We observed an issue with NXP 5.15 LTS kernel that dma_alloc_coherent() may fail sometimes when there're multiple processes trying to allocate CMA memory.
This issue can be very easily reproduced on MX6Q SDB board with latest linux-next kernel by writing a test module creating 16 or 32 threads allocating random size of CMA memory in parallel at the background. Or simply enabling CONFIG_CMA_DEBUG, you can see endless of CMA alloc retries during booting: [ 1.452124] cma: cma_alloc(): memory range at (ptrval) is busy,retrying .... (thousands of reties) NOTE: MX6 has CONFIG_FORCE_MAX_ZONEORDER=14 which means MAX_ORDER is 13 (32M).
The root cause of this issue is that since commit a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock"), CMA supports concurrent memory allocation. It's possible that the pageblock process A try to alloc has already been isolated by the allocation of process B during memory migration.
When there're multi process allocating CMA memory in parallel, it's likely that other the remain pageblocks may have also been isolated, then CMA alloc fail finally during the first round of scanning of the whole available CMA bitmap.
This patchset introduces a retry mechanism to rescan CMA bitmap for -EBUSY error in case the target pageblock may has been temporarily isolated by others and released later. It also improves the CMA allocation performance by trying the next pageblock during reties rather than looping in the same pageblock which is in -EBUSY state.
Theoretically, this issue can be easily reproduced on ARMv7 platforms with big MAX_ORDER/pageblock e.g. 1G RAM(320M reserved CMA) and 32M pageblock ARM platform: Page block order: 13 Pages per block: 8192
The following test is based on linux-next: next-20211213.
Without the fix, it's easily fail. # insmod cma_alloc.ko pnum=16 [ 274.322369] CMA alloc test enter: thread number: 16 [ 274.329948] cpu: 0, pid: 692, index 4 pages 144 [ 274.330143] cpu: 1, pid: 694, index 2 pages 44 [ 274.330359] cpu: 2, pid: 695, index 7 pages 757 [ 274.330760] cpu: 2, pid: 696, index 4 pages 144 [ 274.330974] cpu: 2, pid: 697, index 6 pages 512 [ 274.331223] cpu: 2, pid: 698, index 6 pages 512 [ 274.331499] cpu: 2, pid: 699, index 2 pages 44 [ 274.332228] cpu: 2, pid: 700, index 0 pages 7 [ 274.337421] cpu: 0, pid: 701, index 1 pages 38 [ 274.337618] cpu: 2, pid: 702, index 0 pages 7 [ 274.344669] cpu: 1, pid: 703, index 0 pages 7 [ 274.344807] cpu: 3, pid: 704, index 6 pages 512 [ 274.348269] cpu: 2, pid: 705, index 5 pages 148 [ 274.349490] cma: cma_alloc: reserved: alloc failed, req-size: 38 pages, ret: -16 [ 274.366292] cpu: 1, pid: 706, index 4 pages 144 [ 274.366562] cpu: 0, pid: 707, index 3 pages 128 [ 274.367356] cma: cma_alloc: reserved: alloc failed, req-size: 128 pages, ret: -16 [ 274.367370] cpu: 0, pid: 707, index 3 pages 128 failed [ 274.371148] cma: cma_alloc: reserved: alloc failed, req-size: 148 pages, ret: -16 [ 274.375348] cma: cma_alloc: reserved: alloc failed, req-size: 144 pages, ret: -16 [ 274.384256] cpu: 2, pid: 708, index 0 pages 7 ....
With the fix, 32 threads allocating in parallel can pass overnight stress test.
root@imx6qpdlsolox:~# insmod cma_alloc.ko pnum=32 [ 112.976809] cma_alloc: loading out-of-tree module taints kernel. [ 112.984128] CMA alloc test enter: thread number: 32 [ 112.989748] cpu: 2, pid: 707, index 6 pages 512 [ 112.994342] cpu: 1, pid: 708, index 6 pages 512 [ 112.995162] cpu: 0, pid: 709, index 3 pages 128 [ 112.995867] cpu: 2, pid: 710, index 0 pages 7 [ 112.995910] cpu: 3, pid: 711, index 2 pages 44 [ 112.996005] cpu: 3, pid: 712, index 7 pages 757 [ 112.996098] cpu: 3, pid: 713, index 7 pages 757 ... [41877.368163] cpu: 1, pid: 737, index 2 pages 44 [41877.369388] cpu: 1, pid: 736, index 3 pages 128 [41878.486516] cpu: 0, pid: 737, index 2 pages 44 [41878.486515] cpu: 2, pid: 739, index 4 pages 144 [41878.486622] cpu: 1, pid: 736, index 3 pages 128 [41878.486948] cpu: 2, pid: 735, index 7 pages 757 [41878.487279] cpu: 2, pid: 738, index 4 pages 144 [41879.526603] cpu: 1, pid: 739, index 3 pages 128 [41879.606491] cpu: 2, pid: 737, index 3 pages 128 [41879.606550] cpu: 0, pid: 736, index 0 pages 7 [41879.612271] cpu: 2, pid: 738, index 4 pages 144 ...
Dong Aisheng (2): mm: cma: fix allocation may fail sometimes mm: cma: try next pageblock during retry
mm/cma.c | 15 ++++++++++++++- 1 file changed, 14 insertions(+), 1 deletion(-)
We met dma_alloc_coherent() fail sometimes when doing 8 VPU decoder test in parallel on a MX6Q SDB board.
Error log: cma: cma_alloc: linux,cma: alloc failed, req-size: 148 pages, ret: -16 cma: number of available pages: 3@125+20@172+12@236+4@380+32@736+17@2287+23@2473+20@36076+99@40477+108@40852+44@41108+20@41196+108@41364+108@41620+ 108@42900+108@43156+483@44061+1763@45341+1440@47712+20@49324+20@49388+5076@49452+2304@55040+35@58141+20@58220+20@58284+ 7188@58348+84@66220+7276@66452+227@74525+6371@75549=> 33161 free of 81920 total pages
When issue happened, we saw there were still 33161 pages (129M) free CMA memory and a lot available free slots for 148 pages in CMA bitmap that we want to allocate.
If dumping memory info, we found that there was also ~342M normal memory, but only 1352K CMA memory left in buddy system while a lot of pageblocks were isolated.
Memory info log: Normal free:351096kB min:30000kB low:37500kB high:45000kB reserved_highatomic:0KB active_anon:98060kB inactive_anon:98948kB active_file:60864kB inactive_file:31776kB unevictable:0kB writepending:0kB present:1048576kB managed:1018328kB mlocked:0kB bounce:0kB free_pcp:220kB local_pcp:192kB free_cma:1352kB lowmem_reserve[]: 0 0 0 Normal: 78*4kB (UECI) 1772*8kB (UMECI) 1335*16kB (UMECI) 360*32kB (UMECI) 65*64kB (UMCI) 36*128kB (UMECI) 16*256kB (UMCI) 6*512kB (EI) 8*1024kB (UEI) 4*2048kB (MI) 8*4096kB (EI) 8*8192kB (UI) 3*16384kB (EI) 8*32768kB (M) = 489288kB
The root cause of this issue is that since commit a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock"), CMA supports concurrent memory allocation. It's possible that the pageblock process A try to alloc has already been isolated by the allocation of process B during memory migration.
When there're multi process allocating CMA memory in parallel, it's likely that other the remain pageblocks may have also been isolated, then CMA alloc fail finally during the first round of scanning of the whole available CMA bitmap.
This patch introduces a retry mechanism to rescan CMA bitmap for -EBUSY error in case the target pageblock may has been temporarily isolated by others and released later.
Theoretically, this issue can be easily reproduced on ARMv7 platforms with big MAX_ORDER/pageblock e.g. 1G RAM(320M reserved CMA) and 32M pageblock ARM platform: Page block order: 13 Pages per block: 8192
Cc: Andrew Morton akpm@linux-foundation.org Cc: Marek Szyprowski m.szyprowski@samsung.com Cc: Lecopzer Chen lecopzer.chen@mediatek.com Cc: David Hildenbrand david@redhat.com Cc: Vlastimil Babka vbabka@suse.cz CC: stable@vger.kernel.org # 5.11+ Fixes: a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock") Signed-off-by: Dong Aisheng aisheng.dong@nxp.com --- mm/cma.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/mm/cma.c b/mm/cma.c index bc9ca8f3c487..1c13a729d274 100644 --- a/mm/cma.c +++ b/mm/cma.c @@ -433,6 +433,7 @@ struct page *cma_alloc(struct cma *cma, unsigned long count, unsigned long i; struct page *page = NULL; int ret = -ENOMEM; + int loop = 0;
if (!cma || !cma->count || !cma->bitmap) goto out; @@ -460,6 +461,16 @@ struct page *cma_alloc(struct cma *cma, unsigned long count, offset); if (bitmap_no >= bitmap_maxno) { spin_unlock_irq(&cma->lock); + pr_debug("%s(): alloc fail, retry loop %d\n", __func__, loop++); + /* + * rescan as others may finish the memory migration + * and quit if no available CMA memory found finally + */ + if (start) { + schedule(); + start = 0; + continue; + } break; } bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
On 15.12.21 09:02, Dong Aisheng wrote:
We met dma_alloc_coherent() fail sometimes when doing 8 VPU decoder test in parallel on a MX6Q SDB board.
Error log: cma: cma_alloc: linux,cma: alloc failed, req-size: 148 pages, ret: -16 cma: number of available pages: 3@125+20@172+12@236+4@380+32@736+17@2287+23@2473+20@36076+99@40477+108@40852+44@41108+20@41196+108@41364+108@41620+ 108@42900+108@43156+483@44061+1763@45341+1440@47712+20@49324+20@49388+5076@49452+2304@55040+35@58141+20@58220+20@58284+ 7188@58348+84@66220+7276@66452+227@74525+6371@75549=> 33161 free of 81920 total pages
When issue happened, we saw there were still 33161 pages (129M) free CMA memory and a lot available free slots for 148 pages in CMA bitmap that we want to allocate.
If dumping memory info, we found that there was also ~342M normal memory, but only 1352K CMA memory left in buddy system while a lot of pageblocks were isolated.
Memory info log: Normal free:351096kB min:30000kB low:37500kB high:45000kB reserved_highatomic:0KB active_anon:98060kB inactive_anon:98948kB active_file:60864kB inactive_file:31776kB unevictable:0kB writepending:0kB present:1048576kB managed:1018328kB mlocked:0kB bounce:0kB free_pcp:220kB local_pcp:192kB free_cma:1352kB lowmem_reserve[]: 0 0 0 Normal: 78*4kB (UECI) 1772*8kB (UMECI) 1335*16kB (UMECI) 360*32kB (UMECI) 65*64kB (UMCI) 36*128kB (UMECI) 16*256kB (UMCI) 6*512kB (EI) 8*1024kB (UEI) 4*2048kB (MI) 8*4096kB (EI) 8*8192kB (UI) 3*16384kB (EI) 8*32768kB (M) = 489288kB
The root cause of this issue is that since commit a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock"), CMA supports concurrent memory allocation. It's possible that the pageblock process A try to alloc has already been isolated by the allocation of process B during memory migration.
When there're multi process allocating CMA memory in parallel, it's likely that other the remain pageblocks may have also been isolated, then CMA alloc fail finally during the first round of scanning of the whole available CMA bitmap.
I already raised in different context that we should most probably convert that -EBUSY to -EAGAIN -- to differentiate an actual migration problem from a simple "concurrent allocations that target the same MAX_ORDER -1 range".
From: David Hildenbrand david@redhat.com Sent: Wednesday, December 15, 2021 8:31 PM
On 15.12.21 09:02, Dong Aisheng wrote:
We met dma_alloc_coherent() fail sometimes when doing 8 VPU decoder test in parallel on a MX6Q SDB board.
Error log: cma: cma_alloc: linux,cma: alloc failed, req-size: 148 pages, ret: -16 cma: number of available pages:
3@125+20@172+12@236+4@380+32@736+17@2287+23@2473+20@3607 6+99@40477+108
@40852+44@41108+20@41196+108@41364+108@41620+
108@42900+108@43156+483@44061+1763@45341+1440@47712+20@49 324+20@49388+
5076@49452+2304@55040+35@58141+20@58220+20@58284+ 7188@58348+84@66220+7276@66452+227@74525+6371@75549=>
33161 free of
81920 total pages
When issue happened, we saw there were still 33161 pages (129M) free CMA memory and a lot available free slots for 148 pages in CMA bitmap that we want to allocate.
If dumping memory info, we found that there was also ~342M normal memory, but only 1352K CMA memory left in buddy system while a lot of pageblocks were isolated.
Memory info log: Normal free:351096kB min:30000kB low:37500kB high:45000kB
reserved_highatomic:0KB
active_anon:98060kB inactive_anon:98948kB active_file:60864kBinactive_file:31776kB
unevictable:0kB writepending:0kB present:1048576kBmanaged:1018328kB mlocked:0kB
bounce:0kB free_pcp:220kB local_pcp:192kB free_cma:1352kBlowmem_reserve[]: 0 0 0 Normal: 78*4kB (UECI) 1772*8kB (UMECI) 1335*16kB (UMECI) 360*32kB
(UMECI) 65*64kB (UMCI)
36*128kB (UMECI) 16*256kB (UMCI) 6*512kB (EI) 8*1024kB (UEI)
4*2048kB (MI) 8*4096kB (EI)
8*8192kB (UI) 3*16384kB (EI) 8*32768kB (M) = 489288kB
The root cause of this issue is that since commit a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock"), CMA supports
concurrent
memory allocation. It's possible that the pageblock process A try to alloc has already been isolated by the allocation of process B during memory migration.
When there're multi process allocating CMA memory in parallel, it's likely that other the remain pageblocks may have also been isolated, then CMA alloc fail finally during the first round of scanning of the whole available CMA bitmap.
I already raised in different context that we should most probably convert that -EBUSY to -EAGAIN -- to differentiate an actual migration problem from a simple "concurrent allocations that target the same MAX_ORDER -1 range".
Thanks for the info. Is there a patch under review? BTW i wonder that probably makes no much difference for my patch since we may prefer retry the next pageblock rather than busy waiting on the same isolated pageblock. Otherwise, we may meet the same issue as the patch 2/2 wants to address.
How do you think?
Regards Aisheng
-- Thanks,
David / dhildenb
On 16.12.21 03:54, Aisheng Dong wrote:
From: David Hildenbrand david@redhat.com Sent: Wednesday, December 15, 2021 8:31 PM
On 15.12.21 09:02, Dong Aisheng wrote:
We met dma_alloc_coherent() fail sometimes when doing 8 VPU decoder test in parallel on a MX6Q SDB board.
Error log: cma: cma_alloc: linux,cma: alloc failed, req-size: 148 pages, ret: -16 cma: number of available pages:
3@125+20@172+12@236+4@380+32@736+17@2287+23@2473+20@3607 6+99@40477+108
@40852+44@41108+20@41196+108@41364+108@41620+
108@42900+108@43156+483@44061+1763@45341+1440@47712+20@49 324+20@49388+
5076@49452+2304@55040+35@58141+20@58220+20@58284+ 7188@58348+84@66220+7276@66452+227@74525+6371@75549=>
33161 free of
81920 total pages
When issue happened, we saw there were still 33161 pages (129M) free CMA memory and a lot available free slots for 148 pages in CMA bitmap that we want to allocate.
If dumping memory info, we found that there was also ~342M normal memory, but only 1352K CMA memory left in buddy system while a lot of pageblocks were isolated.
Memory info log: Normal free:351096kB min:30000kB low:37500kB high:45000kB
reserved_highatomic:0KB
active_anon:98060kB inactive_anon:98948kB active_file:60864kBinactive_file:31776kB
unevictable:0kB writepending:0kB present:1048576kBmanaged:1018328kB mlocked:0kB
bounce:0kB free_pcp:220kB local_pcp:192kB free_cma:1352kBlowmem_reserve[]: 0 0 0 Normal: 78*4kB (UECI) 1772*8kB (UMECI) 1335*16kB (UMECI) 360*32kB
(UMECI) 65*64kB (UMCI)
36*128kB (UMECI) 16*256kB (UMCI) 6*512kB (EI) 8*1024kB (UEI)
4*2048kB (MI) 8*4096kB (EI)
8*8192kB (UI) 3*16384kB (EI) 8*32768kB (M) = 489288kB
The root cause of this issue is that since commit a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock"), CMA supports
concurrent
memory allocation. It's possible that the pageblock process A try to alloc has already been isolated by the allocation of process B during memory migration.
When there're multi process allocating CMA memory in parallel, it's likely that other the remain pageblocks may have also been isolated, then CMA alloc fail finally during the first round of scanning of the whole available CMA bitmap.
I already raised in different context that we should most probably convert that -EBUSY to -EAGAIN -- to differentiate an actual migration problem from a simple "concurrent allocations that target the same MAX_ORDER -1 range".
Thanks for the info. Is there a patch under review?
No, and I was too busy for now to send it out.
BTW i wonder that probably makes no much difference for my patch since we may prefer retry the next pageblock rather than busy waiting on the same isolated pageblock.
Makes sense. BUT as of now we isolate not only a pageblock but a MAX_ORDER -1 page (e.g., 2 pageblocks on x86-64 (!) ). So you'll have the same issue in that case.
From: David Hildenbrand david@redhat.com Sent: Thursday, December 16, 2021 6:57 PM
On 16.12.21 03:54, Aisheng Dong wrote:
From: David Hildenbrand david@redhat.com Sent: Wednesday, December 15, 2021 8:31 PM
On 15.12.21 09:02, Dong Aisheng wrote:
We met dma_alloc_coherent() fail sometimes when doing 8 VPU decoder test in parallel on a MX6Q SDB board.
Error log: cma: cma_alloc: linux,cma: alloc failed, req-size: 148 pages, ret: -16 cma: number of available pages:
3@125+20@172+12@236+4@380+32@736+17@2287+23@2473+20@3607
6+99@40477+108
@40852+44@41108+20@41196+108@41364+108@41620+
108@42900+108@43156+483@44061+1763@45341+1440@47712+20@49
324+20@49388+
5076@49452+2304@55040+35@58141+20@58220+20@58284+ 7188@58348+84@66220+7276@66452+227@74525+6371@75549=>
33161 free of
81920 total pages
When issue happened, we saw there were still 33161 pages (129M) free CMA memory and a lot available free slots for 148 pages in CMA bitmap that we want to allocate.
If dumping memory info, we found that there was also ~342M normal memory, but only 1352K CMA memory left in buddy system while a lot of pageblocks were isolated.
Memory info log: Normal free:351096kB min:30000kB low:37500kB high:45000kB
reserved_highatomic:0KB
active_anon:98060kB inactive_anon:98948kB active_file:60864kBinactive_file:31776kB
unevictable:0kB writepending:0kB present:1048576kBmanaged:1018328kB mlocked:0kB
bounce:0kB free_pcp:220kB local_pcp:192kB free_cma:1352kBlowmem_reserve[]: 0 0 0 Normal: 78*4kB (UECI) 1772*8kB (UMECI) 1335*16kB (UMECI) 360*32kB
(UMECI) 65*64kB (UMCI)
36*128kB (UMECI) 16*256kB (UMCI) 6*512kB (EI) 8*1024kB (UEI)
4*2048kB (MI) 8*4096kB (EI)
8*8192kB (UI) 3*16384kB (EI) 8*32768kB (M) = 489288kB
The root cause of this issue is that since commit a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock"), CMA supports
concurrent
memory allocation. It's possible that the pageblock process A try to alloc has already been isolated by the allocation of process B during memory migration.
When there're multi process allocating CMA memory in parallel, it's likely that other the remain pageblocks may have also been isolated, then CMA alloc fail finally during the first round of scanning of the whole available CMA bitmap.
I already raised in different context that we should most probably convert that -EBUSY to -EAGAIN -- to differentiate an actual migration problem from a simple "concurrent allocations that target the
same MAX_ORDER -1 range".
Thanks for the info. Is there a patch under review?
No, and I was too busy for now to send it out.
BTW i wonder that probably makes no much difference for my patch since we may prefer retry the next pageblock rather than busy waiting on the
same isolated pageblock.
Makes sense. BUT as of now we isolate not only a pageblock but a MAX_ORDER -1 page (e.g., 2 pageblocks on x86-64 (!) ). So you'll have the same issue in that case.
Yes, should I change to try next MAX_ORDER_NR_PAGES or keep as it is and let the core to improve it later?
I saw there's a patchset under review which is going to remove the MAX_ORDER - 1 alignment requirement for CMA. https://patchwork.kernel.org/project/linux-mm/cover/20211209230414.2766515-1...
Once it's merged, I guess we can back to align with pageblock rather than MAX_ORDER-1.
Regards Aisheng
-- Thanks,
David / dhildenb
On 17.12.21 04:44, Aisheng Dong wrote:
From: David Hildenbrand david@redhat.com Sent: Thursday, December 16, 2021 6:57 PM
On 16.12.21 03:54, Aisheng Dong wrote:
From: David Hildenbrand david@redhat.com Sent: Wednesday, December 15, 2021 8:31 PM
On 15.12.21 09:02, Dong Aisheng wrote:
We met dma_alloc_coherent() fail sometimes when doing 8 VPU decoder test in parallel on a MX6Q SDB board.
Error log: cma: cma_alloc: linux,cma: alloc failed, req-size: 148 pages, ret: -16 cma: number of available pages:
3@125+20@172+12@236+4@380+32@736+17@2287+23@2473+20@3607
6+99@40477+108
@40852+44@41108+20@41196+108@41364+108@41620+
108@42900+108@43156+483@44061+1763@45341+1440@47712+20@49
324+20@49388+
5076@49452+2304@55040+35@58141+20@58220+20@58284+ 7188@58348+84@66220+7276@66452+227@74525+6371@75549=>
33161 free of
81920 total pages
When issue happened, we saw there were still 33161 pages (129M) free CMA memory and a lot available free slots for 148 pages in CMA bitmap that we want to allocate.
If dumping memory info, we found that there was also ~342M normal memory, but only 1352K CMA memory left in buddy system while a lot of pageblocks were isolated.
Memory info log: Normal free:351096kB min:30000kB low:37500kB high:45000kB
reserved_highatomic:0KB
active_anon:98060kB inactive_anon:98948kB active_file:60864kBinactive_file:31776kB
unevictable:0kB writepending:0kB present:1048576kBmanaged:1018328kB mlocked:0kB
bounce:0kB free_pcp:220kB local_pcp:192kB free_cma:1352kBlowmem_reserve[]: 0 0 0 Normal: 78*4kB (UECI) 1772*8kB (UMECI) 1335*16kB (UMECI) 360*32kB
(UMECI) 65*64kB (UMCI)
36*128kB (UMECI) 16*256kB (UMCI) 6*512kB (EI) 8*1024kB (UEI)
4*2048kB (MI) 8*4096kB (EI)
8*8192kB (UI) 3*16384kB (EI) 8*32768kB (M) = 489288kB
The root cause of this issue is that since commit a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock"), CMA supports
concurrent
memory allocation. It's possible that the pageblock process A try to alloc has already been isolated by the allocation of process B during memory migration.
When there're multi process allocating CMA memory in parallel, it's likely that other the remain pageblocks may have also been isolated, then CMA alloc fail finally during the first round of scanning of the whole available CMA bitmap.
I already raised in different context that we should most probably convert that -EBUSY to -EAGAIN -- to differentiate an actual migration problem from a simple "concurrent allocations that target the
same MAX_ORDER -1 range".
Thanks for the info. Is there a patch under review?
No, and I was too busy for now to send it out.
BTW i wonder that probably makes no much difference for my patch since we may prefer retry the next pageblock rather than busy waiting on the
same isolated pageblock.
Makes sense. BUT as of now we isolate not only a pageblock but a MAX_ORDER -1 page (e.g., 2 pageblocks on x86-64 (!) ). So you'll have the same issue in that case.
Yes, should I change to try next MAX_ORDER_NR_PAGES or keep as it is and let the core to improve it later?
I saw there's a patchset under review which is going to remove the MAX_ORDER - 1 alignment requirement for CMA. https://patchwork.kernel.org/project/linux-mm/cover/20211209230414.2766515-1...
Once it's merged, I guess we can back to align with pageblock rather than MAX_ORDER-1.
While the goal is to get rid of the alignment requirement, we might still have to isolate all applicable MAX_ORDER-1 pageblocks. Depends on what we can or cannot achieve easily :)
On Fri, Dec 17, 2021 at 8:27 PM David Hildenbrand david@redhat.com wrote:
On 17.12.21 04:44, Aisheng Dong wrote:
From: David Hildenbrand david@redhat.com Sent: Thursday, December 16, 2021 6:57 PM
On 16.12.21 03:54, Aisheng Dong wrote:
From: David Hildenbrand david@redhat.com Sent: Wednesday, December 15, 2021 8:31 PM
On 15.12.21 09:02, Dong Aisheng wrote:
We met dma_alloc_coherent() fail sometimes when doing 8 VPU decoder test in parallel on a MX6Q SDB board.
Error log: cma: cma_alloc: linux,cma: alloc failed, req-size: 148 pages, ret: -16 cma: number of available pages:
3@125+20@172+12@236+4@380+32@736+17@2287+23@2473+20@3607
6+99@40477+108
@40852+44@41108+20@41196+108@41364+108@41620+
108@42900+108@43156+483@44061+1763@45341+1440@47712+20@49
324+20@49388+
5076@49452+2304@55040+35@58141+20@58220+20@58284+ 7188@58348+84@66220+7276@66452+227@74525+6371@75549=>
33161 free of
81920 total pages
When issue happened, we saw there were still 33161 pages (129M) free CMA memory and a lot available free slots for 148 pages in CMA bitmap that we want to allocate.
If dumping memory info, we found that there was also ~342M normal memory, but only 1352K CMA memory left in buddy system while a lot of pageblocks were isolated.
Memory info log: Normal free:351096kB min:30000kB low:37500kB high:45000kB
reserved_highatomic:0KB
active_anon:98060kB inactive_anon:98948kB active_file:60864kBinactive_file:31776kB
unevictable:0kB writepending:0kB present:1048576kBmanaged:1018328kB mlocked:0kB
bounce:0kB free_pcp:220kB local_pcp:192kB free_cma:1352kBlowmem_reserve[]: 0 0 0 Normal: 78*4kB (UECI) 1772*8kB (UMECI) 1335*16kB (UMECI) 360*32kB
(UMECI) 65*64kB (UMCI)
36*128kB (UMECI) 16*256kB (UMCI) 6*512kB (EI) 8*1024kB (UEI)
4*2048kB (MI) 8*4096kB (EI)
8*8192kB (UI) 3*16384kB (EI) 8*32768kB (M) = 489288kB
The root cause of this issue is that since commit a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock"), CMA supports
concurrent
memory allocation. It's possible that the pageblock process A try to alloc has already been isolated by the allocation of process B during memory migration.
When there're multi process allocating CMA memory in parallel, it's likely that other the remain pageblocks may have also been isolated, then CMA alloc fail finally during the first round of scanning of the whole available CMA bitmap.
I already raised in different context that we should most probably convert that -EBUSY to -EAGAIN -- to differentiate an actual migration problem from a simple "concurrent allocations that target the
same MAX_ORDER -1 range".
Thanks for the info. Is there a patch under review?
No, and I was too busy for now to send it out.
BTW i wonder that probably makes no much difference for my patch since we may prefer retry the next pageblock rather than busy waiting on the
same isolated pageblock.
Makes sense. BUT as of now we isolate not only a pageblock but a MAX_ORDER -1 page (e.g., 2 pageblocks on x86-64 (!) ). So you'll have the same issue in that case.
Yes, should I change to try next MAX_ORDER_NR_PAGES or keep as it is and let the core to improve it later?
I saw there's a patchset under review which is going to remove the MAX_ORDER - 1 alignment requirement for CMA. https://patchwork.kernel.org/project/linux-mm/cover/20211209230414.2766515-1...
Once it's merged, I guess we can back to align with pageblock rather than MAX_ORDER-1.
While the goal is to get rid of the alignment requirement, we might still have to isolate all applicable MAX_ORDER-1 pageblocks. Depends on what we can or cannot achieve easily :)
Ok, got it. As that's another story and does not affect us to fix the current kernel problem first that CMA alloc may fail occasionally, I'm going to change to align with MAX_ORDER_NR_PAGES for retries as you pointed out in the next version. Do you have more suggestions for this patchset?
Regards Aisheng
-- Thanks,
David / dhildenb
On an ARMv7 platform with 32M pageblock(MAX_ORDER 14), we observed a huge number of retries of CMA allocation (1k+) during booting when allocating one page for each of 3 mmc instance probe.
This is caused by CMA now supports cocurrent allocation since commit a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock"). The pageblock we tried to allocate may have already been acquired and isolated by others, then cma_alloc() will retry the next area of the same size by bitmap_no + mask + 1. However, the pageblock order could be big and pageblock_nr_pages is huge (e.g. 8192), then keep retrying in a small step become meaningless because it's likely known to fail again due to within the same pageblock.
Instread of looping in the same pageblock and wasting CPU mips, especially for big pageblock system (e.g. 16M or 32M), we try the next pageblock directly.
Doing this way can greatly mitigate the situtation.
Below is the original error log during booting: [ 2.004804] cma: cma_alloc(cma (ptrval), count 1, align 0) [ 2.010318] cma: cma_alloc(cma (ptrval), count 1, align 0) [ 2.010776] cma: cma_alloc(): memory range at (ptrval) is busy, retrying [ 2.010785] cma: cma_alloc(): memory range at (ptrval) is busy, retrying [ 2.010793] cma: cma_alloc(): memory range at (ptrval) is busy, retrying [ 2.010800] cma: cma_alloc(): memory range at (ptrval) is busy, retrying [ 2.010807] cma: cma_alloc(): memory range at (ptrval) is busy, retrying [ 2.010814] cma: cma_alloc(): memory range at (ptrval) is busy, retrying .... (+1K retries)
After fix, the 1200+ reties can be reduced to 0. Another test running 8 VPU decoder in parallel shows that 1500+ retries dropped to ~145.
IOW this patch can improve the CMA allocation speed a lot when there're enough CMA memory by reducing retries significantly.
Cc: Andrew Morton akpm@linux-foundation.org Cc: Marek Szyprowski m.szyprowski@samsung.com Cc: Lecopzer Chen lecopzer.chen@mediatek.com Cc: David Hildenbrand david@redhat.com Cc: Vlastimil Babka vbabka@suse.cz CC: stable@vger.kernel.org # 5.11+ Fixes: a4efc174b382 ("mm/cma.c: remove redundant cma_mutex lock") Signed-off-by: Dong Aisheng aisheng.dong@nxp.com --- mm/cma.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/mm/cma.c b/mm/cma.c index 1c13a729d274..108a1ceacbe7 100644 --- a/mm/cma.c +++ b/mm/cma.c @@ -500,7 +500,9 @@ struct page *cma_alloc(struct cma *cma, unsigned long count, trace_cma_alloc_busy_retry(cma->name, pfn, pfn_to_page(pfn), count, align); /* try again with a bit different memory target */ - start = bitmap_no + mask + 1; + start = ALIGN(bitmap_no + mask + 1, + pageblock_nr_pages >> cma->order_per_bit); + }
trace_cma_alloc_finish(cma->name, pfn, page, count, align);
linux-stable-mirror@lists.linaro.org
-
Aisheng Dong -
David Hildenbrand -
Dong Aisheng
-
Dong Aisheng