324,7 +326,12 @@ static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
 
 	pcpu_unmap_pages(chunk, pages, page_start, page_end);
 
-	/* no need to flush tlb, vmalloc will handle it lazily */
+	/*
+	 * We need to flush the tlb unless the caller will pass it to vmalloc,
+	 * which will handle flushing for us.
+	 */
+	if (flush_tlb)
+		pcpu_post_unmap_tlb_flush(chunk, page_start, page_end);
 
 	pcpu_free_pages(chunk, pages, page_start, page_end);
 }
mm/percpu.c +4 -3
index b4cebeca4c0c..e23ba0d22220 100644
@@ -1580,7 +1580,8 @@ static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
 static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
 			       int page_start, int page_end, gfp_t gfp);
 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
-				  int page_start, int page_end);
+				  int page_start, int page_end,
+				  bool flush_tlb);
 static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
 static struct page *pcpu_addr_to_page(void *addr);
@@ -2016,7 +2017,7 @@ static void pcpu_balance_free(bool empty_only)
 
 		bitmap_for_each_set_region(chunk->populated, rs, re, 0,
 					   chunk->nr_pages) {
-			pcpu_depopulate_chunk(chunk, rs, re);
+			pcpu_depopulate_chunk(chunk, rs, re, false);
 			spin_lock_irq(&pcpu_lock);
 			pcpu_chunk_depopulated(chunk, rs, re);
 			spin_unlock_irq(&pcpu_lock);
@@ -2189,7 +2190,7 @@ static void pcpu_reclaim_populated(void)
 				continue;
 
 			spin_unlock_irq(&pcpu_lock);
-			pcpu_depopulate_chunk(chunk, i + 1, end + 1);
+			pcpu_depopulate_chunk(chunk, i + 1, end + 1, true);
 			cond_resched();
 			spin_lock_irq(&pcpu_lock);
 
Signature · 1 line
2.32.0.93.g670b81a890-goog
#1 lore raw
Quoted 14 lines · Guenter Roeck
Quoted 13 lines · Dennis Zhou

Prior to "percpu: implement partial chunk depopulation", pcpu_depopulate_chunk() was called only on the destruction path. This meant the virtual address range was on its way back to vmalloc which will handle flushing the tlbs for us.

However, now that we call pcpu_depopulate_chunk() during the active lifecycle of a chunk, we need to flush the tlb as well otherwise we can end up accessing the wrong page through an invalid tlb mapping.

This was reported in [1].

[1] https://lore.kernel.org/lkml/[email protected]/

Quoted 9 lines
Quoted 3 lines
Fixes f183324133ea ("percpu: implement partial chunk depopulation")

Thanks!
Guenter

I updated the tag and have applied this to for-5.14-fixes so it can get some for-next exposure before I send a fix to Linus.

Thanks for setting up the super easy repro! It made debugging this significantly easier.

Thanks,
Dennis

#2 lore raw
Quoted 25 lines · Dennis Zhou
Quoted 14 lines · Guenter Roeck
Quoted 13 lines · Dennis Zhou

Prior to "percpu: implement partial chunk depopulation", pcpu_depopulate_chunk() was called only on the destruction path. This meant the virtual address range was on its way back to vmalloc which will handle flushing the tlbs for us.

However, now that we call pcpu_depopulate_chunk() during the active lifecycle of a chunk, we need to flush the tlb as well otherwise we can end up accessing the wrong page through an invalid tlb mapping.

This was reported in [1].

[1] https://lore.kernel.org/lkml/[email protected]/

Quoted 9 lines
Quoted 3 lines
Fixes f183324133ea ("percpu: implement partial chunk depopulation")

Thanks!
Guenter

Quoted 6 lines

I updated the tag and have applied this to for-5.14-fixes so it can get some for-next exposure before I send a fix to Linus.

Thanks for setting up the super easy repro! It made debugging this significantly easier.

I'm pulling this because I think I'll need to do some aggregation to make this less expensive. I'll send some variant soon hopefully.

Thanks,
Dennis