Re: [PATCH v2 00/21] Refine memblock API
On Wed, Oct 2, 2019 at 2:36 AM Mike Rapoport wrote: > > Hi Adam, > > On Tue, Oct 01, 2019 at 07:14:13PM -0500, Adam Ford wrote: > > On Sun, Sep 29, 2019 at 8:33 AM Adam Ford wrote: > > > > > > I am attaching two logs. I now the mailing lists will be unhappy, but > > > don't want to try and spam a bunch of log through the mailing liast. > > > The two logs show the differences between the working and non-working > > > imx6q 3D accelerator when trying to run a simple glmark2-es2-drm demo. > > > > > > The only change between them is the 2 line code change you suggested. > > > > > > In both cases, I have cma=128M set in my bootargs. Historically this > > > has been sufficient, but cma=256M has not made a difference. > > > > > > > Mike any suggestions on how to move forward? > > I was hoping to get the fixes tested and pushed before 5.4 is released > > if at all possible > > I have a fix (below) that kinda restores the original behaviour, but I > still would like to double check to make sure it's not a band aid and I > haven't missed the actual root cause. > > Can you please send me your device tree definition and the output of > > cat /sys/kernel/debug/memblock/memory > > and > > cat /sys/kernel/debug/memblock/reserved > > Thanks! > Before the patch: # cat /sys/kernel/debug/memblock/memory 0: 0x1000..0x8fff # cat /sys/kernel/debug/memblock/reserved 0: 0x10004000..0x10007fff 1: 0x1010..0x11ab141f 2: 0x1fff1000..0x1fffcfff 3: 0x2ee4..0x2ef53fff 4: 0x2ef56940..0x2ef56c43 5: 0x2ef56c48..0x2fffefff 6: 0x20c0..0x24d8 7: 0x2500..0x255f 8: 0x2580..0x2703 9: 0x2740..0x2918 10: 0x2940..0x29cf 11: 0x2a00..0x2a0f 12: 0x2a40..0x2a43 13: 0x2a80..0x2ad5 14: 0x2b00..0x2b55 15: 0x2b80..0x2bd5 16: 0x2c00..0x2c4e 17: 0x2c50..0x2c6a 18: 0x2c6c..0x2ce6 19: 0x2ce8..0x2d02 20: 0x2d04..0x2d1e 21: 0x2d20..0x2d3a 22: 0x2d3c..0x2d56 23: 0x2d58..0x2e30 24: 0x2e34..0x2e4c 25: 0x2e50..0x2e68 26: 0x2e6c..0x2e84 27: 0x2e88..0x2ea0 28: 0x2ea4..0x2ebc 29: 0x2ec0..0x2edf 30: 0x2ee4..0x2efc 31: 0x2f00..0x2f13 32: 0x2f28..0x2f4b 33: 0x2f50..0x2f84 34: 0x2f88..0x3fff After the patch: # cat /sys/kernel/debug/memblock/memory 0: 0x1000..0x8fff # cat /sys/kernel/debug/memblock/reserved 0: 0x10004000..0x10007fff 1: 0x1010..0x11ab141f 2: 0x1fff1000..0x1fffcfff 3: 0x3eec..0x3efd3fff 4: 0x3efd6940..0x3efd6c43 5: 0x3efd6c48..0x3fffbfff 6: 0x3fffc0c0..0x3fffc4d8 7: 0x3fffc500..0x3fffc55f 8: 0x3fffc580..0x3fffc703 9: 0x3fffc740..0x3fffc918 10: 0x3fffc940..0x3fffc9cf 11: 0x3fffca00..0x3fffca0f 12: 0x3fffca40..0x3fffca43 13: 0x3fffca80..0x3fffca83 14: 0x3fffcac0..0x3fffcb15 15: 0x3fffcb40..0x3fffcb95 16: 0x3fffcbc0..0x3fffcc15 17: 0x3fffcc28..0x3fffcc72 18: 0x3fffcc74..0x3fffcc8e 19: 0x3fffcc90..0x3fffcd0a 20: 0x3fffcd0c..0x3fffcd26 21: 0x3fffcd28..0x3fffcd42 22: 0x3fffcd44..0x3fffcd5e 23: 0x3fffcd60..0x3fffcd7a 24: 0x3fffcd7c..0x3fffce54 25: 0x3fffce58..0x3fffce70 26: 0x3fffce74..0x3fffce8c 27: 0x3fffce90..0x3fffcea8 28: 0x3fffceac..0x3fffcec4 29: 0x3fffcec8..0x3fffcee0 30: 0x3fffcee4..0x3fffcefc 31: 0x3fffcf00..0x3fffcf1f 32: 0x3fffcf28..0x3fffcf53 33: 0x3fffcf68..0x3fffcf8b 34: 0x3fffcf90..0x3fffcfac 35: 0x3fffcfb0..0x3fff 36: 0x8000..0x8fff > From 06529f861772b7dea2912fc2245debe4690139b8 Mon Sep 17 00:00:00 2001 > From: Mike Rapoport > Date: Wed, 2 Oct 2019 10:14:17 +0300 > Subject: [PATCH] mm: memblock: do not enforce current limit for memblock_phys* > family > > Until commit 92d12f9544b7 ("memblock: refactor internal allocation > functions") the maximal address for memblock allocations was forced to > memblock.current_limit only for the allocation functions returning virtual > address. The changes introduced by that commit moved the limit enforcement > into the allocation core and as a result the allocation functions returning > physical address also started to limit allocations to > memblock.current_limit. > > This caused breakage of etnaviv GPU driver: > > [3.682347] etnaviv etnaviv: bound 13.gpu (ops gpu_ops) > [3.688669] etnaviv etnaviv: bound 134000.gpu (ops gpu_ops) > [3.695099] etnaviv etnaviv: bound 2204000.gpu (ops gpu_ops) > [3.700800] etnaviv-gpu 13.gpu: model: GC2000, revision: 5108 > [3.723013] etnaviv-gpu 13.gpu: command buffer outside valid > memory window > [3.731308] etnaviv-gpu 134000.gpu: model: GC320, revision: 5007 > [3.752437] etnaviv-gpu 134000.gpu: command buffer outside valid > memory window > [3.760583] etnaviv-gpu 2204000.gpu: model: GC355, revision: 1215 > [3.766766] etnaviv-gpu 2204000.gpu: Ignoring GPU with VG and FE2.0 > > Resto
Re: [PATCH v2 00/21] Refine memblock API
Hi Adam, On Tue, Oct 01, 2019 at 07:14:13PM -0500, Adam Ford wrote: > On Sun, Sep 29, 2019 at 8:33 AM Adam Ford wrote: > > > > I am attaching two logs. I now the mailing lists will be unhappy, but > > don't want to try and spam a bunch of log through the mailing liast. > > The two logs show the differences between the working and non-working > > imx6q 3D accelerator when trying to run a simple glmark2-es2-drm demo. > > > > The only change between them is the 2 line code change you suggested. > > > > In both cases, I have cma=128M set in my bootargs. Historically this > > has been sufficient, but cma=256M has not made a difference. > > > > Mike any suggestions on how to move forward? > I was hoping to get the fixes tested and pushed before 5.4 is released > if at all possible I have a fix (below) that kinda restores the original behaviour, but I still would like to double check to make sure it's not a band aid and I haven't missed the actual root cause. Can you please send me your device tree definition and the output of cat /sys/kernel/debug/memblock/memory and cat /sys/kernel/debug/memblock/reserved Thanks! >From 06529f861772b7dea2912fc2245debe4690139b8 Mon Sep 17 00:00:00 2001 From: Mike Rapoport Date: Wed, 2 Oct 2019 10:14:17 +0300 Subject: [PATCH] mm: memblock: do not enforce current limit for memblock_phys* family Until commit 92d12f9544b7 ("memblock: refactor internal allocation functions") the maximal address for memblock allocations was forced to memblock.current_limit only for the allocation functions returning virtual address. The changes introduced by that commit moved the limit enforcement into the allocation core and as a result the allocation functions returning physical address also started to limit allocations to memblock.current_limit. This caused breakage of etnaviv GPU driver: [3.682347] etnaviv etnaviv: bound 13.gpu (ops gpu_ops) [3.688669] etnaviv etnaviv: bound 134000.gpu (ops gpu_ops) [3.695099] etnaviv etnaviv: bound 2204000.gpu (ops gpu_ops) [3.700800] etnaviv-gpu 13.gpu: model: GC2000, revision: 5108 [3.723013] etnaviv-gpu 13.gpu: command buffer outside valid memory window [3.731308] etnaviv-gpu 134000.gpu: model: GC320, revision: 5007 [3.752437] etnaviv-gpu 134000.gpu: command buffer outside valid memory window [3.760583] etnaviv-gpu 2204000.gpu: model: GC355, revision: 1215 [3.766766] etnaviv-gpu 2204000.gpu: Ignoring GPU with VG and FE2.0 Restore the behaviour of memblock_phys* family so that these functions will not enforce memblock.current_limit. Fixes: 92d12f9544b7 ("memblock: refactor internal allocation functions") Reported-by: Adam Ford Signed-off-by: Mike Rapoport --- mm/memblock.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/mm/memblock.c b/mm/memblock.c index 7d4f61a..c4b16ca 100644 --- a/mm/memblock.c +++ b/mm/memblock.c @@ -1356,9 +1356,6 @@ static phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size, align = SMP_CACHE_BYTES; } - if (end > memblock.current_limit) - end = memblock.current_limit; - again: found = memblock_find_in_range_node(size, align, start, end, nid, flags); @@ -1469,6 +1466,9 @@ static void * __init memblock_alloc_internal( if (WARN_ON_ONCE(slab_is_available())) return kzalloc_node(size, GFP_NOWAIT, nid); + if (max_addr > memblock.current_limit) + max_addr = memblock.current_limit; + alloc = memblock_alloc_range_nid(size, align, min_addr, max_addr, nid); /* retry allocation without lower limit */ -- 2.7.4 > > adam > > > > On Sat, Sep 28, 2019 at 2:33 AM Mike Rapoport wrote: > > > > > > On Thu, Sep 26, 2019 at 02:35:53PM -0500, Adam Ford wrote: > > > > On Thu, Sep 26, 2019 at 11:04 AM Mike Rapoport > > > > wrote: > > > > > > > > > > Hi, > > > > > > > > > > On Thu, Sep 26, 2019 at 08:09:52AM -0500, Adam Ford wrote: > > > > > > On Wed, Sep 25, 2019 at 10:17 AM Fabio Estevam > > > > > > wrote: > > > > > > > > > > > > > > On Wed, Sep 25, 2019 at 9:17 AM Adam Ford > > > > > > > wrote: > > > > > > > > > > > > > > > I tried cma=256M and noticed the cma dump at the beginning > > > > > > > > didn't > > > > > > > > change. Do we need to setup a reserved-memory node like > > > > > > > > imx6ul-ccimx6ulsom.dtsi did? > > > > > > > > > > > > > > I don't think so. > > > > > > > > > > > > > > Were you able to identify what was the exact commit that caused > > > > > > > such regression? > > > > > > > > > > > > I was able to narrow it down the 92d12f9544b7 ("memblock: refactor > > > > > > internal allocation functions") that caused the regression with > > > > > > Etnaviv. > > > > > > > > > > > > > > > Can you please test with this change: > > > > > > > > > > > > > That appears to have fixed my issue. I am not sure what the impact > > > > is, but is this a safe option? > > > > > > It'
Re: [PATCH v2 00/21] Refine memblock API
On Sun, Sep 29, 2019 at 8:33 AM Adam Ford wrote: > > I am attaching two logs. I now the mailing lists will be unhappy, but > don't want to try and spam a bunch of log through the mailing liast. > The two logs show the differences between the working and non-working > imx6q 3D accelerator when trying to run a simple glmark2-es2-drm demo. > > The only change between them is the 2 line code change you suggested. > > In both cases, I have cma=128M set in my bootargs. Historically this > has been sufficient, but cma=256M has not made a difference. > Mike any suggestions on how to move forward? I was hoping to get the fixes tested and pushed before 5.4 is released if at all possible > adam > > On Sat, Sep 28, 2019 at 2:33 AM Mike Rapoport wrote: > > > > On Thu, Sep 26, 2019 at 02:35:53PM -0500, Adam Ford wrote: > > > On Thu, Sep 26, 2019 at 11:04 AM Mike Rapoport wrote: > > > > > > > > Hi, > > > > > > > > On Thu, Sep 26, 2019 at 08:09:52AM -0500, Adam Ford wrote: > > > > > On Wed, Sep 25, 2019 at 10:17 AM Fabio Estevam > > > > > wrote: > > > > > > > > > > > > On Wed, Sep 25, 2019 at 9:17 AM Adam Ford > > > > > > wrote: > > > > > > > > > > > > > I tried cma=256M and noticed the cma dump at the beginning didn't > > > > > > > change. Do we need to setup a reserved-memory node like > > > > > > > imx6ul-ccimx6ulsom.dtsi did? > > > > > > > > > > > > I don't think so. > > > > > > > > > > > > Were you able to identify what was the exact commit that caused > > > > > > such regression? > > > > > > > > > > I was able to narrow it down the 92d12f9544b7 ("memblock: refactor > > > > > internal allocation functions") that caused the regression with > > > > > Etnaviv. > > > > > > > > > > > > Can you please test with this change: > > > > > > > > > > That appears to have fixed my issue. I am not sure what the impact > > > is, but is this a safe option? > > > > It's not really a fix, I just wanted to see how exactly 92d12f9544b7 > > ("memblock: > > refactor internal allocation functions") broke your setup. > > > > Can you share the dts you are using and the full kernel log? > > > > > adam > > > > > > > diff --git a/mm/memblock.c b/mm/memblock.c > > > > index 7d4f61a..1f5a0eb 100644 > > > > --- a/mm/memblock.c > > > > +++ b/mm/memblock.c > > > > @@ -1356,9 +1356,6 @@ static phys_addr_t __init > > > > memblock_alloc_range_nid(phys_addr_t size, > > > > align = SMP_CACHE_BYTES; > > > > } > > > > > > > > - if (end > memblock.current_limit) > > > > - end = memblock.current_limit; > > > > - > > > > again: > > > > found = memblock_find_in_range_node(size, align, start, end, > > > > nid, > > > > flags); > > > > > > > > > I also noticed that if I create a reserved memory node as was done one > > > > > imx6ul-ccimx6ulsom.dtsi the 3D seems to work again, but without it, I > > > > > was getting errors regardless of the 'cma=256M' or not. > > > > > I don't have a problem using the reserved memory, but I guess I am not > > > > > sure what the amount should be. I know for the video decoding 1080p, > > > > > I have historically used cma=128M, but with the 3D also needing some > > > > > memory allocation, is that enough or should I use 256M? > > > > > > > > > > adam > > > > > > > > -- > > > > Sincerely yours, > > > > Mike. > > > > > > > > -- > > Sincerely yours, > > Mike. > >
Re: [PATCH v2 00/21] Refine memblock API
I am attaching two logs. I now the mailing lists will be unhappy, but don't want to try and spam a bunch of log through the mailing liast. The two logs show the differences between the working and non-working imx6q 3D accelerator when trying to run a simple glmark2-es2-drm demo. The only change between them is the 2 line code change you suggested. In both cases, I have cma=128M set in my bootargs. Historically this has been sufficient, but cma=256M has not made a difference. adam On Sat, Sep 28, 2019 at 2:33 AM Mike Rapoport wrote: > > On Thu, Sep 26, 2019 at 02:35:53PM -0500, Adam Ford wrote: > > On Thu, Sep 26, 2019 at 11:04 AM Mike Rapoport wrote: > > > > > > Hi, > > > > > > On Thu, Sep 26, 2019 at 08:09:52AM -0500, Adam Ford wrote: > > > > On Wed, Sep 25, 2019 at 10:17 AM Fabio Estevam > > > > wrote: > > > > > > > > > > On Wed, Sep 25, 2019 at 9:17 AM Adam Ford wrote: > > > > > > > > > > > I tried cma=256M and noticed the cma dump at the beginning didn't > > > > > > change. Do we need to setup a reserved-memory node like > > > > > > imx6ul-ccimx6ulsom.dtsi did? > > > > > > > > > > I don't think so. > > > > > > > > > > Were you able to identify what was the exact commit that caused such > > > > > regression? > > > > > > > > I was able to narrow it down the 92d12f9544b7 ("memblock: refactor > > > > internal allocation functions") that caused the regression with > > > > Etnaviv. > > > > > > > > > Can you please test with this change: > > > > > > > That appears to have fixed my issue. I am not sure what the impact > > is, but is this a safe option? > > It's not really a fix, I just wanted to see how exactly 92d12f9544b7 > ("memblock: > refactor internal allocation functions") broke your setup. > > Can you share the dts you are using and the full kernel log? > > > adam > > > > > diff --git a/mm/memblock.c b/mm/memblock.c > > > index 7d4f61a..1f5a0eb 100644 > > > --- a/mm/memblock.c > > > +++ b/mm/memblock.c > > > @@ -1356,9 +1356,6 @@ static phys_addr_t __init > > > memblock_alloc_range_nid(phys_addr_t size, > > > align = SMP_CACHE_BYTES; > > > } > > > > > > - if (end > memblock.current_limit) > > > - end = memblock.current_limit; > > > - > > > again: > > > found = memblock_find_in_range_node(size, align, start, end, nid, > > > flags); > > > > > > > I also noticed that if I create a reserved memory node as was done one > > > > imx6ul-ccimx6ulsom.dtsi the 3D seems to work again, but without it, I > > > > was getting errors regardless of the 'cma=256M' or not. > > > > I don't have a problem using the reserved memory, but I guess I am not > > > > sure what the amount should be. I know for the video decoding 1080p, > > > > I have historically used cma=128M, but with the 3D also needing some > > > > memory allocation, is that enough or should I use 256M? > > > > > > > > adam > > > > > > -- > > > Sincerely yours, > > > Mike. > > > > > -- > Sincerely yours, > Mike. > Starting kernel ... [0.00] Booting Linux on physical CPU 0x0 [0.00] Linux version 5.3.1-dirty (aford@aford-IdeaCentre-A730) (gcc version 8.3.0 (Buildroot 2019.02.5-00192-gcd72d5bf57-dirty)) #2 SMP Sun Sep 29 08:26:09 CDT 2019 [0.00] CPU: ARMv7 Processor [412fc09a] revision 10 (ARMv7), cr=10c5387d [0.00] CPU: PIPT / VIPT nonaliasing data cache, VIPT aliasing instruction cache [0.00] OF: fdt: Machine model: Logic PD i.MX6QD SOM-M3 [0.00] printk: debug: ignoring loglevel setting. [0.00] Memory policy: Data cache writealloc [0.00] cma: Reserved 128 MiB at 0x8800 [0.00] On node 0 totalpages: 524288 [0.00] Normal zone: 1536 pages used for memmap [0.00] Normal zone: 0 pages reserved [0.00] Normal zone: 196608 pages, LIFO batch:63 [0.00] HighMem zone: 327680 pages, LIFO batch:63 [0.00] percpu: Embedded 21 pages/cpu s54632 r8192 d23192 u86016 [0.00] pcpu-alloc: s54632 r8192 d23192 u86016 alloc=21*4096 [0.00] pcpu-alloc: [0] 0 [0] 1 [0] 2 [0] 3 [0.00] Built 1 zonelists, mobility grouping on. Total pages: 522752 [0.00] Kernel command line: console=ttymxc0,115200 root=PARTUUID=60f4e103-02 rootwait rw ignore_loglevel cma=128M [0.00] Dentry cache hash table entries: 131072 (order: 7, 524288 bytes, linear) [0.00] Inode-cache hash table entries: 65536 (order: 6, 262144 bytes, linear) [0.00] mem auto-init: stack:off, heap alloc:off, heap free:off [0.00] Memory: 1922048K/2097152K available (12288K kernel code, 956K rwdata, 4252K rodata, 1024K init, 6920K bss, 44032K reserved, 131072K cma-reserved, 1179648K highmem) [0.00] SLUB: HWalign=64, Order=0-3, MinObjects=0, CPUs=4, Nodes=1 [0.00] Running RCU self tests [0.00] rcu: Hierarchical RCU implementation. [0.00] rcu: RCU event tracing is enabled. [0.00] rcu: RCU lockdep che
Re: [PATCH v2 00/21] Refine memblock API
On Thu, Sep 26, 2019 at 02:35:53PM -0500, Adam Ford wrote: > On Thu, Sep 26, 2019 at 11:04 AM Mike Rapoport wrote: > > > > Hi, > > > > On Thu, Sep 26, 2019 at 08:09:52AM -0500, Adam Ford wrote: > > > On Wed, Sep 25, 2019 at 10:17 AM Fabio Estevam wrote: > > > > > > > > On Wed, Sep 25, 2019 at 9:17 AM Adam Ford wrote: > > > > > > > > > I tried cma=256M and noticed the cma dump at the beginning didn't > > > > > change. Do we need to setup a reserved-memory node like > > > > > imx6ul-ccimx6ulsom.dtsi did? > > > > > > > > I don't think so. > > > > > > > > Were you able to identify what was the exact commit that caused such > > > > regression? > > > > > > I was able to narrow it down the 92d12f9544b7 ("memblock: refactor > > > internal allocation functions") that caused the regression with > > > Etnaviv. > > > > > > Can you please test with this change: > > > > That appears to have fixed my issue. I am not sure what the impact > is, but is this a safe option? It's not really a fix, I just wanted to see how exactly 92d12f9544b7 ("memblock: refactor internal allocation functions") broke your setup. Can you share the dts you are using and the full kernel log? > adam > > > diff --git a/mm/memblock.c b/mm/memblock.c > > index 7d4f61a..1f5a0eb 100644 > > --- a/mm/memblock.c > > +++ b/mm/memblock.c > > @@ -1356,9 +1356,6 @@ static phys_addr_t __init > > memblock_alloc_range_nid(phys_addr_t size, > > align = SMP_CACHE_BYTES; > > } > > > > - if (end > memblock.current_limit) > > - end = memblock.current_limit; > > - > > again: > > found = memblock_find_in_range_node(size, align, start, end, nid, > > flags); > > > > > I also noticed that if I create a reserved memory node as was done one > > > imx6ul-ccimx6ulsom.dtsi the 3D seems to work again, but without it, I > > > was getting errors regardless of the 'cma=256M' or not. > > > I don't have a problem using the reserved memory, but I guess I am not > > > sure what the amount should be. I know for the video decoding 1080p, > > > I have historically used cma=128M, but with the 3D also needing some > > > memory allocation, is that enough or should I use 256M? > > > > > > adam > > > > -- > > Sincerely yours, > > Mike. > > -- Sincerely yours, Mike.
Re: [PATCH v2 00/21] Refine memblock API
On Thu, Sep 26, 2019 at 11:04 AM Mike Rapoport wrote: > > Hi, > > On Thu, Sep 26, 2019 at 08:09:52AM -0500, Adam Ford wrote: > > On Wed, Sep 25, 2019 at 10:17 AM Fabio Estevam wrote: > > > > > > On Wed, Sep 25, 2019 at 9:17 AM Adam Ford wrote: > > > > > > > I tried cma=256M and noticed the cma dump at the beginning didn't > > > > change. Do we need to setup a reserved-memory node like > > > > imx6ul-ccimx6ulsom.dtsi did? > > > > > > I don't think so. > > > > > > Were you able to identify what was the exact commit that caused such > > > regression? > > > > I was able to narrow it down the 92d12f9544b7 ("memblock: refactor > > internal allocation functions") that caused the regression with > > Etnaviv. > > > Can you please test with this change: > That appears to have fixed my issue. I am not sure what the impact is, but is this a safe option? adam > diff --git a/mm/memblock.c b/mm/memblock.c > index 7d4f61a..1f5a0eb 100644 > --- a/mm/memblock.c > +++ b/mm/memblock.c > @@ -1356,9 +1356,6 @@ static phys_addr_t __init > memblock_alloc_range_nid(phys_addr_t size, > align = SMP_CACHE_BYTES; > } > > - if (end > memblock.current_limit) > - end = memblock.current_limit; > - > again: > found = memblock_find_in_range_node(size, align, start, end, nid, > flags); > > > I also noticed that if I create a reserved memory node as was done one > > imx6ul-ccimx6ulsom.dtsi the 3D seems to work again, but without it, I > > was getting errors regardless of the 'cma=256M' or not. > > I don't have a problem using the reserved memory, but I guess I am not > > sure what the amount should be. I know for the video decoding 1080p, > > I have historically used cma=128M, but with the 3D also needing some > > memory allocation, is that enough or should I use 256M? > > > > adam > > -- > Sincerely yours, > Mike. >
Re: [PATCH v2 00/21] Refine memblock API
Hi, On Thu, Sep 26, 2019 at 08:09:52AM -0500, Adam Ford wrote: > On Wed, Sep 25, 2019 at 10:17 AM Fabio Estevam wrote: > > > > On Wed, Sep 25, 2019 at 9:17 AM Adam Ford wrote: > > > > > I tried cma=256M and noticed the cma dump at the beginning didn't > > > change. Do we need to setup a reserved-memory node like > > > imx6ul-ccimx6ulsom.dtsi did? > > > > I don't think so. > > > > Were you able to identify what was the exact commit that caused such > > regression? > > I was able to narrow it down the 92d12f9544b7 ("memblock: refactor > internal allocation functions") that caused the regression with > Etnaviv. Can you please test with this change: diff --git a/mm/memblock.c b/mm/memblock.c index 7d4f61a..1f5a0eb 100644 --- a/mm/memblock.c +++ b/mm/memblock.c @@ -1356,9 +1356,6 @@ static phys_addr_t __init memblock_alloc_range_nid(phys_addr_t size, align = SMP_CACHE_BYTES; } - if (end > memblock.current_limit) - end = memblock.current_limit; - again: found = memblock_find_in_range_node(size, align, start, end, nid, flags); > I also noticed that if I create a reserved memory node as was done one > imx6ul-ccimx6ulsom.dtsi the 3D seems to work again, but without it, I > was getting errors regardless of the 'cma=256M' or not. > I don't have a problem using the reserved memory, but I guess I am not > sure what the amount should be. I know for the video decoding 1080p, > I have historically used cma=128M, but with the 3D also needing some > memory allocation, is that enough or should I use 256M? > > adam -- Sincerely yours, Mike.
Re: [PATCH v2 00/21] Refine memblock API
On Wed, Sep 25, 2019 at 10:17 AM Fabio Estevam wrote: > > On Wed, Sep 25, 2019 at 9:17 AM Adam Ford wrote: > > > I tried cma=256M and noticed the cma dump at the beginning didn't > > change. Do we need to setup a reserved-memory node like > > imx6ul-ccimx6ulsom.dtsi did? > > I don't think so. > > Were you able to identify what was the exact commit that caused such > regression? I was able to narrow it down the 92d12f9544b7 ("memblock: refactor internal allocation functions") that caused the regression with Etnaviv. I also noticed that if I create a reserved memory node as was done one imx6ul-ccimx6ulsom.dtsi the 3D seems to work again, but without it, I was getting errors regardless of the 'cma=256M' or not. I don't have a problem using the reserved memory, but I guess I am not sure what the amount should be. I know for the video decoding 1080p, I have historically used cma=128M, but with the 3D also needing some memory allocation, is that enough or should I use 256M? adam
Re: [PATCH v2 00/21] Refine memblock API
On Wed, Sep 25, 2019 at 9:17 AM Adam Ford wrote: > I tried cma=256M and noticed the cma dump at the beginning didn't > change. Do we need to setup a reserved-memory node like > imx6ul-ccimx6ulsom.dtsi did? I don't think so. Were you able to identify what was the exact commit that caused such regression?
Re: [PATCH v2 00/21] Refine memblock API
On Wed, Sep 25, 2019 at 7:12 AM Fabio Estevam wrote: > > Hi Adam, > > On Wed, Sep 25, 2019 at 6:38 AM Adam Ford wrote: > > > I know it's rather late, but this patch broke the Etnaviv 3D graphics > > in my i.MX6Q. > > > > When I try to use the 3D, it returns some errors and the dmesg log > > shows some memory allocation errors too: > > [3.682347] etnaviv etnaviv: bound 13.gpu (ops gpu_ops) > > [3.688669] etnaviv etnaviv: bound 134000.gpu (ops gpu_ops) > > [3.695099] etnaviv etnaviv: bound 2204000.gpu (ops gpu_ops) > > [3.700800] etnaviv-gpu 13.gpu: model: GC2000, revision: 5108 > > [3.723013] etnaviv-gpu 13.gpu: command buffer outside valid > > memory window > > [3.731308] etnaviv-gpu 134000.gpu: model: GC320, revision: 5007 > > [3.752437] etnaviv-gpu 134000.gpu: command buffer outside valid > > memory window > > This looks similar to what was reported at: > https://bugs.freedesktop.org/show_bug.cgi?id=111789 > > Does it help if you use the same suggestion and pass cma=256M in your > kernel command line? I tried cma=256M and noticed the cma dump at the beginning didn't change. Do we need to setup a reserved-memory node like imx6ul-ccimx6ulsom.dtsi did? adam
Re: [PATCH v2 00/21] Refine memblock API
Hi Adam, On Wed, Sep 25, 2019 at 6:38 AM Adam Ford wrote: > I know it's rather late, but this patch broke the Etnaviv 3D graphics > in my i.MX6Q. > > When I try to use the 3D, it returns some errors and the dmesg log > shows some memory allocation errors too: > [3.682347] etnaviv etnaviv: bound 13.gpu (ops gpu_ops) > [3.688669] etnaviv etnaviv: bound 134000.gpu (ops gpu_ops) > [3.695099] etnaviv etnaviv: bound 2204000.gpu (ops gpu_ops) > [3.700800] etnaviv-gpu 13.gpu: model: GC2000, revision: 5108 > [3.723013] etnaviv-gpu 13.gpu: command buffer outside valid > memory window > [3.731308] etnaviv-gpu 134000.gpu: model: GC320, revision: 5007 > [3.752437] etnaviv-gpu 134000.gpu: command buffer outside valid > memory window This looks similar to what was reported at: https://bugs.freedesktop.org/show_bug.cgi?id=111789 Does it help if you use the same suggestion and pass cma=256M in your kernel command line?
Re: [PATCH v2 00/21] Refine memblock API
On Mon, Jan 21, 2019 at 2:05 AM Mike Rapoport wrote: > > Hi, > > Current memblock API is quite extensive and, which is more annoying, > duplicated. Except the low-level functions that allow searching for a free > memory region and marking it as reserved, memblock provides three (well, > two and a half) sets of functions to allocate memory. There are several > overlapping functions that return a physical address and there are > functions that return virtual address. Those that return the virtual > address may also clear the allocated memory. And, on top of all that, some > allocators panic and some return NULL in case of error. > > This set tries to reduce the mess, and trim down the amount of memblock > allocation methods. > > Patches 1-10 consolidate the functions that return physical address of > the allocated memory > > Patches 11-13 are some trivial cleanups > > Patches 14-19 add checks for the return value of memblock_alloc*() and > panics in case of errors. The patches 14-18 include some minor refactoring > to have better readability of the resulting code and patch 19 is a > mechanical addition of > > if (!ptr) > panic(); > > after memblock_alloc*() calls. > > And, finally, patches 20 and 21 remove panic() calls memblock and _nopanic > variants from memblock. > > v2 changes: > * replace some more %lu with %zu > * remove panics where they are not needed in s390 and in printk > * collect Acked-by and Reviewed-by. > > > Christophe Leroy (1): > powerpc: use memblock functions returning virtual address > > Mike Rapoport (20): > openrisc: prefer memblock APIs returning virtual address > memblock: replace memblock_alloc_base(ANYWHERE) with memblock_phys_alloc > memblock: drop memblock_alloc_base_nid() > memblock: emphasize that memblock_alloc_range() returns a physical address > memblock: memblock_phys_alloc_try_nid(): don't panic > memblock: memblock_phys_alloc(): don't panic > memblock: drop __memblock_alloc_base() > memblock: drop memblock_alloc_base() > memblock: refactor internal allocation functions > memblock: make memblock_find_in_range_node() and choose_memblock_flags() > static > arch: use memblock_alloc() instead of memblock_alloc_from(size, align, 0) > arch: don't memset(0) memory returned by memblock_alloc() > ia64: add checks for the return value of memblock_alloc*() > sparc: add checks for the return value of memblock_alloc*() > mm/percpu: add checks for the return value of memblock_alloc*() > init/main: add checks for the return value of memblock_alloc*() > swiotlb: add checks for the return value of memblock_alloc*() > treewide: add checks for the return value of memblock_alloc*() > memblock: memblock_alloc_try_nid: don't panic > memblock: drop memblock_alloc_*_nopanic() variants > I know it's rather late, but this patch broke the Etnaviv 3D graphics in my i.MX6Q. When I try to use the 3D, it returns some errors and the dmesg log shows some memory allocation errors too: [3.682347] etnaviv etnaviv: bound 13.gpu (ops gpu_ops) [3.688669] etnaviv etnaviv: bound 134000.gpu (ops gpu_ops) [3.695099] etnaviv etnaviv: bound 2204000.gpu (ops gpu_ops) [3.700800] etnaviv-gpu 13.gpu: model: GC2000, revision: 5108 [3.723013] etnaviv-gpu 13.gpu: command buffer outside valid memory window [3.731308] etnaviv-gpu 134000.gpu: model: GC320, revision: 5007 [3.752437] etnaviv-gpu 134000.gpu: command buffer outside valid memory window [3.760583] etnaviv-gpu 2204000.gpu: model: GC355, revision: 1215 [3.766766] etnaviv-gpu 2204000.gpu: Ignoring GPU with VG and FE2.0 [3.776131] [drm] Initialized etnaviv 1.2.0 20151214 for etnaviv on minor 0 # glmark2-es2-drm Error creating gpu Error: eglCreateWindowSurface failed with error: 0x3009 Error: eglCreateWindowSurface failed with error: 0x3009 Error: CanvasGeneric: Invalid EGL state Error: main: Could not initialize canvas Before this patch: [3.691995] etnaviv etnaviv: bound 13.gpu (ops gpu_ops) [3.698356] etnaviv etnaviv: bound 134000.gpu (ops gpu_ops) [3.704792] etnaviv etnaviv: bound 2204000.gpu (ops gpu_ops) [3.710488] etnaviv-gpu 13.gpu: model: GC2000, revision: 5108 [3.733649] etnaviv-gpu 134000.gpu: model: GC320, revision: 5007 [3.756115] etnaviv-gpu 2204000.gpu: model: GC355, revision: 1215 [3.762250] etnaviv-gpu 2204000.gpu: Ignoring GPU with VG and FE2.0 [3.771432] [drm] Initialized etnaviv 1.2.0 20151214 for etnaviv on minor 0 and the 3D gemos work without this. I don't know enough about the i.MX6 nor the 3D accelerator to know how to fix it. I am hoping someone in the know might have some suggestions. > arch/alpha/kernel/core_cia.c | 5 +- > arch/alpha/kernel/core_marvel.c | 6 + > arch/alpha/kernel/pci-noop.c | 13 +- > arch/alpha/kernel/pci.c | 11 +- > arch/alpha/kernel/pci_iommu.c | 16 +- > arch/alpha/kernel/setup.c
[PATCH v2 00/21] Refine memblock API
Hi, Current memblock API is quite extensive and, which is more annoying, duplicated. Except the low-level functions that allow searching for a free memory region and marking it as reserved, memblock provides three (well, two and a half) sets of functions to allocate memory. There are several overlapping functions that return a physical address and there are functions that return virtual address. Those that return the virtual address may also clear the allocated memory. And, on top of all that, some allocators panic and some return NULL in case of error. This set tries to reduce the mess, and trim down the amount of memblock allocation methods. Patches 1-10 consolidate the functions that return physical address of the allocated memory Patches 11-13 are some trivial cleanups Patches 14-19 add checks for the return value of memblock_alloc*() and panics in case of errors. The patches 14-18 include some minor refactoring to have better readability of the resulting code and patch 19 is a mechanical addition of if (!ptr) panic(); after memblock_alloc*() calls. And, finally, patches 20 and 21 remove panic() calls memblock and _nopanic variants from memblock. v2 changes: * replace some more %lu with %zu * remove panics where they are not needed in s390 and in printk * collect Acked-by and Reviewed-by. Christophe Leroy (1): powerpc: use memblock functions returning virtual address Mike Rapoport (20): openrisc: prefer memblock APIs returning virtual address memblock: replace memblock_alloc_base(ANYWHERE) with memblock_phys_alloc memblock: drop memblock_alloc_base_nid() memblock: emphasize that memblock_alloc_range() returns a physical address memblock: memblock_phys_alloc_try_nid(): don't panic memblock: memblock_phys_alloc(): don't panic memblock: drop __memblock_alloc_base() memblock: drop memblock_alloc_base() memblock: refactor internal allocation functions memblock: make memblock_find_in_range_node() and choose_memblock_flags() static arch: use memblock_alloc() instead of memblock_alloc_from(size, align, 0) arch: don't memset(0) memory returned by memblock_alloc() ia64: add checks for the return value of memblock_alloc*() sparc: add checks for the return value of memblock_alloc*() mm/percpu: add checks for the return value of memblock_alloc*() init/main: add checks for the return value of memblock_alloc*() swiotlb: add checks for the return value of memblock_alloc*() treewide: add checks for the return value of memblock_alloc*() memblock: memblock_alloc_try_nid: don't panic memblock: drop memblock_alloc_*_nopanic() variants arch/alpha/kernel/core_cia.c | 5 +- arch/alpha/kernel/core_marvel.c | 6 + arch/alpha/kernel/pci-noop.c | 13 +- arch/alpha/kernel/pci.c | 11 +- arch/alpha/kernel/pci_iommu.c | 16 +- arch/alpha/kernel/setup.c | 2 +- arch/arc/kernel/unwind.c | 3 +- arch/arc/mm/highmem.c | 4 + arch/arm/kernel/setup.c | 6 + arch/arm/mm/init.c| 6 +- arch/arm/mm/mmu.c | 14 +- arch/arm64/kernel/setup.c | 8 +- arch/arm64/mm/kasan_init.c| 10 ++ arch/arm64/mm/mmu.c | 2 + arch/arm64/mm/numa.c | 4 + arch/c6x/mm/dma-coherent.c| 4 + arch/c6x/mm/init.c| 4 +- arch/csky/mm/highmem.c| 5 + arch/h8300/mm/init.c | 4 +- arch/ia64/kernel/mca.c| 25 +-- arch/ia64/mm/contig.c | 8 +- arch/ia64/mm/discontig.c | 4 + arch/ia64/mm/init.c | 38 - arch/ia64/mm/tlb.c| 6 + arch/ia64/sn/kernel/io_common.c | 3 + arch/ia64/sn/kernel/setup.c | 12 +- arch/m68k/atari/stram.c | 4 + arch/m68k/mm/init.c | 3 + arch/m68k/mm/mcfmmu.c | 7 +- arch/m68k/mm/motorola.c | 9 ++ arch/m68k/mm/sun3mmu.c| 6 + arch/m68k/sun3/sun3dvma.c | 3 + arch/microblaze/mm/init.c | 10 +- arch/mips/cavium-octeon/dma-octeon.c | 3 + arch/mips/kernel/setup.c | 3 + arch/mips/kernel/traps.c | 5 +- arch/mips/mm/init.c | 5 + arch/nds32/mm/init.c | 12 ++ arch/openrisc/mm/init.c | 5 +- arch/openrisc/mm/ioremap.c| 8 +- arch/powerpc/kernel/dt_cpu_ftrs.c | 8 +- arch/powerpc/kernel/irq.c | 5 - arch/powerpc/kernel/paca.c| 6 +- arch/powerpc/kernel/pci_32.c | 3 + arch/powerpc/kernel/prom.c| 5 +- arch/powerpc/kernel/rtas.c