Re: [PATCH] mm/zsmalloc.c: Fix zsmalloc 32-bit PAE support

On Mon, Dec 10, 2018 at 02:35:55PM +0000, Robin Murphy wrote:

On 10/12/2018 14:21, Rafael David Tinoco wrote:

quoted

On 32-bit systems, zsmalloc uses HIGHMEM and, when PAE is enabled, the
physical frame number might be so big that zsmalloc obj encoding (to
location) will break, causing:

BUG: KASAN: null-ptr-deref in zs_map_object+0xa4/0x2bc
Read of size 4 at addr 00000000 by task mkfs.ext4/623
CPU: 2 PID: 623 Comm: mkfs.ext4 Not tainted 4.19.0-rc8-00017-g8239bc6d3307-dirty #15
Hardware name: Generic DT based system
[<c0418f7c>] (unwind_backtrace) from [<c0410ca4>] (show_stack+0x20/0x24)
[<c0410ca4>] (show_stack) from [<c16bd540>] (dump_stack+0xbc/0xe8)
[<c16bd540>] (dump_stack) from [<c06cab74>] (kasan_report+0x248/0x390)
[<c06cab74>] (kasan_report) from [<c06c94e8>] (__asan_load4+0x78/0xb4)
[<c06c94e8>] (__asan_load4) from [<c06ddc24>] (zs_map_object+0xa4/0x2bc)
[<c06ddc24>] (zs_map_object) from [<bf0bbbd8>] (zram_bvec_rw.constprop.2+0x324/0x8e4 [zram])
[<bf0bbbd8>] (zram_bvec_rw.constprop.2 [zram]) from [<bf0bc3cc>] (zram_make_request+0x234/0x46c [zram])
[<bf0bc3cc>] (zram_make_request [zram]) from [<c09aff9c>] (generic_make_request+0x304/0x63c)
[<c09aff9c>] (generic_make_request) from [<c09b0320>] (submit_bio+0x4c/0x1c8)
[<c09b0320>] (submit_bio) from [<c0743570>] (submit_bh_wbc.constprop.15+0x238/0x26c)
[<c0743570>] (submit_bh_wbc.constprop.15) from [<c0746cf8>] (__block_write_full_page+0x524/0x76c)
[<c0746cf8>] (__block_write_full_page) from [<c07472c4>] (block_write_full_page+0x1bc/0x1d4)
[<c07472c4>] (block_write_full_page) from [<c0748eb4>] (blkdev_writepage+0x24/0x28)
[<c0748eb4>] (blkdev_writepage) from [<c064a780>] (__writepage+0x44/0x78)
[<c064a780>] (__writepage) from [<c064b50c>] (write_cache_pages+0x3b8/0x800)
[<c064b50c>] (write_cache_pages) from [<c064dd78>] (generic_writepages+0x74/0xa0)
[<c064dd78>] (generic_writepages) from [<c0748e64>] (blkdev_writepages+0x18/0x1c)
[<c0748e64>] (blkdev_writepages) from [<c064e890>] (do_writepages+0x68/0x134)
[<c064e890>] (do_writepages) from [<c06368a4>] (__filemap_fdatawrite_range+0xb0/0xf4)
[<c06368a4>] (__filemap_fdatawrite_range) from [<c0636b68>] (file_write_and_wait_range+0x64/0xd0)
[<c0636b68>] (file_write_and_wait_range) from [<c0747af8>] (blkdev_fsync+0x54/0x84)
[<c0747af8>] (blkdev_fsync) from [<c0739dac>] (vfs_fsync_range+0x70/0xd4)
[<c0739dac>] (vfs_fsync_range) from [<c0739e98>] (do_fsync+0x4c/0x80)
[<c0739e98>] (do_fsync) from [<c073a26c>] (sys_fsync+0x1c/0x20)
[<c073a26c>] (sys_fsync) from [<c0401000>] (ret_fast_syscall+0x0/0x2c)

when trying to decode (the pfn) and map the object.

That happens because one architecture might not re-define
MAX_PHYSMEM_BITS, like in this ARM 32-bit w/ LPAE enabled example. For
32-bit systems, if not re-defined, MAX_POSSIBLE_PHYSMEM_BITS will
default to BITS_PER_LONG (32) in most cases, and, with PAE enabled,
_PFN_BITS might be wrong: which may cause obj variable to overflow if
frame number is in HIGHMEM and referencing a page above the 4GB
watermark.

commit 6e00ec00b1a7 ("staging: zsmalloc: calculate MAX_PHYSMEM_BITS if
not defined") realized MAX_PHYSMEM_BITS depended on SPARSEMEM headers
and "fixed" it by calculating it using BITS_PER_LONG if SPARSEMEM wasn't
used, like in the example given above.

Systems with potential for PAE exist for a long time and assuming
BITS_PER_LONG seems inadequate. Defining MAX_PHYSMEM_BITS looks better,
however it is NOT a constant anymore for x86.

SO, instead, MAX_POSSIBLE_PHYSMEM_BITS should be defined by every
architecture using zsmalloc, together with a sanity check for
MAX_POSSIBLE_PHYSMEM_BITS being too big on 32-bit systems.

Link: https://bugs.linaro.org/show_bug.cgi?id=3765#c17
Signed-off-by: Rafael David Tinoco <redacted>
---
 arch/arm/include/asm/pgtable-2level-types.h |  2 ++
 arch/arm/include/asm/pgtable-3level-types.h |  2 ++
 arch/arm64/include/asm/pgtable-types.h      |  2 ++
 arch/ia64/include/asm/page.h                |  2 ++
 arch/mips/include/asm/page.h                |  2 ++
 arch/powerpc/include/asm/mmu.h              |  2 ++
 arch/s390/include/asm/page.h                |  2 ++
 arch/sh/include/asm/page.h                  |  2 ++
 arch/sparc/include/asm/page_32.h            |  2 ++
 arch/sparc/include/asm/page_64.h            |  2 ++
 arch/x86/include/asm/pgtable-2level_types.h |  2 ++
 arch/x86/include/asm/pgtable-3level_types.h |  3 +-
 arch/x86/include/asm/pgtable_64_types.h     |  4 +--
 mm/zsmalloc.c                               | 35 +++++++++++----------
 14 files changed, 45 insertions(+), 19 deletions(-)

diff --git a/arch/arm/include/asm/pgtable-2level-types.h b/arch/arm/include/asm/pgtable-2level-types.h
index 66cb5b0e89c5..552dba411324 100644
--- a/arch/arm/include/asm/pgtable-2level-types.h
+++ b/arch/arm/include/asm/pgtable-2level-types.h

@@ -64,4 +64,6 @@ typedef pteval_t pgprot_t;
 #endif /* STRICT_MM_TYPECHECKS */
+#define MAX_POSSIBLE_PHYSMEM_BITS 32
+
 #endif	/* _ASM_PGTABLE_2LEVEL_TYPES_H */

diff --git a/arch/arm/include/asm/pgtable-3level-types.h b/arch/arm/include/asm/pgtable-3level-types.h
index 921aa30259c4..664c39e6517c 100644
--- a/arch/arm/include/asm/pgtable-3level-types.h
+++ b/arch/arm/include/asm/pgtable-3level-types.h

@@ -67,4 +67,6 @@ typedef pteval_t pgprot_t;
 #endif	/* STRICT_MM_TYPECHECKS */
+#define MAX_POSSIBLE_PHYSMEM_BITS 36

Nit: with LPAE, physical addresses go up to 40 bits, not just 36.

Right, with that set at 40, we get:

#define _PFN_BITS               (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)

== 28

#define OBJ_TAG_BITS 1
#define OBJ_INDEX_BITS  (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)

== 3

#define ZS_MAX_ZSPAGE_ORDER 2
#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)

== 4

#define ZS_MIN_ALLOC_SIZE \
        MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))

== 4 << 12 >> 3 = 2048

or half-page allocations.

Given this in Documentation/vm/zsmalloc.rst:

On the other hand, if we just use single
(0-order) pages, it would suffer from very high fragmentation --
any object of size PAGE_SIZE/2 or larger would occupy an entire page.
This was one of the major issues with its predecessor (xvmalloc).

It seems that would not be acceptable, so I have to ask whether
using an unsigned long to store PFN + object ID is really acceptable.
Maybe the real solution to this problem is to stop using an
unsigned long to contain both the PFN and object ID?

-- 
RMK's Patch system: http://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up