4:27 a.m.
-----Original Message-----
Hi Kazu,
On Wed, Mar 30, 2022 at 08:28:19AM +0000, HAGIO KAZUHITO(萩尾 一仁) wrote:
> -----Original Message-----
> > 1.) The vmcore file maybe very big.
> >
> > For example, I have a vmcore file which is over 23G,
> > and the panic kernel had 767.6G memory,
> > its max_sect_len is 4468736.
> >
> > Current code costs too much time to do the following loop:
> > ..............................................
> > for (i = 1; i < max_sect_len + 1; i++) {
> > dd->valid_pages[i] = dd->valid_pages[i - 1];
> > for (j = 0; j < BITMAP_SECT_LEN; j++, pfn++)
> > if (page_is_dumpable(pfn))
> > dd->valid_pages[i]++;
> > ..............................................
> >
> > For my case, it costs about 56 seconds to finish the
> > big loop.
> >
> > This patch moves the hweightXX macros to defs.h,
> > and uses hweight64 to optimize the loop.
> >
> > For my vmcore, the loop only costs about one second now.
> >
> > 2.) Tests result:
> > # cat ./commands.txt
> > quit
> >
> > Before:
> >
> > #echo 3 > /proc/sys/vm/drop_caches;
> > #time ./crash -i ./commands.txt /root/t/vmlinux /root/t/vmcore > /dev/null
2>&1
> > ............................
> > real 1m54.259s
> > user 1m12.494s
> > sys 0m3.857s
> > ............................
> >
> > After this patch:
> >
> > #echo 3 > /proc/sys/vm/drop_caches;
> > #time ./crash -i ./commands.txt /root/t/vmlinux /root/t/vmcore > /dev/null
2>&1
> > ............................
> > real 0m55.217s
> > user 0m15.114s
> > sys 0m3.560s
> > ............................
>
> Thank you for the improvement!
>
> as far as I tested on x86_64 it did not give such a big gain, but looking at
> the user time, it will do on arm64. Lianbo, can you reproduce on arm64?
>
> with a 192GB x86_64 dumpfile, slightly improved:
>
> $ time echo quit | ./crash vmlinux dump >/dev/null
>
> real 0m5.632s
Thanks for the testing.
I am curious why it costs only 5.632s for a 192G dumpfile?
How much memory of the panic kernel in the dumpfile?
My vmcore has 767.G memory, and the max_sect_len is 4468736.
I got it with makedumpfile -d 0 and tested it without dropping caches
to measure the change of the loop cost. As for memory, which size
are you saying? That machine has 192GB memory.
$ ls -lhs dump
193G -rw-------. 1 root root 193G Mar 30 17:07 dump
$ file dump
dump: Kdump compressed dump v6, system Linux, ...
$ ./crash vmlinux dump
MEMORY: 191.7 GB
crash> help -D
...
block_size: 4096
sub_hdr_size: 10
bitmap_blocks: 3088
max_mapnr: 50593791
...
total_valid_pages: 50178690
max_sect_len: 12352 // added
The max_sect_len looks too small comparing yours.. but
12352 * 4096 = 50593792
is almost same to max_mapnr.
Thanks,
Kazu
> user 0m5.545s
> sys 0m0.185s
>
> $ time echo quit | ./crash vmlinux dump >/dev/null
>
> real 0m5.547s
> user 0m5.477s
> sys 0m0.170s
>
> >
> > Signed-off-by: Huang Shijie <shijie(a)os.amperecomputing.com>
> > ---
> > defs.h | 20 ++++++++++++++++++++
> > diskdump.c | 12 +++++++++---
> > sbitmap.c | 19 -------------------
> > 3 files changed, 29 insertions(+), 22 deletions(-)
> >
> > diff --git a/defs.h b/defs.h
> > index 81ac049..1e8360d 100644
> > --- a/defs.h
> > +++ b/defs.h
> > @@ -4531,6 +4531,26 @@ struct machine_specific {
> > #define NUM_IN_BITMAP(bitmap, x) (bitmap[(x)/BITS_PER_LONG] &
NUM_TO_BIT(x))
> > #define SET_BIT(bitmap, x) (bitmap[(x)/BITS_PER_LONG] |= NUM_TO_BIT(x))
> >
> > +static inline unsigned int __const_hweight8(unsigned long w)
> > +{
> > + return
> > + (!!((w) & (1ULL << 0))) +
> > + (!!((w) & (1ULL << 1))) +
> > + (!!((w) & (1ULL << 2))) +
> > + (!!((w) & (1ULL << 3))) +
> > + (!!((w) & (1ULL << 4))) +
> > + (!!((w) & (1ULL << 5))) +
> > + (!!((w) & (1ULL << 6))) +
> > + (!!((w) & (1ULL << 7)));
> > +}
> > +
> > +#define __const_hweight16(w) (__const_hweight8(w) + __const_hweight8((w)
>> 8))
> > +#define __const_hweight32(w) (__const_hweight16(w) + __const_hweight16((w)
>> 16))
> > +#define __const_hweight64(w) (__const_hweight32(w) + __const_hweight32((w)
>> 32))
> > +
> > +#define hweight32(w) __const_hweight32(w)
> > +#define hweight64(w) __const_hweight64(w)
> > +
> > /*
> > * precision lengths for fprintf
> > */
> > diff --git a/diskdump.c b/diskdump.c
> > index d567427..d30db9d 100644
> > --- a/diskdump.c
> > +++ b/diskdump.c
> > @@ -547,6 +547,7 @@ read_dump_header(char *file)
> > ulong pfn;
> > int i, j, max_sect_len;
> > int is_split = 0;
> > + u64 tmp, *bitmap;
>
> $ make warn
> ...
> cc -c -g -DX86_64 -DLZO -DSNAPPY -DZSTD -DGDB_10_2 diskdump.c -Wall -O2
-Wstrict-prototypes
-Wmissing-prototypes -fstack-protector -Wformat-security
> diskdump.c: In function ‘read_dump_header’:
> diskdump.c:550:2: error: unknown type name ‘u64’
> u64 tmp, *bitmap;
> ^~~
> diskdump.c:905:12: error: ‘u64’ undeclared (first use in this function); did you
mean ‘a64l’?
> bitmap = (u64 *)dd->dumpable_bitmap;
> ^~~
> a64l
>
> It looks like u64 is defined only on arm and arm64, please use ulonglong
> commonly used in crash.
okay, no problem.
I will change it if we need a new version of this patch.
OK.
Thanks,
Kazu
Thanks
Huang Shijie