On Mon, Mar 14, 2022 at 11:57:35PM -0700, Minchan Kim wrote:

On Mon, Mar 14, 2022 at 09:18:43PM -0700, Ivan Babrou wrote:

quoted

On Fri, Mar 11, 2022 at 11:51 AM Ivan Babrou [off-list ref] wrote:

quoted

Hello,

We're looking into using zram, but unfortunately we ran into some
corruption issues. We've seen rocksdb complaining about "Corruption:
bad entry in block", and we've also seen some coredumps that point at
memory being zeroed out. One of our Rust processes coredumps contains
a non-null pointer pointing at zero, among other things:

* core::ptr::non_null::NonNull<u8> {pointer: 0x0}

In fact, a whole bunch of memory around this pointer was all zeros.

Disabling zram resolves all issues, and we can't reproduce any of
these issues with other swap setups. I've tried adding crc32
checksumming for pages that are compressed, but it didn't catch the
issue either, even though userspace facing symptoms were present. My
crc32 code doesn't touch ZRAM_SAME pages, though.

Unfortunately, this isn't trivial to replicate, and I believe that it
depends on zram used for swap specifically, not for zram as a block
device. Specifically, swap_slot_free_notify looks suspicious.

Here's a patch that I have to catch the issue in the act:

diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
index 438ce34ee760..fea46a70a3c9 100644
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c

@@ -1265,6 +1265,9 @@ static int __zram_bvec_read(struct zram *zram,

struct page *page, u32 index,
  unsigned long value;
  void *mem;

+ if (WARN_ON(!handle && !zram_test_flag(zram, index, ZRAM_SAME)))
+ pr_warn("Page %u read from zram without previous write\n", index);
+
  value = handle ? zram_get_element(zram, index) : 0;
  mem = kmap_atomic(page);
  zram_fill_page(mem, PAGE_SIZE, value);

In essence, it warns whenever a page is read from zram that was not
previously written to. To make this work, one needs to zero out zram
prior to running mkswap on it.

I have prepared a GitHub repo with my observations and a reproduction:

* https://github.com/bobrik/zram-corruptor

I'm able to trigger the following in an aarch64 VM with two threads
reading the same memory out of swap:

[ 512.651752][ T7285] ------------[ cut here ]------------
[ 512.652279][ T7285] WARNING: CPU: 0 PID: 7285 at
drivers/block/zram/zram_drv.c:1285 __zram_bvec_read+0x28c/0x2e8 [zram]
[ 512.653923][ T7285] Modules linked in: zram zsmalloc kheaders nfsv3
nfs lockd grace sunrpc xt_conntrack nft_chain_nat xt_MASQUERADE nf_nat
nf_conntrack_netlink nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4
nft_counter xt_addrtype nft_compat nf_tables nfnetlink bridge stp llc
overlay xfs libcrc32c zstd zstd_compress af_packet aes_ce_blk
aes_ce_cipher ghash_ce gf128mul virtio_net sha3_ce net_failover
sha3_generic failover sha512_ce sha512_arm64 sha2_ce sha256_arm64
virtio_mmio virtio_ring qemu_fw_cfg rtc_pl031 virtio fuse ip_tables
x_tables ext4 mbcache crc16 jbd2 nvme nvme_core pci_host_generic
pci_host_common unix [last unloaded: zsmalloc]
[ 512.659238][ T7285] CPU: 0 PID: 7285 Comm: zram-corruptor Tainted: G
W 5.16.0-ivan #1 0877d306c6dc0716835d43cafe4399473d09e406
[ 512.660413][ T7285] Hardware name: linux,dummy-virt (DT)
[ 512.661077][ T7285] pstate: 80400005 (Nzcv daif +PAN -UAO -TCO -DIT
-SSBS BTYPE=--)
[ 512.661788][ T7285] pc : __zram_bvec_read+0x28c/0x2e8 [zram]
[ 512.662099][ T7285] lr : zram_bvec_rw+0x70/0x204 [zram]
[ 512.662422][ T7285] sp : ffffffc01018bac0
[ 512.662720][ T7285] x29: ffffffc01018bae0 x28: ffffff9e4e725280 x27:
ffffff9e4e725280
[ 512.663122][ T7285] x26: ffffff9e4e725280 x25: 00000000000001f6 x24:
0000000100033e6c
[ 512.663601][ T7285] x23: 00000000000001f6 x22: 0000000000000000 x21:
fffffffe7a36d840
[ 512.664252][ T7285] x20: 00000000000001f6 x19: ffffff9e69423c00 x18:
ffffffc010711068
[ 512.664812][ T7285] x17: 0000000000000008 x16: ffffffd34aed51bc x15:
0000000000000000
[ 512.665507][ T7285] x14: 0000000000000a88 x13: 0000000000000000 x12:
0000000000000000
[ 512.666183][ T7285] x11: 0000000100033e6c x10: ffffffc01091d000 x9 :
0000000001000000
[ 512.666627][ T7285] x8 : 0000000000002f10 x7 : 80b75f8fb90b52c4 x6 :
051609fe50833de3
[ 512.667276][ T7285] x5 : 0000000000000000 x4 : 0000000000000000 x3 :
0000000000000000
[ 512.667875][ T7285] x2 : 00000000000001f6 x1 : 00000000000001f6 x0 :
ffffffd305b746af
[ 512.668483][ T7285] Call trace:
[ 512.668682][ T7285] __zram_bvec_read+0x28c/0x2e8 [zram
745969ed35ea0fb382bfd518d6f70e13966e9b52]
[ 512.669405][ T7285] zram_bvec_rw+0x70/0x204 [zram
745969ed35ea0fb382bfd518d6f70e13966e9b52]
[ 512.670066][ T7285] zram_rw_page+0xb4/0x16c [zram
745969ed35ea0fb382bfd518d6f70e13966e9b52]
[ 512.670584][ T7285] bdev_read_page+0x74/0xac
[ 512.670843][ T7285] swap_readpage+0x5c/0x2e4
[ 512.671243][ T7285] do_swap_page+0x2f4/0x988
[ 512.671560][ T7285] handle_pte_fault+0xcc/0x1fc
[ 512.671935][ T7285] handle_mm_fault+0x284/0x4a8
[ 512.672412][ T7285] do_page_fault+0x274/0x428
[ 512.672704][ T7285] do_translation_fault+0x5c/0xf8
[ 512.673083][ T7285] do_mem_abort+0x50/0xc8
[ 512.673293][ T7285] el0_da+0x3c/0x74
[ 512.673549][ T7285] el0t_64_sync_handler+0xc4/0xec
[ 512.673972][ T7285] el0t_64_sync+0x1a4/0x1a8
[ 512.674495][ T7285] ---[ end trace cf983b7507c20343 ]---
[ 512.675359][ T7285] zram: Page 502 read from zram without previous write

I can also trace accesses to zram to catch the unfortunate sequence:

zram_bvec_write index = 502 [cpu = 3, tid = 7286]
zram_free_page index = 502 [cpu = 3, tid = 7286]
zram_bvec_read index = 502 [cpu = 3, tid = 7286]
zram_free_page index = 502 [cpu = 3, tid = 7286] <-- problematic free
zram_bvec_read index = 502 [cpu = 0, tid = 7285] <-- problematic read

With stacks for zram_free_page:

zram_bvec_write index = 502 [cpu = 3, tid = 7286]
zram_free_page  index = 502 [cpu = 3, tid = 7286]

        zram_free_page+0
        $x.97+32
        zram_rw_page+180
        bdev_write_page+124
        __swap_writepage+116
        swap_writepage+160
        pageout+284
        shrink_page_list+2892
        shrink_inactive_list+688
        shrink_lruvec+360
        shrink_node_memcgs+148
        shrink_node+860
        shrink_zones+368
        do_try_to_free_pages+232
        try_to_free_mem_cgroup_pages+292
        try_charge_memcg+608

zram_bvec_read  index = 502 [cpu = 3, tid = 7286]
zram_free_page  index = 502 [cpu = 3, tid = 7286] <-- problematic free

        zram_free_page+0
        swap_range_free+220
        swap_entry_free+244
        swapcache_free_entries+152
        free_swap_slot+288
        __swap_entry_free+216
        swap_free+108
        do_swap_page+1776
        handle_pte_fault+204
        handle_mm_fault+644
        do_page_fault+628
        do_translation_fault+92
        do_mem_abort+80
        el0_da+60
        el0t_64_sync_handler+196
        el0t_64_sync+420

zram_bvec_read  index = 502 [cpu = 0, tid = 7285] <-- problematic read

The very last read is the same one that triggered the warning from my
patch in dmesg. You can see that the slot is freed before reading by
swapcache_free_entries. As far as I can see, only zram implements
swap_slot_free_notify. Swapping in an uninitialized zram page results
in all zeroes copied, which matches the symptoms.

The issue doesn't reproduce if I pin both threads to the same CPU. It
also doesn't reproduce with a single thread. All of this seems to
point at some sort of race condition.

I was able to reproduce this on x86_64 bare metal server as well.

I'm happy to try out mitigation approaches for this. If my
understanding here is incorrect, I'm also happy to try out patches
that could help me catch the issue in the wild.

I poked around the swapping code a bit. In the failing read stack:

[ 1298.167823][ T7004]  swap_readpage+0x60/0x328
[ 1298.168317][ T7004]  do_swap_page+0x438/0x904

You can see that swap_readpage is only called from do_swap_page for
synchronous IO:

if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
    __swap_count(entry) == 1) {
    // ...
    if (page) {
        // ...
        swap_readpage(page, true);

See: https://elixir.bootlin.com/linux/v5.15.28/source/mm/memory.c#L3548

I looked around some more and found 0bcac06f27d7:

* mm, swap: skip swapcache for swapin of synchronous device

Zram is considered fast synchronous storage. Reverting that notion
makes my reproduction not complain anymore:

Yeah, that was the part I was chasing since we had problem there

5df373e95689b, mm/page_io.c: do not free shared swap slots

Initially, I suspected __swap_count race(I still believe it has
swap_slot_free_notify and do_swap_page) and fixed the race
with workaround but the problem still happened. 

Looks like your test program clone the child with CLONE_VM
which never call swap_duplicate to increase swap_map count.
It means the 0bcac06f27d7 and 5df373e95689b couldn't work
with CLONE_VM.

I think reverting them is best at this moment unless someone
has an idea.

I think the problem with CLONE_VM is following race

CPU A                        CPU B

do_swap_page                do_swap_page
SWP_SYNCHRONOUS_IO path     SWP_SYNCHRONOUS_IO path
swap_readpage original data
  swap_slot_free_notify
    delete zram entry
                            swap_readpage zero data
                            pte_lock
                            map the *zero data* to userspace
                            pte_unlock
pte_lock
if (!pte_same)
  goto out_nomap;
pte_unlock
return and next refault will
read zero data

So, CPU A and B see zero data. With patchset below, it changes 


CPU A                        CPU B

do_swap_page                do_swap_page
SWP_SYNCHRONOUS_IO path     SWP_SYNCHRONOUS_IO path
                            swap_readpage original data
                            pte_lock
                            map the original data
                            swap_free
                              swap_range_free
                                bd_disk->fops->swap_slot_free_notify
swap_readpage read zero data
                            pte_unlock
pte_lock
if (!pte_same)
  goto out_nomap;
pte_unlock
return and next refault will
read correct data again

Here, CPU A could read zero data from zram but that's not a bug
(IOW, warning injected doesn't mean bug).

The concern of the patch would increase memory size since it could
increase wasted memory with compressed form in zram and uncompressed
form in address space.  However, most of cases of zram uses no
readahead and then, do_swap_page is followed by swap_free so it will
free the compressed from in zram quickly.

Ivan, with this patch, you can see the warning you added in the zram
but it shouldn't trigger the userspace corruption as mentioned above
if I understand correctly.

Could you test whether the patch prevent userspace broken?

From 71d8b7dce1f5e11fbae4a282ef7d45253ec2087d Mon Sep 17 00:00:00 2001
From: Minchan Kim <minchan@kernel.org>
Date: Tue, 15 Mar 2022 14:14:23 -0700
Subject: [PATCH] fix

Not-Yet-Signed-off-by: Minchan Kim [off-list ref]
---
 mm/page_io.c | 54 ----------------------------------------------------
 1 file changed, 54 deletions(-)

diff --git a/mm/page_io.c b/mm/page_io.c
index 0bf8e40f4e57..d3eea0a3f1af 100644
--- a/mm/page_io.c
+++ b/mm/page_io.c

@@ -51,54 +51,6 @@ void end_swap_bio_write(struct bio *bio)
 	bio_put(bio);
 }
 
-static void swap_slot_free_notify(struct page *page)
-{
-	struct swap_info_struct *sis;
-	struct gendisk *disk;
-	swp_entry_t entry;
-
-	/*
-	 * There is no guarantee that the page is in swap cache - the software
-	 * suspend code (at least) uses end_swap_bio_read() against a non-
-	 * swapcache page.  So we must check PG_swapcache before proceeding with
-	 * this optimization.
-	 */
-	if (unlikely(!PageSwapCache(page)))
-		return;
-
-	sis = page_swap_info(page);
-	if (data_race(!(sis->flags & SWP_BLKDEV)))
-		return;
-
-	/*
-	 * The swap subsystem performs lazy swap slot freeing,
-	 * expecting that the page will be swapped out again.
-	 * So we can avoid an unnecessary write if the page
-	 * isn't redirtied.
-	 * This is good for real swap storage because we can
-	 * reduce unnecessary I/O and enhance wear-leveling
-	 * if an SSD is used as the as swap device.
-	 * But if in-memory swap device (eg zram) is used,
-	 * this causes a duplicated copy between uncompressed
-	 * data in VM-owned memory and compressed data in
-	 * zram-owned memory.  So let's free zram-owned memory
-	 * and make the VM-owned decompressed page *dirty*,
-	 * so the page should be swapped out somewhere again if
-	 * we again wish to reclaim it.
-	 */
-	disk = sis->bdev->bd_disk;
-	entry.val = page_private(page);
-	if (disk->fops->swap_slot_free_notify && __swap_count(entry) == 1) {
-		unsigned long offset;
-
-		offset = swp_offset(entry);
-
-		SetPageDirty(page);
-		disk->fops->swap_slot_free_notify(sis->bdev,
-				offset);
-	}
-}
-
 static void end_swap_bio_read(struct bio *bio)
 {
 	struct page *page = bio_first_page_all(bio);

@@ -114,7 +66,6 @@ static void end_swap_bio_read(struct bio *bio)
 	}
 
 	SetPageUptodate(page);
-	swap_slot_free_notify(page);
 out:
 	unlock_page(page);
 	WRITE_ONCE(bio->bi_private, NULL);

@@ -392,11 +343,6 @@ int swap_readpage(struct page *page, bool synchronous)
 	if (sis->flags & SWP_SYNCHRONOUS_IO) {
 		ret = bdev_read_page(sis->bdev, swap_page_sector(page), page);
 		if (!ret) {
-			if (trylock_page(page)) {
-				swap_slot_free_notify(page);
-				unlock_page(page);
-			}
-
 			count_vm_event(PSWPIN);
 			goto out;
 		}

-- 
2.35.1.723.g4982287a31-goog