| Message ID | 20210508221328.7338-1-42.hyeyoo@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | mm: kmalloc_index: remove case when size is more than 32MB | expand |
On Sun, May 09, 2021 at 07:13:28AM +0900, Hyeonggon Yoo wrote:
> the return value of kmalloc_index is used as index of kmalloc_caches,
it doesn't matter. every few weeks somebody posts a patch to "optimise"
kmalloc_index, failing to appreciate that it's only ever run at compile
time because it's all under __builtin_constant_p().
On Sun, May 09, 2021 at 12:19:40AM +0100, Matthew Wilcox wrote: > On Sun, May 09, 2021 at 07:13:28AM +0900, Hyeonggon Yoo wrote: > > the return value of kmalloc_index is used as index of kmalloc_caches, > > it doesn't matter. every few weeks somebody posts a patch to "optimise" > kmalloc_index, failing to appreciate that it's only ever run at compile > time because it's all under __builtin_constant_p(). Oh thanks, I didn't know about __builtin_constant_p. But I was not optimizing kmalloc_index. isn't it confusing that kmalloc_caches alllows maximum size of 32MB, and kmalloc_index allows maximum size of 64MB? and even if the code I removed is never reached because 64MB is always bigger than KMALLOC_MAX_CACHE_SIZE, it will cause an error if reached.
On 5/9/21 7:33 AM, Hyeonggon Yoo wrote: > On Sun, May 09, 2021 at 12:19:40AM +0100, Matthew Wilcox wrote: >> On Sun, May 09, 2021 at 07:13:28AM +0900, Hyeonggon Yoo wrote: >> > the return value of kmalloc_index is used as index of kmalloc_caches, >> >> it doesn't matter. every few weeks somebody posts a patch to "optimise" >> kmalloc_index, failing to appreciate that it's only ever run at compile >> time because it's all under __builtin_constant_p(). > > Oh thanks, I didn't know about __builtin_constant_p. > > But I was not optimizing kmalloc_index. isn't it confusing that > kmalloc_caches alllows maximum size of 32MB, and kmalloc_index allows > maximum size of 64MB? > > and even if the code I removed is never reached because 64MB is always > bigger than KMALLOC_MAX_CACHE_SIZE, it will cause an error if reached. KMALLOC_MAX_CACHE_SIZE depends on KMALLOC_SHIFT_HIGH size of kmalloc_caches array depends on KMALLOC_SHIFT_HIGH So I don't an easy way how it could become reachable while causing the index to overflow - if someone increased KMALLOC_SHIFT_HIGH from 25 to 26, all should be fine, AFAICS. The problem would be if someone increased it to 27, then we might suddenly get a BUG() in kmalloc_index(). We should probably replace that BUG() with BUILD_BUG_ON(1) to catch that at compile time. Hopefully no supported compiler will break because it's not able to do the proper compile-time evaluation - but if it does, at least we would know. So I would accept the patch if it also changed BUG() to e.g. BUILD_BUG_ON_MSG(1, "unexpected size in kmalloc_index()"); and expanded the function's comment that this is always compile-time evaluated and thus no attempts at "optimizing" the code should be made.
On Mon, May 10, 2021 at 12:09:55PM +0200, Vlastimil Babka wrote: > On 5/9/21 7:33 AM, Hyeonggon Yoo wrote: > > On Sun, May 09, 2021 at 12:19:40AM +0100, Matthew Wilcox wrote: > >> On Sun, May 09, 2021 at 07:13:28AM +0900, Hyeonggon Yoo wrote: > >> > the return value of kmalloc_index is used as index of kmalloc_caches, > >> > >> it doesn't matter. every few weeks somebody posts a patch to "optimise" > >> kmalloc_index, failing to appreciate that it's only ever run at compile > >> time because it's all under __builtin_constant_p(). > > > > Oh thanks, I didn't know about __builtin_constant_p. > > > > But I was not optimizing kmalloc_index. isn't it confusing that > > kmalloc_caches alllows maximum size of 32MB, and kmalloc_index allows > > maximum size of 64MB? > > > > and even if the code I removed is never reached because 64MB is always > > bigger than KMALLOC_MAX_CACHE_SIZE, it will cause an error if reached. > > KMALLOC_MAX_CACHE_SIZE depends on KMALLOC_SHIFT_HIGH > size of kmalloc_caches array depends on KMALLOC_SHIFT_HIGH > > So I don't an easy way how it could become reachable while causing the index to > overflow - if someone increased KMALLOC_SHIFT_HIGH from 25 to 26, all should be > fine, AFAICS. > > The problem would be if someone increased it to 27, then we might suddenly get a > BUG() in kmalloc_index(). We should probably replace that BUG() with > BUILD_BUG_ON(1) to catch that at compile time. Hopefully no supported compiler > will break because it's not able to do the proper compile-time evaluation - but > if it does, at least we would know. > > So I would accept the patch if it also changed BUG() to e.g. BUILD_BUG_ON_MSG(1, > "unexpected size in kmalloc_index()"); > and expanded the function's comment that this is always compile-time evaluated > and thus no attempts at "optimizing" the code should be made. > Thank you so much reviewing and replying to my patch. plecase check if I understood well. Okay, I'll do that work. then the following patch will: - remove case when size is more than 32MB - change "BUG to BUILD_BUG_ON to let compiler know when the size is not supported" - add comment that there's no need to optimize it is it what you mean. right? and I have a question. in the lin 751 of mm/slab_common.c, thre's struct kmalloc_info_struct kmalloc_info. and it initializes kmalloc info up to 64MB, which is currently not supported. should I change it too? in a separate patch?
On 5/10/21 3:58 PM, Hyeonggon Yoo wrote: > On Mon, May 10, 2021 at 12:09:55PM +0200, Vlastimil Babka wrote: >> On 5/9/21 7:33 AM, Hyeonggon Yoo wrote: >> > On Sun, May 09, 2021 at 12:19:40AM +0100, Matthew Wilcox wrote: >> >> On Sun, May 09, 2021 at 07:13:28AM +0900, Hyeonggon Yoo wrote: >> >> > the return value of kmalloc_index is used as index of kmalloc_caches, >> >> >> >> it doesn't matter. every few weeks somebody posts a patch to "optimise" >> >> kmalloc_index, failing to appreciate that it's only ever run at compile >> >> time because it's all under __builtin_constant_p(). >> > >> > Oh thanks, I didn't know about __builtin_constant_p. >> > >> > But I was not optimizing kmalloc_index. isn't it confusing that >> > kmalloc_caches alllows maximum size of 32MB, and kmalloc_index allows >> > maximum size of 64MB? >> > >> > and even if the code I removed is never reached because 64MB is always >> > bigger than KMALLOC_MAX_CACHE_SIZE, it will cause an error if reached. >> >> KMALLOC_MAX_CACHE_SIZE depends on KMALLOC_SHIFT_HIGH >> size of kmalloc_caches array depends on KMALLOC_SHIFT_HIGH >> >> So I don't an easy way how it could become reachable while causing the index to >> overflow - if someone increased KMALLOC_SHIFT_HIGH from 25 to 26, all should be >> fine, AFAICS. >> >> The problem would be if someone increased it to 27, then we might suddenly get a >> BUG() in kmalloc_index(). We should probably replace that BUG() with >> BUILD_BUG_ON(1) to catch that at compile time. Hopefully no supported compiler >> will break because it's not able to do the proper compile-time evaluation - but >> if it does, at least we would know. >> >> So I would accept the patch if it also changed BUG() to e.g. BUILD_BUG_ON_MSG(1, >> "unexpected size in kmalloc_index()"); >> and expanded the function's comment that this is always compile-time evaluated >> and thus no attempts at "optimizing" the code should be made. >> > > Thank you so much reviewing and replying to my patch. > plecase check if I understood well. > > Okay, I'll do that work. then the following patch will: > - remove case when size is more than 32MB > - change "BUG to BUILD_BUG_ON to let compiler know when the size is not supported" > - add comment that there's no need to optimize it > > is it what you mean. right? Exactly. > and I have a question. in the lin 751 of mm/slab_common.c, > thre's struct kmalloc_info_struct kmalloc_info. and it initializes kmalloc info > up to 64MB, which is currently not supported. should I change it too? in a separate patch? Yeah that could be also changed, in the same patch.
diff --git a/include/linux/slab.h b/include/linux/slab.h index 0c97d788762c..4694b1db4cb2 100644 --- a/include/linux/slab.h +++ b/include/linux/slab.h @@ -382,7 +382,6 @@ static __always_inline unsigned int kmalloc_index(size_t size) if (size <= 8 * 1024 * 1024) return 23; if (size <= 16 * 1024 * 1024) return 24; if (size <= 32 * 1024 * 1024) return 25; - if (size <= 64 * 1024 * 1024) return 26; BUG(); /* Will never be reached. Needed because the compiler may complain */
the return value of kmalloc_index is used as index of kmalloc_caches, which is defined as: struct kmem_cache * kmalloc_caches[NR_KMALLOC_TYPES][KMALLOC_SHIFT_HIGH + 1] and KMALLOC_SHIFT_HIGH is defined as: #define KMALLOC_SHIFT_HIGH ((MAX_ORDER + PAGE_SHIFT - 1) <= 25 ? \ (MAX_ORDER + PAGE_SHIFT - 1) : 25) KMALLOC_SHIFT_HIGH is maximum 25 by its definition. thus index of kmalloc_caches cannot be 26. so this case should be removed. Signed-off-by: Hyeonggon Yoo <42.hyeyoo@gmail.com> --- include/linux/slab.h | 1 - 1 file changed, 1 deletion(-)