[RFC PATCH 00/22] riscv: s64ilp32: Running 32-bit Linux kernel on 64-bit supervisor mode

Guo Ren guoren at kernel.org
Sat May 20 08:57:45 PDT 2023


On Sat, May 20, 2023 at 6:13 PM Arnd Bergmann <arnd at arndb.de> wrote:
>
> On Sat, May 20, 2023, at 04:53, Guo Ren wrote:
> > On Sat, May 20, 2023 at 4:20 AM Arnd Bergmann <arnd at arndb.de> wrote:
> >> On Thu, May 18, 2023, at 15:09, guoren at kernel.org wrote:
> >>
> >> I've tried to run the same numbers for the debate about running
> >> 32-bit vs 64-bit arm kernels in the past, but focused mostly on
> >> slightly larger systems, but I looked mainly at the 512MB case,
> >> as that is the most cost-efficient DDR3 memory configuration
> >> and fairly common.
> > 512MB is extravagant, in my opinion. In the IPC market, 32/64MB is for
> > 480P/720P/1080p, 128/256MB is for 1080p/2k, and 512/1024MB is for 4K.
> >> 512MB chips is less than 5% of the total (I guess). Even in 512MB
> > chips, the additional memory is for the frame buffer, not the Linux
> > system.
>
> This depends a lot on the target application of course. For
> a phone or NAS box, 512MB is probably the lower limit.
>
> What I observe in arch/arm/ devicetree submissions, in board-db.org,
> and when looking at industrial Arm board vendor websites  is that
> 512MB is the most common configuration, and I think 1GB is still
> more common than 256MB even for 32-bit machines. There is of course
> a difference between number of individual products, and number of
> machines shipped in a given configuration, and I guess you have
> a good point that the cheapest ones are also the ones that ship
> in the highest volume.
>
> >> What I'd like to understand better in your example is where
> >> the 14MB of memory went. I assume this is for 128MB of total
> >> RAM, so we know that 1MB went into additional 'struct page'
> >> objects (32 bytes * 32768 pages). It would be good to know
> >> where the dynamic allocations went and if they are  reclaimable
> >> (e.g. inodes) or non-reclaimable (e.g. kmalloc-128).
> >>
> >> For the vmlinux size, is this already a minimal config
> >> that one would run on a board with 128MB of RAM, or a
> >> defconfig that includes a lot of stuff that is only relevant
> >> for other platforms but also grows on 64-bit?
> > It's not minimal config, it's defconfig. So I say it's a roungh
> > measurement :)
> >
> > I admit I wanted a little bit to exaggerate it, but that's the
> > starting point for cutting down memory usage for most people, right?
> > During the past year, we have been convincing our customers to use the
> > s64lp64 + u32ilp32, but they can't tolerate even 1% memory additional
> > cost in 64MB/128MB scenarios and then chose cortex-a7/a35, which could
> > run 32-bit Linux. I think it's too early to talk about throwing 32-bit
> > Linux into the garbage, not only for the reason of memory footprint
> > but also for the ingrained opinion of the people. Changing their mind
> > needs a long time.
> >
> >>
> >> What do you see in /proc/slabinfo, /proc/meminfo/, and
> >> 'size vmlinux' for the s64ilp32 and s64lp64 kernels here?
> > Both s64ilp32 & s64lp64 use the same u32ilp32_rootfs.ext2 binary and
> > the same opensbi binary.
> > All are opensbi(2MB) + Linux(126MB) memory layout.
> >
> > Here is the result:
> >
> > s64ilp32:
> > [    0.000000] Virtual kernel memory layout:
> > [    0.000000]       fixmap : 0x9ce00000 - 0x9d000000   (2048 kB)
> > [    0.000000]       pci io : 0x9d000000 - 0x9e000000   (  16 MB)
> > [    0.000000]      vmemmap : 0x9e000000 - 0xa0000000   (  32 MB)
> > [    0.000000]      vmalloc : 0xa0000000 - 0xc0000000   ( 512 MB)
> > [    0.000000]       lowmem : 0xc0000000 - 0xc7e00000   ( 126 MB)
> > [    0.000000] Memory: 97748K/129024K available (8699K kernel code,
> > 8867K rwdata, 4096K rodata, 4204K init, 361K bss, 31276K reserved, 0K
> > cma-reserved)
>
> Ok, so it saves only a little bit on .text/.init/.bss/.rodata, but
> there is a 4MB difference in rwdata, and a total of 10.4MB difference
> in "reserved" size, which I think includes all of the above plus
> the mem_map[] array.
>
> 89380K/131072K available (8638K kernel code, 4979K rwdata, 4096K rodata, 2191K init, 477K bss, 41692K reserved, 0K cma-reserved)
>
> Oddly, I don't see anywhere close to 8KB in a riscv64 defconfig
> build (linux-next, gcc-13), so I don't know where that comes
> from:
>
> $ size -A build/tmp/vmlinux | sort -k2 -nr | head
> Total                   13518684
> .text                    8896058   18446744071562076160
> .rodata                  2219008   18446744071576748032
> .data                     933760   18446744071583039488
> .bss                      476080   18446744071584092160
> .init.text                264718   18446744071572553728
> __ksymtab_strings         183986   18446744071579214312
> __ksymtab_gpl             122928   18446744071579091384
> __ksymtab                 109080   18446744071578982304
> __bug_table                98352   18446744071583973248
>
>
>
> > KReclaimable:        644 kB
> > Slab:               4536 kB
> > SReclaimable:        644 kB
> > SUnreclaim:         3892 kB
> > KernelStack:         344 kB
>
> These look like the only notable differences in meminfo:
>
>   KReclaimable:       1092 kB
>   Slab:               6900 kB
>   SReclaimable:       1092 kB
>   SUnreclaim:         5808 kB
>   KernelStack:         688 kB
>
> The largest chunk here is 2MB in non-reclaimable slab allocations,
> or a 50% growth of those.
>
> The kernel stacks are doubled as expected, but that's only 344KB,
> similarly for reclaimable slabs.
>
> > # cat /proc/slabinfo
> >
> >                                                              [68/1691]
> > slabinfo - version: 2.1
> > # name            <active_objs> <num_objs> <objsize> <objperslab>
> > <pagesperslab> : tunables <limit> <batchcount> <sharedfactor> :
> > slabdata <active_slabs> <num_slabs> <sharedavail>
> > ext4_groupinfo_1k     28     28    144   28    1 : tunables    0    0
> >   0 : slabdata      1      1      0
> > p9_req_t               0      0    104   39    1 : tunables    0    0
>
> Did you perhaps miss a few lines while pasting these? It seems
> odd that some caches only show up in the ilp32 case (proc_dir_entry,
> bd2_journa_handle, buffer_head, biovec_max, anon_vma_chain, ...) and
> some others are only in the lp64 case (UNIX, ext4_prealloc_space,
> files_cache, filp, ip_fib_alias, task_struct, uid_cache, ...).
>
> Looking at the ones that are in both and have the largest size
> increase, I see
>
> # lp64
> 1788 kernfs_node_cache 14304 128
>  590 shmem_inode_cache 646 936
>  272 inode_cache 360 776
>  153 ext4_inode_cache 105 1496
>  250 dentry 1188 216
>  192 names_cache 48 4096
>  199 radix_tree_node 350 584
>  307 kmalloc-64 4912 64
>   60 kmalloc-128 480 128
>   47 kmalloc-192 252 192
>  204 kmalloc-256 816 256
>   72 kmalloc-512 144 512
>  840 kmalloc-1k 840 1024
>
> # ilp32
> 1197 kernfs_node_cache 13938 88
>  373 shmem_inode_cache 637 600
>  174 inode_cache 360 496
>   84 ext4_inode_cache 88 984
>  177 dentry 1196 152
>   32 names_cache 8 4096
>  100 radix_tree_node 338 304
>  331 kmalloc-64 5302 64
>  132 kmalloc-128 1056 128
>   23 kmalloc-192 126 192
>   16 kmalloc-256 64 256
>  428 kmalloc-512 856 512
>   88 kmalloc-1k 88 1024
>
> So sysfs (kernfs_node_cache) has the largest chunk of the
> 2MB non-reclaimable slab, grown 50% from 1.2MB to 1.8MB.
> In some cases, this could be avoided entirely by turning
> off sysfs, but most users can't do that.
> shmem_inode_cache is probably mostly devtmpfs, the
> other inode caches ones are smaller and likely reclaimable.
>
> It's interesting how the largest slab cache ends up
> being the kmalloc-1k cache (840 1K objects) on lp64,
> but the kmalloc-512 cache (856 512B objects) on ilp32.
> My guess is that the majority of this is from a single
> callsite that has an allocation groing just beyond 512B.
> This alone seems significant enough to need further
> investigation, I would hope we can completely avoid
> these by adding a custom slab cache. I don't see this
> effect on an arm64 boot though, for me the 512B allocations
> are much higher the 1K ones.
>
> Maybe you can identify the culprit using the boot-time traces
> as listed in https://elinux.org/Kernel_dynamic_memory_analysis#Dynamic
> That might help everyone running a 64-bit kernel on
> low-memory configurations, though it would of course slightly
> weaken your argument for an ilp32 kernel ;-)
Thx for the detailed reply, I would try your approches mentioned
lately. But these about traditional CONFIG_32BIT v.s. CONFIG_64BIT
comparation.

Besides the detailed analysis data, we also would meet the people's
concept problem. Such as struct page, struct list_head, and some
variables containing pointers, ilp32's would be significantly smaller
than lp64. That means ilp32 is smaller than lp64 in people's minds.
This concept would prevent vendors from accepting lp64 as a cost-down
solution. They even won't try, which I've met these years. I was an
lp64 kernel supporter last year, but I met a lot of arguments on
s64lp64 + u32ilp32. Some guys are using arm32 Linux; they want to stay
on 32-bit Linux to ensure their complex C code can work. So our
argument about "ilp32 v.s. lp64" won't have a result.

Let's change another view, cache utilization. These 64/128MB SoCs also
have limited cache capacities (L1-32KB+L2-128KB/only L1-64KB). Such as
List walk and stack saving/restoring are very common in Linux. What do
you think about "32-bit v.s. 64-bit" cache utilization?

>
>      Arnd



-- 
Best Regards
 Guo Ren



More information about the linux-riscv mailing list