Re: [PATCH] Properly align memory allocations and temporary buffers

[Jessica Clarke <jrtc27@jrtc27.com>]

On 7 Jan 2022, at 16:08, René Scharfe <l.s.r@web.de> wrote:
> 
> Am 05.01.22 um 14:23 schrieb Jessica Clarke:
>> Currently git_qsort_s allocates a buffer on the stack that has no
>> alignment, and mem_pool_alloc assumes uintmax_t's size is adequate
>> alignment for any type.
>> 
>> On CHERI, and thus Arm's Morello prototype, pointers are implemented as
>> hardware capabilities which, as well as having a normal integer address,
>> have additional bounds, permissions and other metadata in a second word,
>> so on a 64-bit architecture they are 128-bit quantities, including their
>> alignment requirements. Despite being 128-bit, their integer component
>> is still only a 64-bit field, so uintmax_t remains 64-bit, and therefore
>> uintmax_t does not sufficiently align an allocation.
>> 
>> Moreover, these capabilities have an additional "129th" tag bit, which
>> tracks the validity of the capability and is cleared on any invalid
>> operation that doesn't trap (e.g. partially overwriting a capability
>> will invalidate it) which, combined with the architecture's strict
>> checks on capability manipulation instructions, ensures it is
>> architecturally impossible to construct a capability that gives more
>> rights than those you were given in the first place. To store these tag
>> bits, each capability sized and aligned word in memory gains a single
>> tag bit that is stored in unaddressable (to the processor) memory. This
>> means that it is impossible to store a capability at an unaligned
>> address: a normal load or store of a capability will always take an
>> alignment fault even if the (micro)architecture supports unaligned
>> loads/stores for other data types, and a memcpy will, if the destination
>> is not appropriately aligned, copy the byte representation but lose the
>> tag, meaning that if it is eventually copied back and loaded from an
>> aligned location any attempt to dereference it will trap with a tag
>> fault. Thus, even char buffers that are memcpy'ed to or from must be
>> properly aligned on CHERI architectures if they are to hold pointers.
>> 
>> Address both of these by introducing a new git_max_align type put in a
>> union with the on-stack buffer to force its alignment, as well as a new
>> GIT_MAX_ALIGNMENT macro whose value is the alignment of git_max_align
>> that gets used for mem_pool_alloc. As well as making the code work on
>> CHERI, the former change likely also improves performance on some
>> architectures by making memcpy faster (either because it can use larger
>> block sizes or because the microarchitecture has inefficient unaligned
>> accesses).
>> 
>> Signed-off-by: Jessica Clarke <jrtc27@jrtc27.com>
>> ---
>> compat/qsort_s.c  | 11 +++++++----
>> git-compat-util.h | 11 +++++++++++
>> mem-pool.c        |  6 +++---
>> 3 files changed, 21 insertions(+), 7 deletions(-)
>> 
>> diff --git a/compat/qsort_s.c b/compat/qsort_s.c
>> index 52d1f0a73d..1ccdb87451 100644
>> --- a/compat/qsort_s.c
>> +++ b/compat/qsort_s.c
>> @@ -49,16 +49,19 @@ int git_qsort_s(void *b, size_t n, size_t s,
>> 		int (*cmp)(const void *, const void *, void *), void *ctx)
>> {
>> 	const size_t size = st_mult(n, s);
>> -	char buf[1024];
>> +	union {
>> +		char buf[1024];
>> +		git_max_align align;
>> +	} u;
>> 
>> 	if (!n)
>> 		return 0;
>> 	if (!b || !cmp)
>> 		return -1;
>> 
>> -	if (size < sizeof(buf)) {
>> -		/* The temporary array fits on the small on-stack buffer. */
>> -		msort_with_tmp(b, n, s, cmp, buf, ctx);
>> +	if (size < sizeof(u.buf)) {
>> +		/* The temporary array fits in the small on-stack buffer. */
>> +		msort_with_tmp(b, n, s, cmp, u.buf, ctx);
> 
> So buf gets maximum alignment instead of char alignment (i.e. none)
> because some callers use it to sort pointers, which need that on your
> platform.  Makes sense.

Yes, exactly.

>> 	} else {
>> 		/* It's somewhat large, so malloc it.  */
>> 		char *tmp = xmalloc(size);
> 
> tmp is used instead of buf if the latter is not big enough, so it can
> also contain pointers.  No problem, because malloc(3) returns memory
> that is properly aligned for anything already.

Yes.

> stable-qsort.c uses the same algorithm as compat/qsort_s.c, it just
> lacks the context pointer.  Shouldn't it get the same treatment?  It
> is used e.g. (via the macro STABLE_QSORT) in merge-ort.c to sort
> pointers..

Indeed it should; I guess git_stable_qsort just isn’t used for most
common commands, or at least not with structures containing pointers,
as I haven’t seen it break yet.

>> diff --git a/git-compat-util.h b/git-compat-util.h
>> index 5fa54a7afe..28581a45c5 100644
>> --- a/git-compat-util.h
>> +++ b/git-compat-util.h
>> @@ -274,6 +274,17 @@ typedef unsigned long uintptr_t;
>> #define _ALL_SOURCE 1
>> #endif
>> 
>> +typedef union {
>> +	uintmax_t max_align_uintmax;
>> +	void *max_align_pointer;
>> +} git_max_align;
> 
> For your purposes just the void * member would suffice here, right?  And
> with the added uintmax_t this currently gets maximum alignment, suitable
> for any of our objects.  If we were to start using __int128 etc. then
> we'd have to add that to this union as well to really get the maximum
> possible alignment, though.

For us, yes, uintmax_t’s alignment is never greater than void *’s. On
most 32-bit non-CHERI architectures though that isn’t the case, so
whilst omitting uintmax_t would be fine for qsort, it would break
mem_pool_alloc’s alignment, as this gets used for both. Plus aligning
to uintmax_t may still help performance for memcpy there (e.g. maybe it
guarantees the memcpy implementation can use a load/store pair
instruction that it wouldn’t otherwise necessarily be able to).

Regarding __int128, yes, you would, though both GCC’s and FreeBSD’s
stddef.h don’t admit __int128 exists. They do include long double in
their max_align_t, though, which generally has the same alignment
requirements as __int128 on 64-bit architectures, though I have a funny
feeling it might not on some.

>> +
>> +typedef struct {
>> +	char unalign;
>> +	git_max_align aligned;
>> +} git_max_alignment;
>> +#define GIT_MAX_ALIGNMENT offsetof(git_max_alignment, aligned)
> 
> C11 has alignas, alignof and max_align_t.  We only recently started to
> depend on some C99 features, so perhaps it's a bit early to use
> stdalign.h in Git's code base.  That's a pity, though.  The
> GIT_MAX_ALIGNMENT macro is sightly enough, but using a union to get
> pointer alignment is a bit more cumbersome than something like
> 
> 	alignas(alignof(max_align_t)) char buf[1024];

It is; max_align_t is our preferred solution and what our programming
guide recommends, but not everyone is ready to jump to C11, so until
then various bodges like this are needed.

Jess

>> +
>> /* used on Mac OS X */
>> #ifdef PRECOMPOSE_UNICODE
>> #include "compat/precompose_utf8.h"
>> diff --git a/mem-pool.c b/mem-pool.c
>> index ccdcad2e3d..748eff925a 100644
>> --- a/mem-pool.c
>> +++ b/mem-pool.c
>> @@ -69,9 +69,9 @@ void *mem_pool_alloc(struct mem_pool *pool, size_t len)
>> 	struct mp_block *p = NULL;
>> 	void *r;
>> 
>> -	/* round up to a 'uintmax_t' alignment */
>> -	if (len & (sizeof(uintmax_t) - 1))
>> -		len += sizeof(uintmax_t) - (len & (sizeof(uintmax_t) - 1));
>> +	/* round up to a 'GIT_MAX_ALIGNMENT' alignment */
>> +	if (len & (GIT_MAX_ALIGNMENT - 1))
>> +		len += GIT_MAX_ALIGNMENT - (len & (GIT_MAX_ALIGNMENT - 1));
> 
> OK.
> 
>> 
>> 	if (pool->mp_block &&
>> 	    pool->mp_block->end - pool->mp_block->next_free >= len)