malloc() alignment on 32 bit
Hello,
on an iMX6ull I stumpled across
| zstd_decomp_init:536 workspace=8ff1a004+161320
| ERROR: initcall ubifs_init+0x1/0xc4 failed: Invalid argument
which is caused by
| static int zstd_decomp_init(void) | void *wksp = malloc(wksp_size); | ... | ZSTD_DCtx* ZSTD_initStaticDCtx(void *workspace, size_t workspaceSize) | if ((size_t)workspace & 7) return NULL; /* 8-aligned */
Trivial fix would be 'memalign(8, wksp_size)', but is it really ok that malloc() for 32 bit has only an alignment of 4?
Relevant code seems to be in common/tlsf.c
| enum tlsf_private
| {
| #if defined (TLSF_64BIT)
| /* All allocation sizes and addresses are aligned to 8 bytes. */
| ALIGN_SIZE_LOG2 = 3,
| #else
| /* All allocation sizes and addresses are aligned to 4 bytes. */
| ALIGN_SIZE_LOG2 = 2,
| #endif'ldrd/strd' require 8 byte alignment which might break with such alignment.
Enrico
Quoted 31 lines · Enrico Scholz
Hello,
on an iMX6ull I stumpled across
| zstd_decomp_init:536 workspace=8ff1a004+161320
| ERROR: initcallubifs_init+0x1/0xc4failed: Invalid argumentwhich is caused by
| static int zstd_decomp_init(void) | void *wksp = malloc(wksp_size); | ... | ZSTD_DCtx* ZSTD_initStaticDCtx(void *workspace, size_t workspaceSize) | if ((size_t)workspace & 7) return NULL; /* 8-aligned */Trivial fix would be '
memalign(8, wksp_size)', but is it really ok thatmalloc()for 32 bit has only an alignment of 4?Relevant code seems to be in
common/tlsf.c| enum tlsf_private | { | #if defined (TLSF_64BIT) | /* All allocation sizes and addresses are aligned to 8 bytes. */ | ALIGN_SIZE_LOG2 = 3, | #else | /* All allocation sizes and addresses are aligned to 4 bytes. */ | ALIGN_SIZE_LOG2 = 2, | #endif
Quoted 2 lines
'ldrd/strd' require 8 byte alignment which might break with such alignment.
I recently learnt too that on 32-bit TLSF only has 4 byte alignment. I also think that this is too low. 8 byte alignment sounds good IMO.
Cheers,
Ahmad
Quoted 5 lines
Enrico
Signature · 4 lines
Pengutronix e.K. | | Steuerwalder Str. 21 | http://www.pengutronix.de/ | 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Hi Enrico,
Quoted 31 lines · Enrico Scholz
Hello,
on an iMX6ull I stumpled across
| zstd_decomp_init:536 workspace=8ff1a004+161320
| ERROR: initcallubifs_init+0x1/0xc4failed: Invalid argumentwhich is caused by
| static int zstd_decomp_init(void) | void *wksp = malloc(wksp_size); | ... | ZSTD_DCtx* ZSTD_initStaticDCtx(void *workspace, size_t workspaceSize) | if ((size_t)workspace & 7) return NULL; /* 8-aligned */Trivial fix would be '
memalign(8, wksp_size)', but is it really ok thatmalloc()for 32 bit has only an alignment of 4?Relevant code seems to be in
common/tlsf.c| enum tlsf_private | { | #if defined (TLSF_64BIT) | /* All allocation sizes and addresses are aligned to 8 bytes. */ | ALIGN_SIZE_LOG2 = 3, | #else | /* All allocation sizes and addresses are aligned to 4 bytes. */ | ALIGN_SIZE_LOG2 = 2, | #endif
Quoted 2 lines
'ldrd/strd' require 8 byte alignment which might break with such alignment.
If you had asked me which alignment we have then I would have said it's bigger. OTOH I never received any reports about insufficient alignment on ARM or any other 32bit architecture.
I suspect we could just drop the check without any harm, but that's just a gut feeling because we never had any alignment issues.
BTW are you sure ldrd/strd need 8 byte alignment? I just tested it with the following patch and this works without problems. I verified the compiler indeed generates ldrd/strd for accessing the 64bit field.
Sascha
-------------------------8<----------------------------
common/startup.c +24 -0
@@ -334,10 +334,31 @@ static void do_ctors(void)int (*barebox_main)(void ); +struct bar { +uint64_t foo; +}; + +struct bar *myfoo (void ) +{ +struct bar *x; +void *ptr; + + ptr =malloc (16 ); + + ptr = (void *)((unsigned long )ptr |4 ); + + x = ptr; + + x->foo =get_time_ns (); + +return x; +} +void __noreturn start_barebox (void ) {initcall_t *initcall;int result; +struct bar *b;if (!IS_ENABLED (CONFIG_SHELL_NONE ) &&IS_ENABLED (CONFIG_COMMAND_SUPPORT )) barebox_main = run_init;
@@ -355,6 +376,9 @@ void __noreturn start_barebox(void)pr_debug ("initcalls done\n" ); + b =myfoo (); +printf ("V: %lld\n" , b->foo); +if (IS_ENABLED (CONFIG_SELFTEST_AUTORUN ))selftests_run ();
Signature · 4 lines
Pengutronix e.K. | | Steuerwalder Str. 21 | http://www.pengutronix.de/ | 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Quoted 6 lines · Sascha Hauer
Quoted 3 lines
| zstd_decomp_init:536 workspace=8ff1a004+161320
| ERROR: initcallubifs_init+0x1/0xc4failed: Invalid argumentIf you had asked me which alignment we have then I would have said it's bigger. OTOH I never received any reports about insufficient alignment on ARM or any other 32bit architecture.
The code which failed for me was added 3 months ago
| commit b4a9782d4f56333e897dccc35c2c27e2605f6b93 | Author: Ahmad Fatoum <[email protected]> | Date: Wed Jul 13 12:09:18 2022 +0200 | | lib: zstd: sync with Linux
and the Kconfig options (FS_UBIFS_COMPRESSION_ZSTD) are set to "off"...
Quoted 4 lines
I suspect we could just drop the check without any harm, but that's just a gut feeling because we never had any alignment issues.
BTW are you sure ldrd/strd need 8 byte alignment?
Their EX variants (LDREXD + STREXD; see [1]). Unaligned access on plain LDRD/STRD is allowed on ARMv7-A. But not on ARMv7-M or ARMv6 and earlier.
Enrico
Quoted 11 lines · Enrico Scholz
Quoted 6 lines · Sascha Hauer
Quoted 3 lines
| zstd_decomp_init:536 workspace=8ff1a004+161320
| ERROR: initcallubifs_init+0x1/0xc4failed: Invalid argumentIf you had asked me which alignment we have then I would have said it's bigger. OTOH I never received any reports about insufficient alignment on ARM or any other 32bit architecture.
The code which failed for me was added 3 months ago
Quoted 7 lines
| commit b4a9782d4f56333e897dccc35c2c27e2605f6b93 | Author: Ahmad Fatoum <[email protected]> | Date: Wed Jul 13 12:09:18 2022 +0200 | | lib: zstd: sync with Linuxand the
Kconfigoptions (FS_UBIFS_COMPRESSION_ZSTD) are set to "off"...
When I said I received no reports about insufficient malloc alignment I meant reports about erroneous accesses, like wrong data read/writes or data aborts.
Quoted 10 lines
Quoted 6 lines
I suspect we could just drop the check without any harm, but that's just a gut feeling because we never had any alignment issues.
BTW are you sure ldrd/strd need 8 byte alignment?
Their EX variants (LDREXD + STREXD; see [1]). Unaligned access on plain LDRD/STRD is allowed on ARMv7-A. But not on ARMv7-M or ARMv6 and earlier.
Ok, I found this for the LDRD instruction:
| Prior to ARMv6, if the memory address is not 64-bit aligned, the data read from memory is | UNPREDICTABLE. Alignment checking (taking a data abort), and support for a big-endian | (BE-32) data format are implementation options.
So it seems it's really a good idea to increase malloc alignment accordingly.
Sascha
Signature · 4 lines
Pengutronix e.K. | | Steuerwalder Str. 21 | http://www.pengutronix.de/ | 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
Quoted 2 lines · Sascha Hauer
So it seems it's really a good idea to increase malloc alignment accordingly.
But this does not seem to be trivial :(
A naive
| - ALIGN_SIZE_LOG2 = 2,
| + ALIGN_SIZE_LOG2 = 3,
triggers a lot of warnings.
I think, ALIGN_SIZE in tlsf_add_pool() and in the higher level functions (tlsf_malloc(), _memalign(), _realloc()) should be replaced by a new, bigger constant.
But this appears invasive and I have no idea about tlsf and the sideeffects :(
There exists https://github.com/mattconte/tlsf/issues/16 which asks for proper alignment; but project seems to be unmaintained.
Replacing 'malloc()' with 'memalign()' on problems, seems the best idea for now.
Enrico