Re: [RFC PATCH 1/1] mm/percpu.c: fix several trivial issues
On 2016/10/14 8:34, Tejun Heo wrote: > On Tue, Oct 11, 2016 at 09:29:27PM +0800, zijun_hu wrote: >> From: zijun_hu >> >> as shown by pcpu_setup_first_chunk(), the first chunk is same as the >> reserved chunk if the reserved size is nonzero but the dynamic is zero >> this special scenario is referred as the special case by below content >> >> fix several trivial issues: >> >> 1) correct or fix several comments >> the LSB of a chunk->map element is used as free/in-use flag and is cleared >> for free area and set for in-use, rather than use positive/negative number >> to mark area state. >> >> 2) change atomic size to PAGE_SIZE for consistency when CONFIG_SMP == n >> both default setup_per_cpu_areas() and pcpu_page_first_chunk() >> use PAGE_SIZE as atomic size when CONFIG_SMP == y; however >> setup_per_cpu_areas() allocates memory for the only unit with alignment >> PAGE_SIZE but assigns unit size to atomic size when CONFIG_SMP == n, so the >> atomic size isn't consistent with either the alignment or the SMP ones. >> fix it by changing atomic size to PAGE_SIZE when CONFIG_SMP == n >> >> 3) correct empty and populated pages statistic error >> in order to service dynamic atomic memory allocation, the number of empty >> and populated pages of chunks is counted to maintain above a low threshold. >> however, for the special case, the first chunk is took into account by >> pcpu_setup_first_chunk(), it is meaningless since the chunk don't include >> any dynamic areas. >> fix it by excluding the reserved chunk before statistic as the other >> contexts do. >> >> 4) fix potential memory leakage for percpu_init_late() >> in order to manage chunk->map memory uniformly, for the first and reserved >> chunks, percpu_init_late() will allocate memory to replace the static >> chunk->map array within section .init.data after slab is brought up >> however, for the special case, memory are allocated for the same chunk->map >> twice since the first chunk reference is same as the reserved, so the >> memory allocated at the first time are leaked obviously. >> fix it by eliminating the second memory allocation under the special case >> >> Signed-off-by: zijun_hu > > Can you please break the changes into separate patches? > > Thanks. >
Re: [RFC PATCH 1/1] mm/percpu.c: fix several trivial issues
On 2016/10/14 8:34, Tejun Heo wrote: > On Tue, Oct 11, 2016 at 09:29:27PM +0800, zijun_hu wrote: >> From: zijun_hu >> >> as shown by pcpu_setup_first_chunk(), the first chunk is same as the >> reserved chunk if the reserved size is nonzero but the dynamic is zero >> this special scenario is referred as the special case by below content >> >> fix several trivial issues: >> >> 1) correct or fix several comments >> the LSB of a chunk->map element is used as free/in-use flag and is cleared >> for free area and set for in-use, rather than use positive/negative number >> to mark area state. >> >> 2) change atomic size to PAGE_SIZE for consistency when CONFIG_SMP == n >> both default setup_per_cpu_areas() and pcpu_page_first_chunk() >> use PAGE_SIZE as atomic size when CONFIG_SMP == y; however >> setup_per_cpu_areas() allocates memory for the only unit with alignment >> PAGE_SIZE but assigns unit size to atomic size when CONFIG_SMP == n, so the >> atomic size isn't consistent with either the alignment or the SMP ones. >> fix it by changing atomic size to PAGE_SIZE when CONFIG_SMP == n >> >> 3) correct empty and populated pages statistic error >> in order to service dynamic atomic memory allocation, the number of empty >> and populated pages of chunks is counted to maintain above a low threshold. >> however, for the special case, the first chunk is took into account by >> pcpu_setup_first_chunk(), it is meaningless since the chunk don't include >> any dynamic areas. >> fix it by excluding the reserved chunk before statistic as the other >> contexts do. >> >> 4) fix potential memory leakage for percpu_init_late() >> in order to manage chunk->map memory uniformly, for the first and reserved >> chunks, percpu_init_late() will allocate memory to replace the static >> chunk->map array within section .init.data after slab is brought up >> however, for the special case, memory are allocated for the same chunk->map >> twice since the first chunk reference is same as the reserved, so the >> memory allocated at the first time are leaked obviously. >> fix it by eliminating the second memory allocation under the special case >> >> Signed-off-by: zijun_hu > > Can you please break the changes into separate patches? yes, i can could you give many comments for that trivial issues firstly? i will separate and product which you thinks reasonable formally. thanks > > Thanks. >
Re: [RFC PATCH 1/1] mm/percpu.c: fix several trivial issues
On Tue, Oct 11, 2016 at 09:29:27PM +0800, zijun_hu wrote: > From: zijun_hu > > as shown by pcpu_setup_first_chunk(), the first chunk is same as the > reserved chunk if the reserved size is nonzero but the dynamic is zero > this special scenario is referred as the special case by below content > > fix several trivial issues: > > 1) correct or fix several comments > the LSB of a chunk->map element is used as free/in-use flag and is cleared > for free area and set for in-use, rather than use positive/negative number > to mark area state. > > 2) change atomic size to PAGE_SIZE for consistency when CONFIG_SMP == n > both default setup_per_cpu_areas() and pcpu_page_first_chunk() > use PAGE_SIZE as atomic size when CONFIG_SMP == y; however > setup_per_cpu_areas() allocates memory for the only unit with alignment > PAGE_SIZE but assigns unit size to atomic size when CONFIG_SMP == n, so the > atomic size isn't consistent with either the alignment or the SMP ones. > fix it by changing atomic size to PAGE_SIZE when CONFIG_SMP == n > > 3) correct empty and populated pages statistic error > in order to service dynamic atomic memory allocation, the number of empty > and populated pages of chunks is counted to maintain above a low threshold. > however, for the special case, the first chunk is took into account by > pcpu_setup_first_chunk(), it is meaningless since the chunk don't include > any dynamic areas. > fix it by excluding the reserved chunk before statistic as the other > contexts do. > > 4) fix potential memory leakage for percpu_init_late() > in order to manage chunk->map memory uniformly, for the first and reserved > chunks, percpu_init_late() will allocate memory to replace the static > chunk->map array within section .init.data after slab is brought up > however, for the special case, memory are allocated for the same chunk->map > twice since the first chunk reference is same as the reserved, so the > memory allocated at the first time are leaked obviously. > fix it by eliminating the second memory allocation under the special case > > Signed-off-by: zijun_hu Can you please break the changes into separate patches? Thanks. -- tejun