Re: [LTP] [linux-next:master] [mm/migrate] b28dd7507f: ltp.move_pages04.fail

[David Hildenbrand <david@redhat.com>]

On 28.08.24 12:37, Cyril Hrubis wrote:
> Hi!
>> Okay, and it even looks like the test caught the unintended change for
>> "unpopulated memory", but instead we decided to change the test to
>> expect the other return code ... because there was some confusion about
>> "zero page".
>>
>> Long story short: the test needs to be fixed.
> 
> Will do, but we need the patch to land into some kernel version first so
> taht we can add the range of kernels where the kernel wrongly returns
> EFAULT.
> 
> Or alternatively if you are going to backport this to stable trees we
> can revert the test change that expect -EFAULT so the test expects only
> -ENOENT.
> 

I am not yet sure if we should simply allow either -EFAULT or -ENOENT for
the "nothing mapped" case in the check (below).

Alternatively, I agree, wee need to have this in the kernel so we can
check for versions.

What would be your preference?


Currently I have (WIP):

 From c152535cdfae194819b6df450cdc29a60f8cdb8d Mon Sep 17 00:00:00 2001
From: David Hildenbrand <david@redhat.com>
Date: Wed, 21 Aug 2024 10:58:34 +0200
Subject: [PATCH] move_pages04: properly check for "no page mapped" and "shared
  zero page mapped"

While the kernel commit d899844e9c98 ("mm: fix status code which
move_pages() returns for zero page") fixed the return value when the
shared zero page was encountered to match what was state in the man page,
it unfortunately also changed the behavior when no page is mapped yet --
when no page was faulted in/populated on demand.

Then, this test started failing, and we thought we would be testing for
the "zero page" case, but actually we were testing for the "no page mapped"
yet case, and didn't realize that the kernel commit had unintended side
effects.

As we are changing the behavior back to return "-ENOENT" for the "no
page mapped" case, while still making keeping the shared zero page to
return "-EFAULT" the test starts failing again ...

The man page clearly spells out that the expectation for the zero page is
"-EFAULT", and that "-EFAULT" can also be returned if "the memory area is
not mapped by the process" -- which means that there is no VMA/mmap()
covering that address.

The man page isn't completely clear what the expected return value for the
"no page mapped" case is. It documents "-ENOENT" for "The page is not
present", which can be interpreted to include "there is nothing mapped"
and not just "there is something, it' just simply not suitable".

We'll clarify the man page soon, to be clearer that the expectation is
to get "-ENOENT" in that case, like the kernel originally did. But we'll
also add a note that some kernel versions will return either -ENOENT
or -EFAULT.

So let's test for both cases, and make sure we allow both -ENOENT and
-EFAULT for the "no page mapped" case.

Signed-off-by: David Hildenbrand <david@redhat.com>
---
  .../kernel/syscalls/move_pages/move_pages04.c | 82 ++++++++++++++-----
  1 file changed, 61 insertions(+), 21 deletions(-)

diff --git a/testcases/kernel/syscalls/move_pages/move_pages04.c b/testcases/kernel/syscalls/move_pages/move_pages04.c
index f53453ab4..f51a73b6c 100644
--- a/testcases/kernel/syscalls/move_pages/move_pages04.c
+++ b/testcases/kernel/syscalls/move_pages/move_pages04.c
@@ -26,13 +26,16 @@
   *	move_pages04.c
   *
   * DESCRIPTION
- *      Failure when page does not exit.
+ *      Failure when no page is mapped or the shared zero page is mapped.
   *
   * ALGORITHM
   *
- *      1. Pass zero page (allocated, but not written to) as one of the
- *         page addresses to move_pages().
- *      2. Check if the corresponding status is set to:
+ *      1. Pass the address of a valid memory area where no page is mapped yet
+ *         (not read/written) and the address of the shared zero page
+ *         (read, but not written to) as page addresses to move_pages().
+ *      2. Check if the corresponding status for "no page mapped" is set to
+ *         either -ENOENT or -EFAULT.
+ *      3. Check if the corresponding status for "shared zero page" is set to:
   *         -ENOENT for kernels < 4.3
   *         -EFAULT for kernels >= 4.3 [1]
   *
@@ -64,10 +67,11 @@
  #include "test.h"
  #include "move_pages_support.h"
  
-#define TEST_PAGES 2
+#define TEST_PAGES 3
  #define TEST_NODES 2
  #define TOUCHED_PAGES 1
-#define UNTOUCHED_PAGE (TEST_PAGES - 1)
+#define NO_PAGE (TEST_PAGES - 1)
+#define ZERO_PAGE (NO_PAGE - 1)
  
  void setup(void);
  void cleanup(void);
@@ -89,12 +93,12 @@ int main(int argc, char **argv)
  	int lc;
  	unsigned int from_node;
  	unsigned int to_node;
-	int ret, exp_status;
+	int ret, exp_zero_page_status;
  
  	if ((tst_kvercmp(4, 3, 0)) >= 0)
-		exp_status = -EFAULT;
+		exp_zero_page_status = -EFAULT;
  	else
-		exp_status = -ENOENT;
+		exp_zero_page_status = -ENOENT;
  
  	ret = get_allowed_nodes(NH_MEMS, 2, &from_node, &to_node);
  	if (ret < 0)
@@ -106,6 +110,7 @@ int main(int argc, char **argv)
  		int nodes[TEST_PAGES];
  		int status[TEST_PAGES];
  		unsigned long onepage = get_page_size();
+		char tmp;
  
  		/* reset tst_count in case we are looping */
  		tst_count = 0;
@@ -114,13 +119,30 @@ int main(int argc, char **argv)
  		if (ret == -1)
  			continue;
  
-		/* Allocate page and do not touch it. */
-		pages[UNTOUCHED_PAGE] = numa_alloc_onnode(onepage, from_node);
-		if (pages[UNTOUCHED_PAGE] == NULL) {
-			tst_resm(TBROK, "failed allocating page on node %d",
+		/*
+		 * Allocate memory and do not touch it. Consequently, no
+		 * page will be faulted in / mapped into the page tables.
+		 */
+		pages[NO_PAGE] = numa_alloc_onnode(onepage, from_node);
+		if (pages[NO_PAGE] == NULL) {
+			tst_resm(TBROK, "failed allocating memory on node %d",
+				 from_node);
+			goto err_free_pages;
+		}
+
+		/*
+		 * Allocate memory, read from it, but do not write to it. This
+		 * will populate the shared zeropage.
+		 */
+		pages[ZERO_PAGE] = numa_alloc_onnode(onepage, from_node);
+		if (pages[ZERO_PAGE] == NULL) {
+			tst_resm(TBROK, "failed allocating memory on node %d",
  				 from_node);
  			goto err_free_pages;
  		}
+		/* Make the compiler not optimize-out the read. */
+		tmp = *((char *)pages[ZERO_PAGE]);
+		asm volatile("" : "+r" (tmp));
  
  		for (i = 0; i < TEST_PAGES; i++)
  			nodes[i] = to_node;
@@ -135,20 +157,38 @@ int main(int argc, char **argv)
  			tst_resm(TINFO, "move_pages() returned %d", ret);
  		}
  
-		if (status[UNTOUCHED_PAGE] == exp_status) {
+		switch (status[NO_PAGE]) {
+		case -ENOENT:
+		case -EFAULT:
+			/*
+			 * Before 4.3, the kernel returned -ENOENT. With 4.3
+			 * that behavior was changed by accident to return
+			 * -EFAULT for some mapping types (including anonymous
+			 * memory we use here). Newer kernels are expected to
+			 * change that behavior again back to -ENOENT.
+			 */
  			tst_resm(TPASS, "status[%d] has expected value",
-				 UNTOUCHED_PAGE);
+				 NO_PAGE);
+			break;
+		default:
+			tst_resm(TFAIL, "status[%d] is %s, expected %s or %s",
+				NO_PAGE,
+				tst_strerrno(-status[NO_PAGE]),
+				tst_strerrno(ENOENT), tst_strerrno(EFAULT));
+		}
+
+		if (status[ZERO_PAGE] == exp_zero_page_status) {
+			tst_resm(TPASS, "status[%d] has expected value",
+				 ZERO_PAGE);
  		} else {
  			tst_resm(TFAIL, "status[%d] is %s, expected %s",
-				UNTOUCHED_PAGE,
-				tst_strerrno(-status[UNTOUCHED_PAGE]),
-				tst_strerrno(-exp_status));
+				ZERO_PAGE,
+				tst_strerrno(-status[ZERO_PAGE]),
+				tst_strerrno(-exp_zero_page_status));
  		}
  
  err_free_pages:
-		/* This is capable of freeing both the touched and
-		 * untouched pages.
-		 */
+		/* This is capable of freeing all memory we allocated. */
  		free_pages(pages, TEST_PAGES);
  	}
  #else
-- 
2.46.0


-- 
Cheers,

David / dhildenb