RE: [PATCH v7 2/4] topology: use bin_attribute to break the size limitation of cpumap ABI

[Song Bao Hua (Barry Song)]

> -----Original Message-----
> From: Yury Norov [mailto:yury.norov@xxxxxxxxx]
> Sent: Tuesday, July 20, 2021 5:10 AM
> To: Song Bao Hua (Barry Song) <song.bao.hua@xxxxxxxxxxxxx>
> Cc: andriy.shevchenko@xxxxxxxxxxxxxxx; gregkh@xxxxxxxxxxxxxxxxxxx;
> akpm@xxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> dave.hansen@xxxxxxxxx; linux@xxxxxxxxxxxxxxxxxx; rafael@xxxxxxxxxx;
> rdunlap@xxxxxxxxxxxxx; agordeev@xxxxxxxxxxxxx; sbrivio@xxxxxxxxxx;
> jianpeng.ma@xxxxxxxxx; valentin.schneider@xxxxxxx; peterz@xxxxxxxxxxxxx;
> bristot@xxxxxxxxxx; guodong.xu@xxxxxxxxxx; tangchengchang
> <tangchengchang@xxxxxxxxxx>; Zengtao (B) <prime.zeng@xxxxxxxxxxxxx>;
> yangyicong <yangyicong@xxxxxxxxxx>; tim.c.chen@xxxxxxxxxxxxxxx; Linuxarm
> <linuxarm@xxxxxxxxxx>; tiantao (H) <tiantao6@xxxxxxxxxxxxx>
> Subject: Re: [PATCH v7 2/4] topology: use bin_attribute to break the size
> limitation of cpumap ABI
> 
> On Mon, Jul 19, 2021 at 11:10:45AM +0000, Song Bao Hua (Barry Song) wrote:
> >
> >
> > > -----Original Message-----
> > > From: andriy.shevchenko@xxxxxxxxxxxxxxx
> > > [mailto:andriy.shevchenko@xxxxxxxxxxxxxxx]
> > > Sent: Monday, July 19, 2021 9:07 PM
> > > To: Yury Norov <yury.norov@xxxxxxxxx>
> > > Cc: Song Bao Hua (Barry Song) <song.bao.hua@xxxxxxxxxxxxx>;
> > > gregkh@xxxxxxxxxxxxxxxxxxx; akpm@xxxxxxxxxxxxxxxxxxxx;
> > > linux-kernel@xxxxxxxxxxxxxxx; dave.hansen@xxxxxxxxx;
> > > linux@xxxxxxxxxxxxxxxxxx; rafael@xxxxxxxxxx; rdunlap@xxxxxxxxxxxxx;
> > > agordeev@xxxxxxxxxxxxx; sbrivio@xxxxxxxxxx; jianpeng.ma@xxxxxxxxx;
> > > valentin.schneider@xxxxxxx; peterz@xxxxxxxxxxxxx; bristot@xxxxxxxxxx;
> > > guodong.xu@xxxxxxxxxx; tangchengchang <tangchengchang@xxxxxxxxxx>;
> Zengtao (B)
> > > <prime.zeng@xxxxxxxxxxxxx>; yangyicong <yangyicong@xxxxxxxxxx>;
> > > tim.c.chen@xxxxxxxxxxxxxxx; Linuxarm <linuxarm@xxxxxxxxxx>; tiantao (H)
> > > <tiantao6@xxxxxxxxxxxxx>
> > > Subject: Re: [PATCH v7 2/4] topology: use bin_attribute to break the size
> > > limitation of cpumap ABI
> > >
> > > On Fri, Jul 16, 2021 at 06:12:21PM -0700, Yury Norov wrote:
> > > > On Sat, Jul 17, 2021 at 12:16:48AM +0000, Song Bao Hua (Barry Song) wrote:
> > > > > > From: Yury Norov [mailto:yury.norov@xxxxxxxxx]
> > > > > > Sent: Saturday, July 17, 2021 8:04 AM
> > > > > > On Fri, Jul 16, 2021 at 08:49:58AM +0000, Song Bao Hua (Barry Song)
> wrote:
> > >
> > > ...
> > >
> > > > > Generally good idea. However, for sysfs ABI entries, it might not be
> > > > > that true.
> > > > >
> > > > > A sysfs entry might never be read for its whole life. As I explained
> > > > > before, a sysfs entry - especially for list,  is randomly "cat" by users.
> > > > > Many of them won't be read forever. And after they are read once, they
> > > > > will probably never be read again. The operations to read ABI could
> be
> > > > > random and rare. Performance wouldn't be a concern.
> > > > >
> > > > > To avoid holding the memory which might never be used, it is better
> to
> > > > > allocate and free the memory during runtime. I mean to allocate in show()
> > > > > and free in show(), aka, to do it on demand.
> > > > >
> > > > > For example, for a server with 256CPU and each cpu has dozens of sysfs
> ABI
> > > > > entries, only a few of sysfs list entries might be randomly "cat" by
> users.
> > > > > Holding 256*entries memory doesn't look good.
> > > >
> > > > Ok, makes sense.
> > > >
> > > > > > This would require to add bitmap_max_string_size(list, bitmap, nbits),
> > > > > > but it's O(1), and I think, others will find it helpful.
> > > > >
> > > > > What about getting size and memory at the same time?
> > > >
> > > > 1. We already have kasprintf()
> > > > 2. It breaks coding style.
> > > >
> > > > Documentation/process/coding-style.rst:
> > > >         Functions should be short and sweet, and do just one thing.
> > > >
> > > > From practical point of view, there should be some balance between
> > > > granularity and ease-of-use. But in this case, bitmap_list cries for
> > > > a function that will help to estimate size of output buffer.
> > >
> > > According to the vsnprintf() logic the estimated size is what it returns.
> If
> > > user supplies too few bytes available, the comparison with the returned
> value
> > > can tell caller that space wasn't big enough.
> 
> snprintf(NULL, 0, "pbl", ...) also works, but it's O(nbits), and user is not
> guaranteed that allocated memory would be always sufficient. I mean max possible
> length for given nbits, not a length of a specific string.
> 
> In case of lists, the length may grow. Consider:
>         0-8 -> 0-3,5-8 -> 0,2,4,6,8
> 
> If we want to allocate a storage for strings that may change, it would be
> helpful to allocate memory for the lengthiest string in advance.
> 
> So, bitmap_max_string_len() may be a convenient O(1) alternative for
> those who interested in printing the same bitmap in the same buffer.
> 
> > As far as my understanding, for estimated size in bitmap_max_string_size()
> > Yury might mean something as below?
> >
> > * For bitmask:
> > Each 32bit needs 9 bytes "11111111,", so the max size of mask would be:
> > 9*nr_cpus / 32 ?
> 
> 11111 -> "f1", but your formula gives 1.
> I think it should be like this (not tested):
>         DIV_ROUND_UP(nbits, 4) + nbits < 32 ? 0 : nbits / 32 - 1

I am not quite sure this is correct, in case nbits=64,
we will need:
ffffffff,ffffffff   9+8=17

your formula is ignoring the "," for each 32bits?

> 
> > * For list:
> > Maximally  cpu0-9 need 2bytes, like "1,"
> > Maximally cpu10-99 need 3bytes, like "50,"
> > Maximally cpu100-999 need 4bytes, like "101,"
> > Maximally cpu1000-9999 need 5 bytes..
> >
> > So once we know the size of the bitmap, we can figure out the maximum
> > size of its string for mask and list?
> >
> > Pls correct me if you don't mean this, yury.
> 
> Assuming that longest possible strings are of the form 0,2,4,6,... I
> think it's correct except for the last comma, so substract 1.

I don't think 0,2,4,6,8 takes maximum length,
0-1,3-4,6-7 are longer.

It seems maximum bitmap pattern is(binary):
110110110110...

In this case, each 3bits takes 4 bytes.

For cpu 10-99, it would be: 10-11,13-14,16-17,19-20,

Each 3bits takes "6 bytes".

For cpu100-999, it would be: 100-101, 103-104, 106-107,....
Each 3bits takes "8 bytes"?

Anyway, It seems be quite tricky :-)

> 
> If we decide to go on with this bitmap_max_strlen(), the list part
> should be tested extensively.
> 

Btw, sysfs core code needs changes to support pre-allocated
memory in customized size and seq read on it. This needs a lot
of efforts I am looking at.

> > > > And it's
> > > > easy to imagine a case where the estimated length of bitmap is needed
> > > > explicitly:
> > > >
> > > >         size_t max_size = bitmap_max_string_size(nbits);
> > > >         char *buf = kmalloc(PAGE_ALIGN(max_size) * nr_cpus);
> > > >
> > > > Thought, I don't insist. In your driver you can do:
> > > >
> > > >         size_t size = snprintf(NULL, 0, ...);
> > > >         void *buf = kmalloc(size);
> > > >
> > > > It will be fully correct, and you already have everything you need.
> > > >
> > > > > ssize_t bitmap_get_print_buf(bool list, char **buf, const unsigned long
> > > > >  *maskp, int nmaskbits)
> > > > >
> > > > > ssize_t cpumap_get_print_buf(bool list, char **buf, const struct cpumask
> > > *mask);
> > > > >
> > > > > This API returns the size of printed buffer, and it also gets the
> > > > > printed result saved in *buf. Then drivers don't need to do three
> > > > > steps:
> > > > >
> > > > > 1. get cpumap buffer size which is your cpumap_max_string_size()
> > > > > 2. allocate memory for buffer according to size got in step 1
> > > > > 3. print bitmap(cpumap) to buffer by "pbl"
> > > > >
> > > > > It will only need to call bitmap_get_print_buf() and all three
> > > > > things are done inside bitmap_get_print_buf().
> > > > >
> > > > > How to use the size and memory allocated in cpumap_get_print_buf
> > > > > will be totally up to users.
> > > > >
> > > > > The other benefit for this is that if we get string size during
> initialization,
> > > > > and then we print in show() entries, the size got at the beginning might
> > > be not
> > > > > enough as system topology might have changed. Sysfs ABI reflects the
> status
> > > of
> > > > > system at this moment.
> > >
> > > --
> > > With Best Regards,
> > > Andy Shevchenko
> > >
> >


Thanks
Barry