Skip to contentSkip to conversation

selftests: bpf: mmap question

DISCUSSION Yauheni Kaliuta 2 participants 4 messages Jul 23, 2020 – Aug 11, 2020 BPF

lore
Share & tools
lore
#0 lore raw

Hi!

I have a question about the part of the test:

	/* check some more advanced mmap() manipulations */

	/* map all but last page: pages 1-3 mapped */
	tmp1 = mmap(NULL, 3 * page_size, PROT_READ, MAP_SHARED,
			  data_map_fd, 0);
	if (CHECK(tmp1 == MAP_FAILED, "adv_mmap1", "errno %d\n", errno))
		goto cleanup;

	/* unmap second page: pages 1, 3 mapped */
	err = munmap(tmp1 + page_size, page_size);
	if (CHECK(err, "adv_mmap2", "errno %d\n", errno)) {
		munmap(tmp1, map_sz);
		goto cleanup;
	}

	/* map page 2 back */
	tmp2 = mmap(tmp1 + page_size, page_size, PROT_READ,
		    MAP_SHARED | MAP_FIXED, data_map_fd, 0);
	if (CHECK(tmp2 == MAP_FAILED, "adv_mmap3", "errno %d\n", errno)) {
		munmap(tmp1, page_size);
		munmap(tmp1 + 2*page_size, page_size);
		goto cleanup;
	}
	CHECK(tmp1 + page_size != tmp2, "adv_mmap4",
	      "tmp1: %p, tmp2: %p\n", tmp1, tmp2);

	/* re-map all 4 pages */
	tmp2 = mmap(tmp1, 4 * page_size, PROT_READ, MAP_SHARED | MAP_FIXED,
		    data_map_fd, 0);
	if (CHECK(tmp2 == MAP_FAILED, "adv_mmap5", "errno %d\n", errno)) {
		munmap(tmp1, 3 * page_size); /* unmap page 1 */
		goto cleanup;
	}
	CHECK(tmp1 != tmp2, "adv_mmap6", "tmp1: %p, tmp2: %p\n", tmp1, tmp2);

In my configuration the first mapping

	/* map all but last page: pages 1-3 mapped */
	tmp1 = mmap(NULL, 3 * page_size, PROT_READ, MAP_SHARED,
			  data_map_fd, 0);

maps the area to the 3 pages right before the TLS page. I find it's pretty ok.

But then the 4 page mapping

	/* re-map all 4 pages */
	tmp2 = mmap(tmp1, 4 * page_size, PROT_READ, MAP_SHARED | MAP_FIXED,
		    data_map_fd, 0);

since it has MAP_FIXED flag, unmaps TLS and maps the former TLS address BPF map.

Which is again exactly the behaviour of MAP_FIXED, but it breaks the test.

Using MAP_FIXED_NOREPLACE fails the check:

CHECK(tmp1 != tmp2, "adv_mmap6", "tmp1: %p, tmp2: %p\n", tmp1, tmp2);

as expected.

Should the test be modified to be a bit more relaxed? Since the kernel behaviour looks correct or I'm missing something?

PS: BTW, the previous data_map mapping left unmmaped. Is it expected?

Signature · 2 lines
WBR,
Yauheni Kaliuta
#1 lore raw
Quoted 5 lines · Yauheni Kaliuta

Hi!

I have a question about the part of the test:

[...]

Quoted 10 lines

In my configuration the first mapping

        /* map all but last page: pages 1-3 mapped */
        tmp1 = mmap(NULL, 3 * page_size, PROT_READ, MAP_SHARED,
                          data_map_fd, 0);

maps the area to the 3 pages right before the TLS page. I find it's pretty ok.

Hm... I never ran into this problem. The point here is to be able to re-mmap partial ranges. One way would be to re-write all those manipulations to start with a full range map, and then do partial un-mmaps/re-mmaps, eventually just re-mmaping everything back. I think that would work, right, as long as we never unmmap the last page? Do you mind trying to fix the test in such a fashion?

Quoted 18 lines

But then the 4 page mapping

        /* re-map all 4 pages */
        tmp2 = mmap(tmp1, 4 * page_size, PROT_READ, MAP_SHARED | MAP_FIXED,
                    data_map_fd, 0);

since it has MAP_FIXED flag, unmaps TLS and maps the former TLS address BPF map.

Which is again exactly the behaviour of MAP_FIXED, but it breaks the test.

Using MAP_FIXED_NOREPLACE fails the check:

CHECK(tmp1 != tmp2, "adv_mmap6", "tmp1: %p, tmp2: %p\n", tmp1, tmp2);

Quoted 8 lines

as expected.

Should the test be modified to be a bit more relaxed? Since the kernel behaviour looks correct or I'm missing something?

PS: BTW, the previous data_map mapping left unmmaped. Is it expected?

Not intentional, the idea is that each test exits with a clean state.

Quoted 5 lines

--
WBR,
Yauheni Kaliuta

#2 lore raw

Hi!

Quoted 22 lines · Andrii Nakryiko
Quoted 5 lines · Yauheni Kaliuta

Hi!

I have a question about the part of the test:

[...]

Quoted 10 lines

In my configuration the first mapping

        /* map all but last page: pages 1-3 mapped */
        tmp1 = mmap(NULL, 3 * page_size, PROT_READ, MAP_SHARED,
                          data_map_fd, 0);

maps the area to the 3 pages right before the TLS page. I find it's pretty ok.

Quoted 6 lines

Hm... I never ran into this problem. The point here is to be able to re-mmap partial ranges. One way would be to re-write all those manipulations to start with a full range map, and then do partial un-mmaps/re-mmaps, eventually just re-mmaping everything back. I think that would work, right, as long as we never unmmap the last page? Do you mind trying to fix the test in such a fashion?

Sorry for the delay. I don't sure, sending the patch.

#3 lore raw
Quoted 27 lines · Yauheni Kaliuta

Hi!

Quoted 22 lines · Andrii Nakryiko
Quoted 5 lines · Yauheni Kaliuta

Hi!

I have a question about the part of the test:

[...]

Quoted 10 lines

In my configuration the first mapping

        /* map all but last page: pages 1-3 mapped */
        tmp1 = mmap(NULL, 3 * page_size, PROT_READ, MAP_SHARED,
                          data_map_fd, 0);

maps the area to the 3 pages right before the TLS page. I find it's pretty ok.

Quoted 9 lines
Quoted 6 lines

Hm... I never ran into this problem. The point here is to be able to re-mmap partial ranges. One way would be to re-write all those manipulations to start with a full range map, and then do partial un-mmaps/re-mmaps, eventually just re-mmaping everything back. I think that would work, right, as long as we never unmmap the last page? Do you mind trying to fix the test in such a fashion?

Sorry for the delay. I don't sure, sending the patch.

No worries! Meanwhile there has been an alternative fix for the same bug: [0]!

[0] https://patchwork.ozlabs.org/project/netdev/patch/[email protected]/