Re: [PATCH v6 1/4] cpumask: introduce cpumap_print_to_buf to support large bitmask and list

From: Andy Shevchenko
Date: Fri Jul 09 2021 - 07:23:59 EST


On Fri, Jul 09, 2021 at 07:55:41PM +1200, Barry Song wrote:
> From: Tian Tao <tiantao6@xxxxxxxxxxxxx>
>
> The existing cpumap_print_to_pagebuf() is used by cpu topology and other
> drivers to export hexadecimal bitmask and decimal list to userspace by
> sysfs ABI.
>
> Right now, those drivers are using a normal attribute for this kind of
> ABIs. A normal attribute typically has show entry as below:
>
> static ssize_t example_dev_show(struct device *dev,
> struct device_attribute *attr, char *buf)
> {
> ...
> return cpumap_print_to_pagebuf(true, buf, &pmu_mmdc->cpu);
> }
> show entry of attribute has no offset and count parameters and this
> means the file is limited to one page only.
>
> cpumap_print_to_pagebuf() API works terribly well for this kind of
> normal attribute with buf parameter and without offset, count:
>
> static inline ssize_t
> cpumap_print_to_pagebuf(bool list, char *buf, const struct cpumask *mask)
> {
> return bitmap_print_to_pagebuf(list, buf, cpumask_bits(mask),
> nr_cpu_ids);
> }
>
> The problem is once we have many cpus, we have a chance to make bitmask
> or list more than one page. Especially for list, it could be as complex
> as 0,3,5,7,9,...... We have no simple way to know it exact size.
>
> It turns out bin_attribute is a way to break this limit. bin_attribute
> has show entry as below:
> static ssize_t
> example_bin_attribute_show(struct file *filp, struct kobject *kobj,
> struct bin_attribute *attr, char *buf,
> loff_t offset, size_t count)
> {
> ...
> }
>
> With the new offset and count parameters, this makes sysfs ABI be able
> to support file size more than one page. For example, offset could be
> >= 4096.
>
> This patch introduces cpumap_print_to_buf() so that those drivers can
> move to bin_attribute to support large bitmask and list. In result,
> we have to pass the corresponding parameters from bin_attribute to this
> new API.

This is rather good and explanatory commit message. I'm wondering if we can
make parts of it in some documentation.


> Signed-off-by: Tian Tao <tiantao6@xxxxxxxxxxxxx>
> Reviewed-by: Jonathan Cameron <jonathan.cameron@xxxxxxxxxx>
> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Cc: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> Cc: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
> Cc: Stefano Brivio <sbrivio@xxxxxxxxxx>
> Cc: Alexander Gordeev <agordeev@xxxxxxxxxxxxx>
> Cc: "Ma, Jianpeng" <jianpeng.ma@xxxxxxxxx>
> Cc: Yury Norov <yury.norov@xxxxxxxxx>
> Cc: Valentin Schneider <valentin.schneider@xxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Cc: Daniel Bristot de Oliveira <bristot@xxxxxxxxxx>
> Signed-off-by: Barry Song <song.bao.hua@xxxxxxxxxxxxx>
> ---
> -v6:
> -minor cleanup doc according to Andy Shevchenko's comment;
> -take bitmap_print_to_buf back according to Yury Norov's comment and
> fix the documents;
> -Sorry, Yury, I don't think it is doable to move memory allocation
> to drivers.
> Considering a driver like topology.c, we have M CPUs and each CPU
> has N various nodes like core_siblings, package_cpus, die_cpus etc,
> we can't know the size of each node of each CPU in advance. The best
> time to get the size of each node is really when users read the sysfs
> node. otherwise, we have to scan M*N nodes in drivers in advance to
> figure out the exact size of buffers we need.
> On the other hand, it is crazily tricky to ask a bundle of drivers to
> find a proper place to save the pointer of allocated buffers so that
> they can be re-used in second read of the same bin_attribute node.
> And I doubt it is really useful to save the address of buffers
> somewhere. Immediately freeing it seems to be a correct option to
> avoid runtime waste of memory. We can't predict when users will read
> topology ABI and which node users will read.
> Finally, reading topology things wouldn't be the really cpu-bound
> things in user applications, hardly this kind of ABI things can be
> a performance bottleneck. Users use numactl and lstopo commands to
> read ABIs but nobody will do it again and again. And a normal
> application won't read topology repeatly. So the overhead caused by
> malloc/free in the new bitmap API doesn't really matter.
> if we really want a place to re-used the buffer and avoid malloc/free,
> it seems this should be done in some common place rather than each
> driver. still it is hard to find the best place.
>
> Thanks for the comments of Yury and Andy in v5.
>
> include/linux/bitmap.h | 2 ++
> include/linux/cpumask.h | 24 +++++++++++++++++++++++
> lib/bitmap.c | 43 +++++++++++++++++++++++++++++++++++++++++
> 3 files changed, 69 insertions(+)
>
> diff --git a/include/linux/bitmap.h b/include/linux/bitmap.h
> index a36cfcec4e77..0de6effa2797 100644
> --- a/include/linux/bitmap.h
> +++ b/include/linux/bitmap.h
> @@ -226,6 +226,8 @@ void bitmap_copy_le(unsigned long *dst, const unsigned long *src, unsigned int n
> unsigned int bitmap_ord_to_pos(const unsigned long *bitmap, unsigned int ord, unsigned int nbits);
> int bitmap_print_to_pagebuf(bool list, char *buf,
> const unsigned long *maskp, int nmaskbits);
> +int bitmap_print_to_buf(bool list, char *buf, const unsigned long *maskp,
> + int nmaskbits, loff_t off, size_t count);
>
> #define BITMAP_FIRST_WORD_MASK(start) (~0UL << ((start) & (BITS_PER_LONG - 1)))
> #define BITMAP_LAST_WORD_MASK(nbits) (~0UL >> (-(nbits) & (BITS_PER_LONG - 1)))
> diff --git a/include/linux/cpumask.h b/include/linux/cpumask.h
> index bfc4690de4f4..8a89d133fa2d 100644
> --- a/include/linux/cpumask.h
> +++ b/include/linux/cpumask.h
> @@ -983,6 +983,30 @@ cpumap_print_to_pagebuf(bool list, char *buf, const struct cpumask *mask)
> nr_cpu_ids);
> }
>
> +/**
> + * cpumap_print_to_buf - copies the cpumask into the buffer
> + * @list: indicates whether the cpumap must be list

a list

> + * true: print in decimal list format
> + * false: print in hexadecimal bitmask format

Perhaps explanation deserves to be a paragraph below?

> + * @mask: the cpumask to copy
> + * @buf: the buffer to copy into
> + * @off: in the string from which we are copying, We copy to @buf
> + * @count: the maximum number of bytes to print
> + *
> + * The function copies the cpumask into the buffer either as comma-separated
> + * list of cpus or hex values of cpumask; Typically used by bin_attribute to
> + * export cpumask bitmask and list ABI.
> + *
> + * Returns the length of how many bytes have been copied.
> + */
> +static inline ssize_t
> +cpumap_print_to_buf(bool list, char *buf, const struct cpumask *mask,
> + loff_t off, size_t count)
> +{
> + return bitmap_print_to_buf(list, buf, cpumask_bits(mask),
> + nr_cpu_ids, off, count);
> +}
> +
> #if NR_CPUS <= BITS_PER_LONG
> #define CPU_MASK_ALL \
> (cpumask_t) { { \
> diff --git a/lib/bitmap.c b/lib/bitmap.c
> index 9401d39e4722..c64baa3a8606 100644
> --- a/lib/bitmap.c
> +++ b/lib/bitmap.c
> @@ -487,6 +487,49 @@ int bitmap_print_to_pagebuf(bool list, char *buf, const unsigned long *maskp,
> }
> EXPORT_SYMBOL(bitmap_print_to_pagebuf);
>
> +/**
> + * bitmap_print_to_buf - convert bitmap to list or hex format ASCII string
> + * @list: indicates whether the bitmap must be list

a list

> + * true: print in decimal list format
> + * false: print in hexadecimal bitmask format

Perhaps explanation deserves to be a paragraph below?

> + * @buf: buffer into which string is placed
> + * @maskp: pointer to bitmap to convert
> + * @nmaskbits: size of bitmap, in bits
> + * @off: in the string from which we are copying, We copy to @buf
> + * @count: the maximum number of bytes to print
> + *
> + * The role of cpumap_print_to_buf() and cpumap_print_to_pagebuf() is similar,
> + * the difference is that bitmap_print_to_pagebuf() mainly serves sysfs
> + * attribute with the assumption the destination buffer is exactly one page
> + * aligned with PAGE_SIZE and it won't be more than one page, thus,
> + * bitmap_print_to_pagebuf() needs neither offset to copy from nor count
> + * which is the length we are going to copy. cpumap_print_to_buf(), on the
> + * other hand, mainly serves bin_attribute which doesn't work with exact
> + * one page, and it has explicit parameters like "offset" to copy from and
> + * "count" bytes to copy. So cpumap_print_to_buf() can break the size limit
> + * of converted decimal list and hexadecimal bitmask. And buf doesn't have
> + * to be exactly one page.
> + *
> + * Returns the number of characters actually printed to @buf
> + */
> +int bitmap_print_to_buf(bool list, char *buf, const unsigned long *maskp,
> + int nmaskbits, loff_t off, size_t count)
> +{
> + const char *fmt = list ? "%*pbl\n" : "%*pb\n";
> + ssize_t size;
> + void *data;
> +
> + data = kasprintf(GFP_KERNEL, fmt, nmaskbits, maskp);
> + if (!data)
> + return -ENOMEM;
> +
> + size = memory_read_from_buffer(buf, count, &off, data, strlen(data) + 1);
> + kfree(data);
> +
> + return size;
> +}
> +EXPORT_SYMBOL(bitmap_print_to_buf);
> +
> /*
> * Region 9-38:4/10 describes the following bitmap structure:
> * 0 9 12 18 38 N
> --
> 2.25.1
>

--
With Best Regards,
Andy Shevchenko