: system-heap use PAGE_ALLOC_COSTLY_ORDER for high order
Hi,
Quoted 37 lines
________________________________
From: Laura Abbott <[email protected]>
To: Heesub Shin <[email protected]>; Pintu Kumar <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]
Sent: Monday, 6 October 2014 7:37 PM
Subject: Re: [PATCH 1/1] [ion]: system-heap usePAGE_ALLOC_COSTLY_ORDERfor high orderQuoted 15 lines · Heesub Shin
Hello Kumar,
Quoted 9 lines · Pintu Kumar
The Android
ion_system_heapuses allocation fallback mechanism based on 8,4,0 order pages available in the system.
It changes gfp flags based on higher order allocation request. This higher order value is hard-coded as 4, instead of using the system defined higher order value.
Thus replacing this hard-coded value withPAGE_ALLOC_COSTLY_ORDERwhich is defined as 3.
This will help mapping the higher order request in system heap with the actual allocation request.Quite reasonable.
Quoted 8 lines
Reviewed-by Heesub Shin <[email protected]>BTW, Anyone knows how the allocation order (8,4 and 0) was decided? I think only Google guys might know the answer.
regards,
heesub
Quoted 6 lines
My understanding was this was completely unrelated to the costly order and was related to the page sizes corresponding to IOMMU page sizes (1MB, 64K, 4K). This won't make a difference for the uncached page pool case but for the not page pool case, I'm not sure if there would be a benefit for trying to get 32K pages with some effort vs. just going back to 4K pages.
No, it is not just related to IOMMU case. It comes into picture also for normal system-heap allocation (without iommu cases).
Also, it is applicable for both uncached and page_pool cases.
Please also check the changes under ion_system_heap_create.
Here the gfp_flags are set under the pool structure.
This value is used in ion_page_pool_alloc_pages.
In both the cases, it internally calls alloc_pages, with this gfp_flags. Now, during memory pressure scenario, when alloc_pages moves to slowpath this gfp_flags will be used to decide allocation retry.
In the current code, the higher-order flag is set only when order is greater than 4. But, in MM, the order 4 is also considered as higher-order request. This higher-order is decided based on PAGE_ALLOC_COSTLY_ORDER (3) value. Hence, I think this value should be in sync with the MM code.
Quoted 2 lines
Do you have any data/metrics that show a benefit from this patch?
I think it is not related to any data or metrics.
It is about replacing the hard-coded higher-order check to be in sync with the MM code.
ÿôèº{.nÇ+‰·Ÿ®‰†+%ŠËÿ±éݶ¥Šwÿº{.nÇ+‰·¥Š{±þG«�éÿŠ{ayºÊ‡Ú™ë,j¢f£¢·hš�ï�êÿ‘êçz_è®(階ŠÝ¢j"�ú¶m§ÿÿ¾«þG«�éÿ¢¸?™¨èÚ&£ø§~�á¶iO•æ¬z·švØ^¶m§ÿÿÃÿ¶ìÿ¢¸?–I¥
Quoted 60 lines · PINTU KUMAR
Hi,
Quoted 37 lines
________________________________
From: Laura Abbott <[email protected]>
To: Heesub Shin <[email protected]>; Pintu Kumar <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]
Sent: Monday, 6 October 2014 7:37 PM
Subject: Re: [PATCH 1/1] [ion]: system-heap usePAGE_ALLOC_COSTLY_ORDERfor high orderQuoted 15 lines · Heesub Shin
Hello Kumar,
Quoted 9 lines · Pintu Kumar
The Android
ion_system_heapuses allocation fallback mechanism based on 8,4,0 order pages available in the system.
It changes gfp flags based on higher order allocation request. This higher order value is hard-coded as 4, instead of using the system defined higher order value.
Thus replacing this hard-coded value withPAGE_ALLOC_COSTLY_ORDERwhich is defined as 3.
This will help mapping the higher order request in system heap with the actual allocation request.Quite reasonable.
Quoted 8 lines
Reviewed-by Heesub Shin <[email protected]>BTW, Anyone knows how the allocation order (8,4 and 0) was decided? I think only Google guys might know the answer.
regards,
heesubQuoted 6 lines
My understanding was this was completely unrelated to the costly order and was related to the page sizes corresponding to IOMMU page sizes (1MB, 64K, 4K). This won't make a difference for the uncached page pool case but for the not page pool case, I'm not sure if there would be a benefit for trying to get 32K pages with some effort vs. just going back to 4K pages.
No, it is not just related to IOMMU case. It comes into picture also for normal system-heap allocation (without iommu cases).
Also, it is applicable for both uncached andpage_poolcases.
Please also check the changes underion_system_heap_create.
Here thegfp_flagsare set under the pool structure.
This value is used inion_page_pool_alloc_pages.
In both the cases, it internally callsalloc_pages, with thisgfp_flags. Now, during memory pressure scenario, whenalloc_pagesmoves to slowpath thisgfp_flagswill be used to decide allocation retry.
In the current code, the higher-order flag is set only when order is greater than 4. But, in MM, the order 4 is also considered as higher-order request. This higher-order is decided based onPAGE_ALLOC_COSTLY_ORDER(3) value. Hence, I think this value should be in sync with the MM code.Quoted 1 line
Quoted 5 lines
Quoted 1 line
Do you have any data/metrics that show a benefit from this patch?
I think it is not related to any data or metrics.
It is about replacing the hard-coded higher-order check to be in sync with the MM code.
The selection of the orders used for allocation (8, then 4, then 0) is designed to match with the sizes often found in IOMMUs, but this isn't changing the order of the allocation, it is changing the GFP flags used for the order 4 allocation. Right now we are using the low_order_gfp_flags for order 4, this patch would change it to use high_order_gfp_flags. We originally used low_order_gfp_flags here because the MM subsystem can usually satisfy these allocations, and the additional load placed on the MM subsystem to kick off kswapd to free up more order 4 chunks is generally worth it. Using order 4 pages instead of order 0 pages can significantly improve the performance of many IOMMUs by reducing TLB pressure and time spent updating page tables. Unless you have data showing that this improves something, and doesn't just cause all allocations to be order 0 when under memory pressure, I don't suggest merging this.
----- Original Message -----
Quoted 84 lines
From: Colin Cross <[email protected]>
To: [email protected]
Cc: Laura Abbott <[email protected]>; Heesub Shin <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; IQBAL SHAREEF <[email protected]>; "[email protected]" <[email protected]>; Vishnu Pratap Singh <[email protected]>; "[email protected]" <[email protected]>
Sent: Monday, 6 October 2014 11:01 PM
Subject: Re: Re: [PATCH 1/1] [ion]: system-heap usePAGE_ALLOC_COSTLY_ORDERfor high orderQuoted 5 lines · PINTU KUMAR
Hi,
Quoted 3 lines
________________________________
From: Laura Abbott <[email protected]>
To: Heesub Shin <[email protected]>; Pintu Kumar<[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Quoted 1 line
Quoted 1 line
Quoted 2 lines
Quoted 2 lines
Sent: Monday, 6 October 2014 7:37 PM
Subject: Re: [PATCH 1/1] [ion]: system-heap usePAGE_ALLOC_COSTLY_ORDERfor high order
Quoted 9 lines
Quoted 9 lines · Heesub Shin
Quoted 6 lines
Hello Kumar,
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines · Pintu Kumar
Quoted 6 lines
The Android
ion_system_heapuses allocation fallback mechanism based on 8,4,0 order pages available in the system. It changes gfp flags based on higher order allocation request. This higher order value is hard-coded as 4, instead of using the system defined higher order value.
Thus replacing this hard-coded value with
PAGE_ALLOC_COSTLY_ORDERQuoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
which is defined as 3.
This will help mapping the higher order request in system heapwith
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 1 line
the actual allocation request.
Quite reasonable.
Reviewed-by Heesub Shin <[email protected]>BTW, Anyone knows how the allocation order (8,4 and 0) was
decided? I
Quoted 10 lines
Quoted 10 lines
Quoted 5 lines
think only Google guys might know the answer.
regards,
heesubMy understanding was this was completely unrelated to the costly order and was related to the page sizes corresponding to IOMMU page sizes (1MB, 64K, 4K). This won't make a difference for the uncached page pool case but for the not page pool case, I'm not sure if there
would
Quoted 13 lines
Quoted 2 lines
be a benefit for trying to get 32K pages with some effort vs. just going back to 4K pages.
No, it is not just related to IOMMU case. It comes into picture also for normal system-heap allocation (without iommu cases).
Also, it is applicable for both uncached andpage_poolcases. Please also check the changes underion_system_heap_create. Here thegfp_flagsare set under the pool structure.
This value is used inion_page_pool_alloc_pages.
In both the cases, it internally callsalloc_pages, with thisgfp_flags. Now, during memory pressure scenario, whenalloc_pagesmoves to slowpath thisgfp_flagswill be used to decide allocation retry.
In the current code, the higher-order flag is set only when order isgreater than 4.
Quoted 9 lines
But, in MM, the order 4 is also considered as higher-order request. This higher-order is decided based on
PAGE_ALLOC_COSTLY_ORDER(3) value. Hence, I think this value should be in sync with the MM code.Quoted 2 lines
Do you have any data/metrics that show a benefit from this patch?
I think it is not related to any data or metrics.
It is about replacing the hard-coded higher-order check to be in sync with the MM code.
Quoted 15 lines
The selection of the orders used for allocation (8, then 4, then 0) is designed to match with the sizes often found in IOMMUs, but this isn't changing the order of the allocation, it is changing the GFP flags used for the order 4 allocation. Right now we are using the
low_order_gfp_flagsfor order 4, this patch would change it to usehigh_order_gfp_flags. We originally usedlow_order_gfp_flagshere because the MM subsystem can usually satisfy these allocations, and the additional load placed on the MM subsystem to kick off kswapd to free up more order 4 chunks is generally worth it. Using order 4 pages instead of order 0 pages can significantly improve the performance of many IOMMUs by reducing TLB pressure and time spent updating page tables. Unless you have data showing that this improves something, and doesn't just cause all allocations to be order 0 when under memory pressure, I don't suggest merging this.
Ok agree. It is worth retrying the allocation with order-4 pages.
But, since 4 is considered higher order for MM and is greater than PAGE_ALLOC_COSTLY_ORDER. I guess the retrying will not happen, because of the following check in page_alloc:
if (order > PAGE_ALLOC_COSTLY_ORDER)
goto nopage;
Quoted 101 lines · PINTU KUMAR
----- Original Message -----
Quoted 84 lines
From: Colin Cross <[email protected]>
To: [email protected]
Cc: Laura Abbott <[email protected]>; Heesub Shin <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; "[email protected]" <[email protected]>; IQBAL SHAREEF <[email protected]>; "[email protected]" <[email protected]>; Vishnu Pratap Singh <[email protected]>; "[email protected]" <[email protected]>
Sent: Monday, 6 October 2014 11:01 PM
Subject: Re: Re: [PATCH 1/1] [ion]: system-heap usePAGE_ALLOC_COSTLY_ORDERfor high orderQuoted 5 lines · PINTU KUMAR
Hi,
Quoted 3 lines
________________________________
From: Laura Abbott <[email protected]>
To: Heesub Shin <[email protected]>; Pintu Kumar<[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Quoted 1 line
Quoted 1 line
Quoted 2 lines
Quoted 2 lines
Sent: Monday, 6 October 2014 7:37 PM
Subject: Re: [PATCH 1/1] [ion]: system-heap usePAGE_ALLOC_COSTLY_ORDERfor high order
Quoted 9 lines
Quoted 9 lines · Heesub Shin
Quoted 6 lines
Hello Kumar,
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines · Pintu Kumar
Quoted 6 lines
The Android
ion_system_heapuses allocation fallback mechanism based on 8,4,0 order pages available in the system. It changes gfp flags based on higher order allocation request. This higher order value is hard-coded as 4, instead of using the system defined higher order value.
Thus replacing this hard-coded value with
PAGE_ALLOC_COSTLY_ORDERQuoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
which is defined as 3.
This will help mapping the higher order request in system heapwith
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 1 line
the actual allocation request.
Quite reasonable.
Reviewed-by Heesub Shin <[email protected]>BTW, Anyone knows how the allocation order (8,4 and 0) was
decided? I
Quoted 10 lines
Quoted 10 lines
Quoted 5 lines
think only Google guys might know the answer.
regards,
heesubMy understanding was this was completely unrelated to the costly order and was related to the page sizes corresponding to IOMMU page sizes (1MB, 64K, 4K). This won't make a difference for the uncached page pool case but for the not page pool case, I'm not sure if there
would
Quoted 13 lines
Quoted 2 lines
be a benefit for trying to get 32K pages with some effort vs. just going back to 4K pages.
No, it is not just related to IOMMU case. It comes into picture also for normal system-heap allocation (without iommu cases).
Also, it is applicable for both uncached andpage_poolcases. Please also check the changes underion_system_heap_create. Here thegfp_flagsare set under the pool structure.
This value is used inion_page_pool_alloc_pages.
In both the cases, it internally callsalloc_pages, with thisgfp_flags. Now, during memory pressure scenario, whenalloc_pagesmoves to slowpath thisgfp_flagswill be used to decide allocation retry.
In the current code, the higher-order flag is set only when order isgreater than 4.
Quoted 9 lines
But, in MM, the order 4 is also considered as higher-order request. This higher-order is decided based on
PAGE_ALLOC_COSTLY_ORDER(3) value. Hence, I think this value should be in sync with the MM code.Quoted 2 lines
Do you have any data/metrics that show a benefit from this patch?
I think it is not related to any data or metrics.
It is about replacing the hard-coded higher-order check to be in sync with the MM code.Quoted 15 lines
The selection of the orders used for allocation (8, then 4, then 0) is designed to match with the sizes often found in IOMMUs, but this isn't changing the order of the allocation, it is changing the GFP flags used for the order 4 allocation. Right now we are using the
low_order_gfp_flagsfor order 4, this patch would change it to usehigh_order_gfp_flags. We originally usedlow_order_gfp_flagshere because the MM subsystem can usually satisfy these allocations, and the additional load placed on the MM subsystem to kick off kswapd to free up more order 4 chunks is generally worth it. Using order 4 pages instead of order 0 pages can significantly improve the performance of many IOMMUs by reducing TLB pressure and time spent updating page tables. Unless you have data showing that this improves something, and doesn't just cause all allocations to be order 0 when under memory pressure, I don't suggest merging this.
Quoted 5 lines
Ok agree. It is worth retrying the allocation with order-4 pages.
But, since 4 is considered higher order for MM and is greater thanPAGE_ALLOC_COSTLY_ORDER. I guess the retrying will not happen, because of the following check inpage_alloc:
if (order >PAGE_ALLOC_COSTLY_ORDER)goto nopage;
It's not the retry that I'd be particularly worried about, it's kicking off kswapd, which will run in the background to replenish the pool of free order 4 pages. However, when c654345924f7cce87bb221b89db91cba890421ba (mm: remove __GFP_NO_KSWAPD) was merged (and then reverted, and then merged again), the __GFP_NO_KSWAPD flag was removed from high_order_gfp_flags, and when it was finally reverted again we did not re-add the flag, so we lost that behavior long ago.