The basic objective here was to add support for "nohz_full=8-N" and/or
"rcu_nocbs="4-N" -- essentially introduce "N" as a portable reference
to the last core, evaluated at boot for anything using a CPU list.
The thinking behind this, is that people carve off a few early CPUs to
support housekeeping tasks, and perhaps dedicate one to a busy I/O
peripheral, and then the remaining pool of CPUs out to the end are a
part of a commonly configured pool used for the real work the user
cares about.
Extend that logic out to a fleet of machines - some new, and some
nearing EOL, and you've probably got a wide range of core counts to
contend with - even though the early number of cores dedicated to the
system overhead probably doesn't vary.
This change would enable sysadmins to have a common bootarg across all
such systems, and would also avoid any off-by-one fencepost errors that
happen for users who might briefly forget that core counts start at zero.
Originally I did this at the CPU subsys level, but Yury suggested it
be moved down further to bitmap level itself, which made the core
implementation smaller and less complex, but the series longer.
New self tests are added to better exercise what bitmap range/region
currently supports, and new tests are added for the new "N" support.
Also tested boot arg and the post-boot cgroup use case as per below:
root@hackbox:~# cat /proc/cmdline
BOOT_IMAGE=/boot/bzImage root=/dev/sda1 rcu_nocbs=2,3,8-N:1/2
root@hackbox:~# dmesg|grep Offl
rcu: Offload RCU callbacks from CPUs: 2-3,8,10,12,14.
root@hackbox:/sys/fs/cgroup/cpuset/foo# cat cpuset.cpus
root@hackbox:/sys/fs/cgroup/cpuset/foo# /bin/echo 10-N > cpuset.cpus
root@hackbox:/sys/fs/cgroup/cpuset/foo# cat cpuset.cpus
10-15
root@hackbox:/sys/fs/cgroup/cpuset/foo# /bin/echo N-N:N/N > cpuset.cpus
root@hackbox:/sys/fs/cgroup/cpuset/foo# cat cpuset.cpus
15
This was on a 16 core machine with CONFIG_NR_CPUS=16 in .config file.
Note that "N" is a dynamic quantity, and can change scope if the bitmap
is changed in size. So at the risk of stating the obvious, don't use it
for "burn_eFuse=128-N" or "secure_erase_firmware=32-N" type stuff.
Paul.
---
I've intentionally not gone down the rabbit hole of whether N or Z or
L is the better letter to mark the end of a mathematical set in the
hope that we can stay focused, and get this closed out here in v4.
Aside from that, I believe all other feedback has been responded to
in one way or another. Note that I didn't add Reviewed/Ack tags to
anything that changed significantly from what was reviewed in v3.
[v4: pair nbits with region, instead of inside it. Split EINVAL and
ERANGE tests. Don't handle start/end/offset within a macro to
abstract away nbits usage. Added some Reviwed-by/Ack tags.]
[v3: Allow "N" to be used anywhere in the region spec, i.e. "N-N:N/N" vs.
just being allowed at end of range like "0-N". Add new self-tests. Drop
"all" and "none" aliases as redundant and not worth the extra complication. ]
https://lore.kernel.org/lkml/[email protected]
[v2: push code down from cpu subsys to core bitmap code as per
Yury's comments. Change "last" to simply be "N" as per PeterZ.]
https://lore.kernel.org/lkml/[email protected]/
[v1: https://lore.kernel.org/lkml/20210106004850.GA11682@paulmck-ThinkPad-P72/
Cc: Li Zefan <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Yury Norov <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Josh Triplett <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: "Paul E. McKenney" <[email protected]>
Cc: Frederic Weisbecker <[email protected]>
Cc: Rasmus Villemoes <[email protected]>
Cc: Andy Shevchenko <[email protected]>
Paul Gortmaker (8):
lib: test_bitmap: clearly separate ERANGE from EINVAL tests.
lib: test_bitmap: add tests to trigger ERANGE case.
lib: test_bitmap: add more start-end:offset/len tests
lib: bitmap: move ERANGE check from set_region to check_region
lib: bitmap: pair nbits value with region struct
lib: bitmap: support "N" as an alias for size of bitmap
lib: test_bitmap: add tests for "N" alias
rcu: deprecate "all" option to rcu_nocbs=
.../admin-guide/kernel-parameters.rst | 7 +++
.../admin-guide/kernel-parameters.txt | 4 +-
kernel/rcu/tree_plugin.h | 6 +-
lib/bitmap.c | 62 +++++++++++++------
lib/test_bitmap.c | 46 ++++++++++++--
5 files changed, 93 insertions(+), 32 deletions(-)
--
2.17.1
While this is done for all bitmaps, the original use case in mind was
for CPU masks and cpulist_parse() as described below.
It seems that a common configuration is to use the 1st couple cores for
housekeeping tasks. This tends to leave the remaining ones to form a
pool of similarly configured cores to take on the real workload of
interest to the user.
So on machine A - with 32 cores, it could be 0-3 for "system" and then
4-31 being used in boot args like nohz_full=, or rcu_nocbs= as part of
setting up the worker pool of CPUs.
But then newer machine B is added, and it has 48 cores, and so while
the 0-3 part remains unchanged, the pool setup cpu list becomes 4-47.
Multiple deployment becomes easier when we can just simply replace 31
and 47 with "N" and let the system substitute in the actual number at
boot; a number that it knows better than we do.
Cc: Yury Norov <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: "Paul E. McKenney" <[email protected]>
Cc: Rasmus Villemoes <[email protected]>
Cc: Andy Shevchenko <[email protected]>
Suggested-by: Yury Norov <[email protected]> # move it from CPU code
Signed-off-by: Paul Gortmaker <[email protected]>
---
.../admin-guide/kernel-parameters.rst | 7 +++++
lib/bitmap.c | 27 ++++++++++++++-----
2 files changed, 27 insertions(+), 7 deletions(-)
diff --git a/Documentation/admin-guide/kernel-parameters.rst b/Documentation/admin-guide/kernel-parameters.rst
index 682ab28b5c94..7733a773f5f8 100644
--- a/Documentation/admin-guide/kernel-parameters.rst
+++ b/Documentation/admin-guide/kernel-parameters.rst
@@ -68,6 +68,13 @@ For example one can add to the command line following parameter:
where the final item represents CPUs 100,101,125,126,150,151,...
+The value "N" can be used to represent the numerically last CPU on the system,
+i.e "foo_cpus=16-N" would be equivalent to "16-31" on a 32 core system.
+
+Keep in mind that "N" is dynamic, so if system changes cause the bitmap width
+to change, such as less cores in the CPU list, then N and any ranges using N
+will also change. Use the same on a small 4 core system, and "16-N" becomes
+"16-3" and now the same boot input will be flagged as invalid (start > end).
This document may not be entirely up to date and comprehensive. The command
diff --git a/lib/bitmap.c b/lib/bitmap.c
index 6b568f98af3d..cc7cb1fca1ac 100644
--- a/lib/bitmap.c
+++ b/lib/bitmap.c
@@ -530,11 +530,17 @@ static int bitmap_check_region(const struct bitmap_region *br)
return 0;
}
-static const char *bitmap_getnum(const char *str, unsigned int *num)
+static const char *bitmap_getnum(const char *str, unsigned int *num,
+ unsigned int lastbit)
{
unsigned long long n;
unsigned int len;
+ if (str[0] == 'N') {
+ *num = lastbit;
+ return str + 1;
+ }
+
len = _parse_integer(str, 10, &n);
if (!len)
return ERR_PTR(-EINVAL);
@@ -580,9 +586,12 @@ static const char *bitmap_find_region_reverse(const char *start, const char *end
return end;
}
-static const char *bitmap_parse_region(const char *str, struct region *r)
+static const char *bitmap_parse_region(const char *str, struct bitmap_region *br)
{
- str = bitmap_getnum(str, &r->start);
+ struct region *r = br->r;
+ unsigned int lastbit = br->nbits - 1;
+
+ str = bitmap_getnum(str, &r->start, lastbit);
if (IS_ERR(str))
return str;
@@ -592,7 +601,7 @@ static const char *bitmap_parse_region(const char *str, struct region *r)
if (*str != '-')
return ERR_PTR(-EINVAL);
- str = bitmap_getnum(str + 1, &r->end);
+ str = bitmap_getnum(str + 1, &r->end, lastbit);
if (IS_ERR(str))
return str;
@@ -602,14 +611,14 @@ static const char *bitmap_parse_region(const char *str, struct region *r)
if (*str != ':')
return ERR_PTR(-EINVAL);
- str = bitmap_getnum(str + 1, &r->off);
+ str = bitmap_getnum(str + 1, &r->off, lastbit);
if (IS_ERR(str))
return str;
if (*str != '/')
return ERR_PTR(-EINVAL);
- return bitmap_getnum(str + 1, &r->group_len);
+ return bitmap_getnum(str + 1, &r->group_len, lastbit);
no_end:
r->end = r->start;
@@ -636,6 +645,10 @@ static const char *bitmap_parse_region(const char *str, struct region *r)
* From each group will be used only defined amount of bits.
* Syntax: range:used_size/group_size
* Example: 0-1023:2/256 ==> 0,1,256,257,512,513,768,769
+ * The value 'N' can be used as a dynamically substituted token for the
+ * maximum allowed value; i.e (nmaskbits - 1). Keep in mind that it is
+ * dynamic, so if system changes cause the bitmap width to change, such
+ * as more cores in a CPU list, then any ranges using N will also change.
*
* Returns: 0 on success, -errno on invalid input strings. Error values:
*
@@ -660,7 +673,7 @@ int bitmap_parselist(const char *buf, unsigned long *maskp, int nmaskbits)
if (buf == NULL)
return 0;
- buf = bitmap_parse_region(buf, &r);
+ buf = bitmap_parse_region(buf, &br);
if (IS_ERR(buf))
return PTR_ERR(buf);
--
2.17.1
It makes sense to do all the checks in check_region() and not 1/2
in check_region and 1/2 in set_region.
Since set_region is called immediately after check_region, the net
effect on runtime is zero, but it gets rid of an if (...) return...
Cc: Yury Norov <[email protected]>
Cc: Rasmus Villemoes <[email protected]>
Cc: Andy Shevchenko <[email protected]>
Acked-by: Yury Norov <[email protected]>
Signed-off-by: Paul Gortmaker <[email protected]>
---
lib/bitmap.c | 19 +++++++------------
1 file changed, 7 insertions(+), 12 deletions(-)
diff --git a/lib/bitmap.c b/lib/bitmap.c
index 75006c4036e9..9596ba53c36b 100644
--- a/lib/bitmap.c
+++ b/lib/bitmap.c
@@ -499,25 +499,22 @@ struct region {
unsigned int end;
};
-static int bitmap_set_region(const struct region *r,
- unsigned long *bitmap, int nbits)
+static void bitmap_set_region(const struct region *r, unsigned long *bitmap)
{
unsigned int start;
- if (r->end >= nbits)
- return -ERANGE;
-
for (start = r->start; start <= r->end; start += r->group_len)
bitmap_set(bitmap, start, min(r->end - start + 1, r->off));
-
- return 0;
}
-static int bitmap_check_region(const struct region *r)
+static int bitmap_check_region(const struct region *r, int nbits)
{
if (r->start > r->end || r->group_len == 0 || r->off > r->group_len)
return -EINVAL;
+ if (r->end >= nbits)
+ return -ERANGE;
+
return 0;
}
@@ -651,13 +648,11 @@ int bitmap_parselist(const char *buf, unsigned long *maskp, int nmaskbits)
if (IS_ERR(buf))
return PTR_ERR(buf);
- ret = bitmap_check_region(&r);
+ ret = bitmap_check_region(&r, nmaskbits);
if (ret)
return ret;
- ret = bitmap_set_region(&r, maskp, nmaskbits);
- if (ret)
- return ret;
+ bitmap_set_region(&r, maskp);
}
return 0;
--
2.17.1
On Tue, Feb 09, 2021 at 05:58:59PM -0500, Paul Gortmaker wrote:
> The basic objective here was to add support for "nohz_full=8-N" and/or
> "rcu_nocbs="4-N" -- essentially introduce "N" as a portable reference
> to the last core, evaluated at boot for anything using a CPU list.
I thought we kinda agreed that N is confusing and L is better.
N to me is equal to 32 on 32 core system as *number of cores / CPUs*. While L
sounds better as *last available CPU number*.
--
With Best Regards,
Andy Shevchenko
On Tue, Feb 09, 2021 at 05:59:03PM -0500, Paul Gortmaker wrote:
> It makes sense to do all the checks in check_region() and not 1/2
> in check_region and 1/2 in set_region.
>
> Since set_region is called immediately after check_region, the net
> effect on runtime is zero, but it gets rid of an if (...) return...
Reviewed-by: Andy Shevchenko <[email protected]>
> Cc: Yury Norov <[email protected]>
> Cc: Rasmus Villemoes <[email protected]>
> Cc: Andy Shevchenko <[email protected]>
> Acked-by: Yury Norov <[email protected]>
> Signed-off-by: Paul Gortmaker <[email protected]>
> ---
> lib/bitmap.c | 19 +++++++------------
> 1 file changed, 7 insertions(+), 12 deletions(-)
>
> diff --git a/lib/bitmap.c b/lib/bitmap.c
> index 75006c4036e9..9596ba53c36b 100644
> --- a/lib/bitmap.c
> +++ b/lib/bitmap.c
> @@ -499,25 +499,22 @@ struct region {
> unsigned int end;
> };
>
> -static int bitmap_set_region(const struct region *r,
> - unsigned long *bitmap, int nbits)
> +static void bitmap_set_region(const struct region *r, unsigned long *bitmap)
> {
> unsigned int start;
>
> - if (r->end >= nbits)
> - return -ERANGE;
> -
> for (start = r->start; start <= r->end; start += r->group_len)
> bitmap_set(bitmap, start, min(r->end - start + 1, r->off));
> -
> - return 0;
> }
>
> -static int bitmap_check_region(const struct region *r)
> +static int bitmap_check_region(const struct region *r, int nbits)
> {
> if (r->start > r->end || r->group_len == 0 || r->off > r->group_len)
> return -EINVAL;
>
> + if (r->end >= nbits)
> + return -ERANGE;
> +
> return 0;
> }
>
> @@ -651,13 +648,11 @@ int bitmap_parselist(const char *buf, unsigned long *maskp, int nmaskbits)
> if (IS_ERR(buf))
> return PTR_ERR(buf);
>
> - ret = bitmap_check_region(&r);
> + ret = bitmap_check_region(&r, nmaskbits);
> if (ret)
> return ret;
>
> - ret = bitmap_set_region(&r, maskp, nmaskbits);
> - if (ret)
> - return ret;
> + bitmap_set_region(&r, maskp);
> }
>
> return 0;
> --
> 2.17.1
>
--
With Best Regards,
Andy Shevchenko
On Wed, Feb 10, 2021 at 06:26:54PM +0200, Andy Shevchenko wrote:
> On Tue, Feb 09, 2021 at 05:58:59PM -0500, Paul Gortmaker wrote:
> > The basic objective here was to add support for "nohz_full=8-N" and/or
> > "rcu_nocbs="4-N" -- essentially introduce "N" as a portable reference
> > to the last core, evaluated at boot for anything using a CPU list.
>
> I thought we kinda agreed that N is confusing and L is better.
> N to me is equal to 32 on 32 core system as *number of cores / CPUs*. While L
> sounds better as *last available CPU number*.
The advantage of "N" is that people will automatically recognize it as
"last thing" or number of things" because "N" has long been used in
both senses. In contrast, someone seeing "0-L" for the first time is
likely to go "What???".
Besides, why would someone interpret "N" as "number of CPUs" when doing
that almost always gets you an invalid CPU number?
Thanx, Paul
On Wed, Feb 10, 2021 at 9:57 AM Paul E. McKenney <[email protected]> wrote:
>
> On Wed, Feb 10, 2021 at 06:26:54PM +0200, Andy Shevchenko wrote:
> > On Tue, Feb 09, 2021 at 05:58:59PM -0500, Paul Gortmaker wrote:
> > > The basic objective here was to add support for "nohz_full=8-N" and/or
> > > "rcu_nocbs="4-N" -- essentially introduce "N" as a portable reference
> > > to the last core, evaluated at boot for anything using a CPU list.
> >
> > I thought we kinda agreed that N is confusing and L is better.
> > N to me is equal to 32 on 32 core system as *number of cores / CPUs*. While L
> > sounds better as *last available CPU number*.
>
> The advantage of "N" is that people will automatically recognize it as
> "last thing" or number of things" because "N" has long been used in
> both senses. In contrast, someone seeing "0-L" for the first time is
> likely to go "What???".
>
> Besides, why would someone interpret "N" as "number of CPUs" when doing
> that almost always gets you an invalid CPU number?
>
> Thanx, Paul
I have no strong opinion about a letter, but I like Andy's idea to make it
case-insensitive.
There is another comment from the previous iteration not addressed so far.
This idea of the N notation is to make the bitmap list interface more robust
when we share the configs between different machines. What we have now
is definitely a good thing, but not completely portable except for cases
'N', '0-N' and 'N-N'.
For example, if one user adds rcu_nocbs= '4-N', and it works perfectly fine for
him, another user with s NR_CPUS == 2 will fail to boot with such a config.
This is not a problem of course in case of absolute values because nobody
guaranteed robustness. But this N feature would be barely useful in practice,
except for 'N', '0-N' and 'N-N' as I mentioned before, because there's always
a chance to end up with a broken config.
We can improve on robustness a lot if we take care about this case.For me,
the more reliable interface would look like this:
1. chunks without N work as before.
2. if 'a-N' is passed where a>=N, we drop chunk and print warning message
3. if 'a-N' is passed where a>=N together with a control key, we set last bit
and print warning.
For example, on 2-core CPU:
"4-2" --> error
"4-4" --> error
"4-N" --> drop and warn
"X, 4-N" --> set last bit and warn
Any comments?
On Wed, Feb 10, 2021 at 03:50:07PM -0800, Yury Norov wrote:
> On Wed, Feb 10, 2021 at 9:57 AM Paul E. McKenney <[email protected]> wrote:
> >
> > On Wed, Feb 10, 2021 at 06:26:54PM +0200, Andy Shevchenko wrote:
> > > On Tue, Feb 09, 2021 at 05:58:59PM -0500, Paul Gortmaker wrote:
> > > > The basic objective here was to add support for "nohz_full=8-N" and/or
> > > > "rcu_nocbs="4-N" -- essentially introduce "N" as a portable reference
> > > > to the last core, evaluated at boot for anything using a CPU list.
> > >
> > > I thought we kinda agreed that N is confusing and L is better.
> > > N to me is equal to 32 on 32 core system as *number of cores / CPUs*. While L
> > > sounds better as *last available CPU number*.
> >
> > The advantage of "N" is that people will automatically recognize it as
> > "last thing" or number of things" because "N" has long been used in
> > both senses. In contrast, someone seeing "0-L" for the first time is
> > likely to go "What???".
> >
> > Besides, why would someone interpret "N" as "number of CPUs" when doing
> > that almost always gets you an invalid CPU number?
> >
> > Thanx, Paul
>
> I have no strong opinion about a letter, but I like Andy's idea to make it
> case-insensitive.
>
> There is another comment from the previous iteration not addressed so far.
>
> This idea of the N notation is to make the bitmap list interface more robust
> when we share the configs between different machines. What we have now
> is definitely a good thing, but not completely portable except for cases
> 'N', '0-N' and 'N-N'.
>
> For example, if one user adds rcu_nocbs= '4-N', and it works perfectly fine for
> him, another user with s NR_CPUS == 2 will fail to boot with such a config.
>
> This is not a problem of course in case of absolute values because nobody
> guaranteed robustness. But this N feature would be barely useful in practice,
> except for 'N', '0-N' and 'N-N' as I mentioned before, because there's always
> a chance to end up with a broken config.
>
> We can improve on robustness a lot if we take care about this case.For me,
> the more reliable interface would look like this:
> 1. chunks without N work as before.
> 2. if 'a-N' is passed where a>=N, we drop chunk and print warning message
> 3. if 'a-N' is passed where a>=N together with a control key, we set last bit
> and print warning.
>
> For example, on 2-core CPU:
> "4-2" --> error
> "4-4" --> error
> "4-N" --> drop and warn
> "X, 4-N" --> set last bit and warn
>
> Any comments?
We really don't know the user's intent, and we cannot have complete
portability without knowing the user's intent. For example, "4-N" means
"all but the first four CPUs", in which case an error is appropriate
because "4-N" makes no more sense on a 2-CPU system than does "4-1".
I could see a potential desire for some notation for "the last two CPUs",
but let's please have a real need for such a thing before overengineering
this patch series any further.
To get the level of portability you seem to be looking for, we need some
higher-level automation that knows how many CPUs there are and what
the intent is. That automation can then generate the cpumasks for a
given system. But for more typical situations, what Paul has now will
work fine.
Paul Gortmaker's patch series is doing something useful. We should
not let potential future desires prevent us from taking a very useful
step forward.
Thanx, Paul
On 10/02/2021 18.57, Paul E. McKenney wrote:
> On Wed, Feb 10, 2021 at 06:26:54PM +0200, Andy Shevchenko wrote:
>> On Tue, Feb 09, 2021 at 05:58:59PM -0500, Paul Gortmaker wrote:
>>> The basic objective here was to add support for "nohz_full=8-N" and/or
>>> "rcu_nocbs="4-N" -- essentially introduce "N" as a portable reference
>>> to the last core, evaluated at boot for anything using a CPU list.
>>
>> I thought we kinda agreed that N is confusing and L is better.
>> N to me is equal to 32 on 32 core system as *number of cores / CPUs*. While L
>> sounds better as *last available CPU number*.
>
> The advantage of "N" is that people will automatically recognize it as
> "last thing" or number of things" because "N" has long been used in
> both senses. In contrast, someone seeing "0-L" for the first time is
> likely to go "What???".
Completely agree. The patch that introduces this even updates
Documentation/ at the same time, and if people are confused just because
they don't RTFM, xkcd#293 applies. So let's please just paint the
bikeshed N. (As for case insensitivity, I don't see the point, it just
makes documentation and implementation more cumbersome and confusing.
Just document and implement _one_ way of doing this.)
As for a future syntax for "last 4 cpus", it's common to accept a
negative index to mean count from the end, so unless we already accept
-4 as a shorthand for 0-4 (haven't checked), that could be -4-N. But
regardless, I also agree with Paul on this point, that's for a future
time when the need arises.
Rasmus
On Wed, Feb 10, 2021 at 04:23:09PM -0800, Paul E. McKenney wrote:
> On Wed, Feb 10, 2021 at 03:50:07PM -0800, Yury Norov wrote:
> > On Wed, Feb 10, 2021 at 9:57 AM Paul E. McKenney <[email protected]> wrote:
> > >
> > > On Wed, Feb 10, 2021 at 06:26:54PM +0200, Andy Shevchenko wrote:
> > > > On Tue, Feb 09, 2021 at 05:58:59PM -0500, Paul Gortmaker wrote:
> > > > > The basic objective here was to add support for "nohz_full=8-N" and/or
> > > > > "rcu_nocbs="4-N" -- essentially introduce "N" as a portable reference
> > > > > to the last core, evaluated at boot for anything using a CPU list.
> > > >
> > > > I thought we kinda agreed that N is confusing and L is better.
> > > > N to me is equal to 32 on 32 core system as *number of cores / CPUs*. While L
> > > > sounds better as *last available CPU number*.
> > >
> > > The advantage of "N" is that people will automatically recognize it as
> > > "last thing" or number of things" because "N" has long been used in
> > > both senses. In contrast, someone seeing "0-L" for the first time is
> > > likely to go "What???".
> > >
> > > Besides, why would someone interpret "N" as "number of CPUs" when doing
> > > that almost always gets you an invalid CPU number?
> > >
> > > Thanx, Paul
> >
> > I have no strong opinion about a letter, but I like Andy's idea to make it
> > case-insensitive.
> >
> > There is another comment from the previous iteration not addressed so far.
> >
> > This idea of the N notation is to make the bitmap list interface more robust
> > when we share the configs between different machines. What we have now
> > is definitely a good thing, but not completely portable except for cases
> > 'N', '0-N' and 'N-N'.
> >
> > For example, if one user adds rcu_nocbs= '4-N', and it works perfectly fine for
> > him, another user with s NR_CPUS == 2 will fail to boot with such a config.
> >
> > This is not a problem of course in case of absolute values because nobody
> > guaranteed robustness. But this N feature would be barely useful in practice,
> > except for 'N', '0-N' and 'N-N' as I mentioned before, because there's always
> > a chance to end up with a broken config.
> >
> > We can improve on robustness a lot if we take care about this case.For me,
> > the more reliable interface would look like this:
> > 1. chunks without N work as before.
> > 2. if 'a-N' is passed where a>=N, we drop chunk and print warning message
> > 3. if 'a-N' is passed where a>=N together with a control key, we set last bit
> > and print warning.
> >
> > For example, on 2-core CPU:
> > "4-2" --> error
> > "4-4" --> error
> > "4-N" --> drop and warn
> > "X, 4-N" --> set last bit and warn
> >
> > Any comments?
>
> We really don't know the user's intent, and we cannot have complete
> portability without knowing the user's intent. For example, "4-N" means
> "all but the first four CPUs", in which case an error is appropriate
> because "4-N" makes no more sense on a 2-CPU system than does "4-1".
> I could see a potential desire for some notation for "the last two CPUs",
> but let's please have a real need for such a thing before overengineering
> this patch series any further.
>
> To get the level of portability you seem to be looking for, we need some
> higher-level automation that knows how many CPUs there are and what
> the intent is. That automation can then generate the cpumasks for a
> given system. But for more typical situations, what Paul has now will
> work fine.
>
> Paul Gortmaker's patch series is doing something useful. We should
> not let potential future desires prevent us from taking a very useful
> step forward.
>
> Thanx, Paul
No problem, we can do it later if it will become a real concern.
Can you please remove this series from linux-next unless we finish
the review? It prevents me from applying the series from the LKML.
Yury
On Thu, Feb 11, 2021 at 04:23:39PM -0800, Yury Norov wrote:
> On Wed, Feb 10, 2021 at 04:23:09PM -0800, Paul E. McKenney wrote:
> > On Wed, Feb 10, 2021 at 03:50:07PM -0800, Yury Norov wrote:
> > > On Wed, Feb 10, 2021 at 9:57 AM Paul E. McKenney <[email protected]> wrote:
> > > >
> > > > On Wed, Feb 10, 2021 at 06:26:54PM +0200, Andy Shevchenko wrote:
> > > > > On Tue, Feb 09, 2021 at 05:58:59PM -0500, Paul Gortmaker wrote:
> > > > > > The basic objective here was to add support for "nohz_full=8-N" and/or
> > > > > > "rcu_nocbs="4-N" -- essentially introduce "N" as a portable reference
> > > > > > to the last core, evaluated at boot for anything using a CPU list.
> > > > >
> > > > > I thought we kinda agreed that N is confusing and L is better.
> > > > > N to me is equal to 32 on 32 core system as *number of cores / CPUs*. While L
> > > > > sounds better as *last available CPU number*.
> > > >
> > > > The advantage of "N" is that people will automatically recognize it as
> > > > "last thing" or number of things" because "N" has long been used in
> > > > both senses. In contrast, someone seeing "0-L" for the first time is
> > > > likely to go "What???".
> > > >
> > > > Besides, why would someone interpret "N" as "number of CPUs" when doing
> > > > that almost always gets you an invalid CPU number?
> > > >
> > > > Thanx, Paul
> > >
> > > I have no strong opinion about a letter, but I like Andy's idea to make it
> > > case-insensitive.
> > >
> > > There is another comment from the previous iteration not addressed so far.
> > >
> > > This idea of the N notation is to make the bitmap list interface more robust
> > > when we share the configs between different machines. What we have now
> > > is definitely a good thing, but not completely portable except for cases
> > > 'N', '0-N' and 'N-N'.
> > >
> > > For example, if one user adds rcu_nocbs= '4-N', and it works perfectly fine for
> > > him, another user with s NR_CPUS == 2 will fail to boot with such a config.
> > >
> > > This is not a problem of course in case of absolute values because nobody
> > > guaranteed robustness. But this N feature would be barely useful in practice,
> > > except for 'N', '0-N' and 'N-N' as I mentioned before, because there's always
> > > a chance to end up with a broken config.
> > >
> > > We can improve on robustness a lot if we take care about this case.For me,
> > > the more reliable interface would look like this:
> > > 1. chunks without N work as before.
> > > 2. if 'a-N' is passed where a>=N, we drop chunk and print warning message
> > > 3. if 'a-N' is passed where a>=N together with a control key, we set last bit
> > > and print warning.
> > >
> > > For example, on 2-core CPU:
> > > "4-2" --> error
> > > "4-4" --> error
> > > "4-N" --> drop and warn
> > > "X, 4-N" --> set last bit and warn
> > >
> > > Any comments?
> >
> > We really don't know the user's intent, and we cannot have complete
> > portability without knowing the user's intent. For example, "4-N" means
> > "all but the first four CPUs", in which case an error is appropriate
> > because "4-N" makes no more sense on a 2-CPU system than does "4-1".
> > I could see a potential desire for some notation for "the last two CPUs",
> > but let's please have a real need for such a thing before overengineering
> > this patch series any further.
> >
> > To get the level of portability you seem to be looking for, we need some
> > higher-level automation that knows how many CPUs there are and what
> > the intent is. That automation can then generate the cpumasks for a
> > given system. But for more typical situations, what Paul has now will
> > work fine.
> >
> > Paul Gortmaker's patch series is doing something useful. We should
> > not let potential future desires prevent us from taking a very useful
> > step forward.
> >
> > Thanx, Paul
>
> No problem, we can do it later if it will become a real concern.
>
> Can you please remove this series from linux-next unless we finish
> the review? It prevents me from applying the series from the LKML.
That will happen shortly, but in the meantime, just do the following on
top of -next before applying Paul's latest series:
git revert b3c314b ed78166 1e792c4 e831c73
Thanx, Paul
[Re: [PATCH v4 0/8] support for bitmap (and hence CPU) list "N" abbreviation] On 10/02/2021 (Wed 15:50) Yury Norov wrote:
> On Wed, Feb 10, 2021 at 9:57 AM Paul E. McKenney <[email protected]> wrote:
> >
> > On Wed, Feb 10, 2021 at 06:26:54PM +0200, Andy Shevchenko wrote:
> > > On Tue, Feb 09, 2021 at 05:58:59PM -0500, Paul Gortmaker wrote:
> > > > The basic objective here was to add support for "nohz_full=8-N" and/or
> > > > "rcu_nocbs="4-N" -- essentially introduce "N" as a portable reference
> > > > to the last core, evaluated at boot for anything using a CPU list.
> > >
> > > I thought we kinda agreed that N is confusing and L is better.
> > > N to me is equal to 32 on 32 core system as *number of cores / CPUs*. While L
> > > sounds better as *last available CPU number*.
> >
> > The advantage of "N" is that people will automatically recognize it as
> > "last thing" or number of things" because "N" has long been used in
> > both senses. In contrast, someone seeing "0-L" for the first time is
> > likely to go "What???".
> >
> > Besides, why would someone interpret "N" as "number of CPUs" when doing
> > that almost always gets you an invalid CPU number?
> >
> > Thanx, Paul
>
> I have no strong opinion about a letter, but I like Andy's idea to make it
> case-insensitive.
It is trivial to add later if someone can prove a genuine need for it,
but it is impossible to remove later if we add it now for no reason.
>
> There is another comment from the previous iteration not addressed so far.
Actually, no - it was addressed in detail already:
https://lore.kernel.org/lkml/[email protected]/
> This idea of the N notation is to make the bitmap list interface more robust
> when we share the configs between different machines. What we have now
> is definitely a good thing, but not completely portable except for cases
> 'N', '0-N' and 'N-N'.
>
> For example, if one user adds rcu_nocbs= '4-N', and it works perfectly fine for
> him, another user with s NR_CPUS == 2 will fail to boot with such a config.
Firstly there is no "fail to boot" from "rcu_nocbs=<invalid>" -- that
just doesn't happen. In any case, as you can see, I added in v4 the
documentation (as you requested) for this case - in several places.
And I explained in the thread above why any attempt to do some kind of
mapping policy was doomed to just add confusion and end up doing the
wrong thing. And the discussion ended with that.
So I'm not clear why it was brought up again here as if I just ignored
your "broken config" concerns and never addressed them.
In any case as others have indicated, it serves no immediate purpose to
over-think this and start adding corner case reactions to use cases that
simply don't exist and probably never will.
Thanks,
Paul.
--
>
> This is not a problem of course in case of absolute values because nobody
> guaranteed robustness. But this N feature would be barely useful in practice,
> except for 'N', '0-N' and 'N-N' as I mentioned before, because there's always
> a chance to end up with a broken config.
>
> We can improve on robustness a lot if we take care about this case.For me,
> the more reliable interface would look like this:
> 1. chunks without N work as before.
> 2. if 'a-N' is passed where a>=N, we drop chunk and print warning message
> 3. if 'a-N' is passed where a>=N together with a control key, we set last bit
> and print warning.
>
> For example, on 2-core CPU:
> "4-2" --> error
> "4-4" --> error
> "4-N" --> drop and warn
> "X, 4-N" --> set last bit and warn
>
> Any comments?