lkml.org 
[lkml]   [2006]   [May]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [Patch 5/6] statistics infrastructure
From
Date
Andrew, please pick up the update that comes in a separate mail.
It contains all changes as indicated below.

The update patch doesn't address parts of your feedback.
For details see also below, please.

Thanks, Martin

On Wed, 2006-05-24 at 15:57 -0700, Andrew Morton wrote:
> Martin Peschke <mp3@de.ibm.com> wrote:
> >
>
> It would be great to have a non-s390 exploiter of this code. So more
> people could try it out. Is that much work?
>
> One assumes that there's some subsytem or driver which has a real-life need
> for such instrumentation, although I don't know which one that would be.
> (And if there is no such subsystem then that's rather a black mark for
> merging all this code!)
>
> Thoughts?

Our initial requirement has been to provide SCSI statistics. I have got
a patch for that (latencies etc.). Unfortunately, it requires more work.
I have found out that SCSI device creation doesn't guarantee process
context. Setup up statistics might call schedule() (see
debugs_create_file).

I am going to discuss the issue with linux-scsi. I have seen patches
that seem to move SCSI device scanning to a kernel thread, which
might be welcome help.

> > ...
> >
> > +struct statistic_discipline {
> > + int (*parse)(struct statistic *, struct statistic_info *, int, char *);
> > + void* (*alloc)(struct statistic *, size_t, gfp_t, int);
> > + void (*free)(struct statistic *, void *);
> > + void (*reset)(struct statistic *, void *);
> > + void (*merge)(struct statistic *, void *, void*);
> > + int (*fdata)(struct statistic *, const char *,
> > + struct statistic_file_private *, void *);
> > + int (*fdef)(struct statistic *, char *);
> > + void (*add)(struct statistic *, int, s64, u64);
> > + void (*set)(struct statistic *, s64, u64);
> > + char *name;
> > + size_t size;
> > +};
>
> This practice of omitting the variable names drives me up the wall, sorry.
> Look at the definition of `add' then fall down dazed and confused.
>
> This is particularly true of these function-pointer style declarations.
> For example, do:
>
> $EDITOR -t aio_read
>
> and you end up here:
>
> ssize_t (*aio_read) (struct kiocb *, char __user *, size_t, loff_t);
>
> which is uninformative. You have to go and hunt down an instance of an
> aio_read() implementation to actually be sure what those args are doing.
>
> So I think putting the nicely-chosen variable names in there is quite
> helpful.

done

> > +#ifdef CONFIG_STATISTICS
> > +
> > +extern int statistic_create(struct statistic_interface *, const char *);
> > +extern int statistic_remove(struct statistic_interface *);
> > +
> > +/**
> > + * statistic_add - update statistic with incremental data in (X, Y) pair
> > + * @stat: struct statistic array
> > + * @i: index of statistic to be updated
> > + * @value: X
> > + * @incr: Y
> > + *
> > + * The actual processing of the (X, Y) data pair is determined by the current
> > + * the definition applied to the statistic. See Documentation/statistics.txt.
> > + *
> > + * This variant takes care of protecting per-cpu data. It is preferred whenever
> > + * exploiters don't update several statistics of the same entity in one go.
> > + */
> > +static inline void statistic_add(struct statistic *stat, int i,
> > + s64 value, u64 incr)
> > +{
> > + unsigned long flags;
> > + local_irq_save(flags);
> > + if (stat[i].state == STATISTIC_STATE_ON)
> > + stat[i].add(&stat[i], smp_processor_id(), value, incr);
> > + local_irq_restore(flags);
> > +}
>
> afaict this isn't actually used?
>
> If it is, and assuming this is only accessed via a function pointer (the
> mysterious `add' method) then there's not a lot of point in inlining it.
>
> Except if this code really isn't called, then inlining it will avoid having
> an unused piece of code in vmlinux.
>
> But if it _is_ used, and it has multiple users then we end up with multiple
> copies in vmlinux.
>
> So what's up with that?

That's one of the few functions that exploiter's are supposed to call
for statistic updates. This inline function then calls the mysterious
add function of the data processing mode that is being used at that
time.

Just revert inlining?

> And elsewhere we have:
>
> > +static inline void statistic_add(struct statistic *stat, int i,
> > + s64 value, u64 incr)
> > +{
> > +}
> > +
>
> Do we expect this to have any callers if !CONFIG_STATISTICS?

This relieves exploiters from the burden of taking care of
!CONFIG_STATISTICS. I think something like this in ten thousand to-be
exploiters would look pretty ugly:

#ifdef CONFIG_STATISTICS
statistic_add(mystat, index, value, incr);
#endif

>
> > +static int statistic_free(struct statistic *stat, struct statistic_info *info)
> > +{
> > + int cpu;
> > + stat->state = STATISTIC_STATE_RELEASED;
> > + if (unlikely(info->flags & STATISTIC_FLAGS_NOINCR)) {
> > + statistic_free_ptr(stat, stat->pdata);
> > + stat->pdata = NULL;
> > + return 0;
> > + }
> > + for_each_cpu(cpu) {
>
> for_each_cpu() is on death row. Replace it with for_each_possible_cpu().
> If that is indeed appropriate - perhaps you meant online_cpu, or
> present_cpu.

done

> > +static void * statistic_alloc_generic(struct statistic *stat, size_t size,
> ^
>
> unwelcome space ;)

done

> > +static int statistic_alloc(struct statistic *stat,
> > + struct statistic_info *info)
> > +{
> > + int cpu;
> > + stat->age = sched_clock();
>
> argh. Didn't we end up finding a way to avoid this?
>
> At the least, we should have statistics_clock(), or nsec_clock(), or
> something which is decoupled from this low-level scheduler-internal thing,
> and which architectures can implement (vis attribute-weak) if they have a
> preferred/better/more-accurate alternative.

That's something to address next. Sorry, didn't manage to change it for
the recent update patch.

> > +static int statistic_transition(struct statistic *stat,
> > + struct statistic_info *info,
> > + enum statistic_state requested_state)
> > +{
> > + int z = (requested_state < stat->state ? 1 : 0);
> > + int retval = -EINVAL;
> > +
> > + while (stat->state != requested_state) {
> > + switch (stat->state) {
> > + case STATISTIC_STATE_INVALID:
> > + retval = ( z ? -EINVAL : statistic_initialise(stat) );
> > + break;
> > + case STATISTIC_STATE_UNCONFIGURED:
> > + retval = ( z ? statistic_uninitialise(stat)
> > + : statistic_define(stat) );
> > + break;
> > + case STATISTIC_STATE_RELEASED:
> > + retval = ( z ? statistic_initialise(stat)
> > + : statistic_alloc(stat, info) );
> > + break;
> > + case STATISTIC_STATE_OFF:
> > + retval = ( z ? statistic_free(stat, info)
> > + : statistic_start(stat) );
> > + break;
> > + case STATISTIC_STATE_ON:
> > + retval = ( z ? statistic_stop(stat) : -EINVAL );
> > + break;
>
> Lots of unneeded parentheses there.

done

> > +static int statistic_reset(struct statistic *stat, struct statistic_info *info)
> > +{
> > + enum statistic_state prev_state = stat->state;
> > + int cpu;
> > +
> > + if (unlikely(stat->state < STATISTIC_STATE_OFF))
> > + return 0;
> > + statistic_transition(stat, info, STATISTIC_STATE_OFF);
> > + if (unlikely(info->flags & STATISTIC_FLAGS_NOINCR))
> > + statistic_reset_ptr(stat, stat->pdata);
> > + else
> > + for_each_cpu(cpu)
>
> for_each_possible_cpu() (maybe)

correct, done

> > +static inline int statistic_fdata(struct statistic_interface *interface, int i,
> > + struct statistic_file_private *fpriv)
> > +{
> > + struct statistic *stat = &interface->stat[i];
> > + struct statistic_info *info = &interface->info[i];
> > + struct statistic_discipline *disc = &statistic_discs[stat->type];
> > + struct statistic_merge_private mpriv;
> > + int retval;
> > +
> > + if (unlikely(stat->state < STATISTIC_STATE_OFF))
> > + return 0;
> > + if (unlikely(info->flags & STATISTIC_FLAGS_NOINCR))
> > + return disc->fdata(stat, info->name, fpriv, stat->pdata);
> > + mpriv.dst = statistic_alloc_ptr(stat, GFP_KERNEL, -1);
> > + if (unlikely(!mpriv.dst))
> > + return -ENOMEM;
> > + spin_lock_init(&mpriv.lock);
> > + mpriv.stat = stat;
> > + on_each_cpu(statistic_merge, &mpriv, 0, 1);
> > + retval = disc->fdata(stat, info->name, fpriv, mpriv.dst);
> > + statistic_free_ptr(stat, mpriv.dst);
> > + return retval;
> > +}
>
> You do like that `inline' thingy ;)

changed

> > +/* cpu hotplug handling for per-cpu data */
> > +
> > +static inline int _statistic_hotcpu(struct statistic_interface *interface,
> > + int i, unsigned long action, int cpu)
> > +{
> > + struct statistic *stat = &interface->stat[i];
> > + struct statistic_info *info = &interface->info[i];
> > +
> > + if (unlikely(info->flags & STATISTIC_FLAGS_NOINCR))
> > + return 0;
> > + if (stat->state < STATISTIC_STATE_OFF)
> > + return 0;
> > + switch (action) {
> > + case CPU_UP_PREPARE:
> > + stat->pdata->ptrs[cpu] = statistic_alloc_ptr(stat, GFP_ATOMIC,
> > + cpu_to_node(cpu));
> > + break;
>
> So this allocation can fail. Does all the other code handle that? If not,
> we should fail the CPU bringup.

done

> Dangit, this is inlined as well. It makes oops-tracing really hard :(

changed

> > +{
> > + statistic_root_dir = debugfs_create_dir(STATISTIC_ROOT_DIR, NULL);
> > + if (unlikely(!statistic_root_dir))
> > + return -ENOMEM;
> > + INIT_LIST_HEAD(&statistic_list);
> > + mutex_init(&statistic_list_mutex);
> > + register_cpu_notifier(&statistic_hotcpu_notifier);
>
> Actually, this can fail too (well, actually it can't, but the API suggests
> it can).

I see. grep indicates that nobody cares. I guess if
register_cpu_notifier ever starts failing, the kernel will do lots of
funny things anyway. I tend to leave my code the way it is for the time
being and to leave this kernel-wide job to someone else.

> > + int offset;
> > + struct list_head line_lh;
> > + char *nl;
> > + size_t line_size = 0;
> > +
> > + INIT_LIST_HEAD(&line_lh);
>
> LIST_HEAD(line_lh);

done

> > +
> > +/* code concerned with utilisation indicator statistic */
> > +
> > +static void statistic_reset_util(struct statistic *stat, void *ptr)
> > +{
> > + struct statistic_entry_util *util = ptr;
> > + util->num = 0;
> > + util->acc = 0;
> > + util->min = (~0ULL >> 1) - 1;
> > + util->max = -(~0ULL >> 1) + 1;
> > +}
>
> `min' is a large positive number and `max' is a large negative one. Is that
> right?
>
> `min' gets 0x7ffffffffffffffe, which seems to be off-by-one.

fixed off-by-one

> Consider using LLONG_MAX and friends.

changed

> > +static int statistic_fdata_util(struct statistic *stat, const char *name,
> > + struct statistic_file_private *fpriv,
> > + void *data)
> > +{
> > + struct sgrb_seg *seg;
> > + struct statistic_entry_util *util = data;
> > + unsigned long long whole = 0;
> > + signed long long min = 0, max = 0, decimal = 0, last_digit;
> > +
> > + seg = sgrb_seg_find(&fpriv->read_seg_lh, 128);
> > + if (unlikely(!seg))
> > + return -ENOMEM;
> > + if (likely(util->num)) {
> > + whole = util->acc;
> > + do_div(whole, util->num);
> > + decimal = util->acc * 10000;
> > + do_div(decimal, util->num);
> > + decimal -= whole * 10000;
> > + if (decimal < 0)
> > + decimal = -decimal;
> > + last_digit = decimal;
> > + do_div(last_digit, 10);
> > + last_digit = decimal - last_digit * 10;
> > + if (last_digit >= 5)
> > + decimal += 10;
> > + do_div(decimal, 10);
> > + min = util->min;
> > + max = util->max;
> > + }
> > + seg->offset += sprintf(seg->address + seg->offset,
> > + "%s %Lu %Ld %Ld.%03lld %Ld\n", name,
> > + (unsigned long long)util->num,
> > + (signed long long)min, whole, decimal,
> > + (signed long long)max);
>
> There's no need to cast `min' and `max' here. A cast would be needed if
> they were u64/s64.

done

> > +
> > +static inline int statistic_add_sparse_new(struct statistic_sparse_list *slist,
> > + s64 value, u64 incr)
> > +{
> > + struct statistic_entry_sparse *entry;
> > +
> > + if (unlikely(slist->entries == slist->entries_max))
> > + return -ENOMEM;
> > + entry = kmalloc(sizeof(struct statistic_entry_sparse), GFP_ATOMIC);
> > + if (unlikely(!entry))
> > + return -ENOMEM;
> > + entry->value = value;
> > + entry->hits = incr;
> > + slist->entries++;
> > + list_add_tail(&entry->list, &slist->entry_lh);
> > + return 0;
> > +}
> >
> > +static inline void _statistic_add_sparse(struct statistic_sparse_list *slist,
> > + s64 value, u64 incr)
> > +{
> > + struct list_head *head = &slist->entry_lh;
> > + struct statistic_entry_sparse *entry;
> > +
> > + list_for_each_entry(entry, head, list) {
> > + if (likely(entry->value == value)) {
> > + entry->hits += incr;
> > + statistic_add_sparse_sort(head, entry);
> > + return;
> > + }
> > + }
> > + if (unlikely(statistic_add_sparse_new(slist, value, incr)))
> > + slist->hits_missed += incr;
> > +}
>
> I hereby revoke your inlining license.

I admit my failure. Changed almost all of these places.

> > +static void statistic_set_sparse(struct statistic *stat, s64 value, u64 total)
> > +{
> > + struct statistic_sparse_list *slist = (struct statistic_sparse_list *)
> > + stat->pdata;
>
> Hang on, what's happening here? statistic.pdata is `struct percpu_data *'.
> That's
>
> struct percpu_data {
> void *ptrs[NR_CPUS];
> };
>
> How can we cast that to a statistic_sparse_list* and then start playing
> with it? We're supposed to use per_cpu_ptr() to get at the actual data.

With regard to the data that a statistic feeds on, there are are two
types of statistics: statistics that accumulate incremental updates
(pushed - probably frequently - through statistic_add() or
statistic_inc()), and statistics that accept total numbers (pulled
through statistic_set() only when read by user). We use per-cpu data for
the former. As to the latter, per-cpu data would be way to heavy.
That is why, my code is capable of dealing with both per-cpu data and
non-per-cpu data. Since a particular statistic is either per-cpu or
non-per-cpu, I use the same data pointer for both cases in order to keep
struct statistic as small as possible.

I admit the cast looks a bit fishy. But lines above are correct.

I should probably rename 'pdata' in struct statistic to 'data' (to
reflect its versatile use), change its type from struct percpu_data* to
void*, and finally use per_cpu_ptr. per_cpu_ptr works fine any pointer
type. The only issue with per_cpu_ptr is that I can't convert

stat->pdata->ptrs[cpu] = some_buf;

to

per_cpu_ptr(stat->pdata, cpu) = some_buf;

as done in my code instead of calling alloc_percpu() in order to avoid
eating up memory for offline or unavailable cpus. See cpu hotplug
handler.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2006-05-30 00:19    [from the cache]
©2003-2011 Jasper Spaans