Message-ID: <CAGAzgsox7bfh4fLBCjJvFeXAJRWx9imsrFVgQYa=q1ooX2Mi1w@mail.gmail.com>
Date:   Wed, 24 Oct 2018 13:36:42 -0700
From:   "dbasehore ." <dbasehore@...omium.org>
To:     jbrunet@...libre.com
Cc:     linux-kernel <linux-kernel@...r.kernel.org>,
        linux-clk@...r.kernel.org, linux-arm-kernel@...ts.infradead.org,
        linux-rockchip@...ts.infradead.org, linux-doc@...r.kernel.org,
        sboyd@...nel.org, Michael Turquette <mturquette@...libre.com>,
        Heiko Stübner <heiko@...ech.de>,
        aisheng.dong@....com, mchehab+samsung@...nel.org,
        Jonathan Corbet <corbet@....net>,
        Stephen Boyd <sboyd@...eaurora.org>
Subject: Re: [PATCH 1/6] clk: Remove recursion in clk_core_{prepare,enable}()

On Wed, Oct 24, 2018 at 2:51 AM Jerome Brunet <jbrunet@...libre.com> wrote:
>
> On Tue, 2018-10-23 at 18:31 -0700, Derek Basehore wrote:
> > From: Stephen Boyd <sboyd@...eaurora.org>
> >
> > Enabling and preparing clocks can be written quite naturally with
> > recursion. We start at some point in the tree and recurse up the
> > tree to find the oldest parent clk that needs to be enabled or
> > prepared. Then we enable/prepare and return to the caller, going
> > back to the clk we started at and enabling/preparing along the
> > way.
> >
> > The problem is recursion isn't great for kernel code where we
> > have a limited stack size. Furthermore, we may be calling this
> > code inside clk_set_rate() which also has recursion in it, so
> > we're really not looking good if we encounter a tall clk tree.
> >
> > Let's create a stack instead by looping over the parent chain and
> > collecting clks of interest. Then the enable/prepare becomes as
> > simple as iterating over that list and calling enable.
>
> Hi Derek,
>
> What about unprepare() and disable() ?
>
> This patch removes the recursion from the enable path but keeps it for the
> disable path ... this is very odd. Assuming doing so works, It certainly makes
> CCF a lot harder to understand.

It wasn't there in the original patch. I asked Stephen, and he thinks
it may have been left that way because unprepare/disable are tail
recursion cases, which the compiler can optimize away.

>
> What about clock protection which essentially works on the same model as prepare
> and enable ?
>
> Overall, this change does not look like something that should be merged as it
> is. If you were just seeking comments, you should add the "RFC" tag to your
> series.
>
> Jerome.
>
> >
> > Cc: Jerome Brunet <jbrunet@...libre.com>
>
> If you don't mind, I would prefer to get the whole series next time. It helps to
> get the context.
>
> > Signed-off-by: Stephen Boyd <sboyd@...eaurora.org>
> > Signed-off-by: Derek Basehore <dbasehore@...omium.org>
> > ---
> >  drivers/clk/clk.c | 113 ++++++++++++++++++++++++++--------------------
> >  1 file changed, 64 insertions(+), 49 deletions(-)
> >
> > diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
> > index af011974d4ec..95d818f5edb2 100644
> > --- a/drivers/clk/clk.c
> > +++ b/drivers/clk/clk.c
> > @@ -71,6 +71,8 @@ struct clk_core {
> >       struct hlist_head       children;
> >       struct hlist_node       child_node;
> >       struct hlist_head       clks;
> > +     struct list_head        prepare_list;
> > +     struct list_head        enable_list;
> >       unsigned int            notifier_count;
> >  #ifdef CONFIG_DEBUG_FS
> >       struct dentry           *dentry;
> > @@ -740,49 +742,48 @@ EXPORT_SYMBOL_GPL(clk_unprepare);
> >  static int clk_core_prepare(struct clk_core *core)
> >  {
> >       int ret = 0;
> > +     struct clk_core *tmp, *parent;
> > +     LIST_HEAD(head);
> >
> >       lockdep_assert_held(&prepare_lock);
> >
> > -     if (!core)
> > -             return 0;
> > +     while (core) {
> > +             list_add(&core->prepare_list, &head);
> > +             /* Stop once we see a clk that is already prepared */
> > +             if (core->prepare_count)
> > +                     break;
> > +             core = core->parent;
> > +     }
> >
> > -     if (core->prepare_count == 0) {
> > -             ret = clk_pm_runtime_get(core);
> > -             if (ret)
> > -                     return ret;
> > +     list_for_each_entry_safe(core, tmp, &head, prepare_list) {
> > +             list_del_init(&core->prepare_list);
>
> Is there any point in removing it from the list ?
> Maybe I missed it but it does not seems useful.
>
> Without this, we could use list_for_each_entry()
>
> >
> > -             ret = clk_core_prepare(core->parent);
> > -             if (ret)
> > -                     goto runtime_put;
> > +             if (core->prepare_count == 0) {
>
> Should we really check the count here ? You are not checking the count when the
> put() counterpart is called below.
>
> Since PM runtime has ref counting as well, either way would work I guess ... but
> we shall be consistent
>
> > +                     ret = clk_pm_runtime_get(core);
> > +                     if (ret)
> > +                             goto err;
> >
> > -             trace_clk_prepare(core);
> > +                     trace_clk_prepare(core);
> >
> > -             if (core->ops->prepare)
> > -                     ret = core->ops->prepare(core->hw);
> > +                     if (core->ops->prepare)
> > +                             ret = core->ops->prepare(core->hw);
> >
> > -             trace_clk_prepare_complete(core);
> > +                     trace_clk_prepare_complete(core);
> >
> > -             if (ret)
> > -                     goto unprepare;
> > +                     if (ret) {
> > +                             clk_pm_runtime_put(core);
> > +                             goto err;
> > +                     }
> > +             }
> > +             core->prepare_count++;
> >       }
> >
> > -     core->prepare_count++;
> > -
> > -     /*
> > -      * CLK_SET_RATE_GATE is a special case of clock protection
> > -      * Instead of a consumer claiming exclusive rate control, it is
> > -      * actually the provider which prevents any consumer from making any
> > -      * operation which could result in a rate change or rate glitch while
> > -      * the clock is prepared.
> > -      */
> > -     if (core->flags & CLK_SET_RATE_GATE)
> > -             clk_core_rate_protect(core);
>
> This gets removed without anything replacing it.
>
> is CLK_SET_RATE_GATE and clock protection support dropped after this change ?
>
> > -
> >       return 0;
> > -unprepare:
> > -     clk_core_unprepare(core->parent);
> > -runtime_put:
> > -     clk_pm_runtime_put(core);
> > +err:
> > +     parent = core->parent;
> > +     list_for_each_entry_safe_continue(core, tmp, &head, prepare_list)
> > +             list_del_init(&core->prepare_list);
> > +     clk_core_unprepare(parent);
>
> If you get here because of failure clk_pm_runtime_get(), you will unprepare a
> clock which may have not been prepared first
>
> Overall the rework of error exit path does not seem right (or necessary)
>
> >       return ret;
> >  }
> >
> > @@ -878,37 +879,49 @@ EXPORT_SYMBOL_GPL(clk_disable);
> >  static int clk_core_enable(struct clk_core *core)
> >  {
> >       int ret = 0;
> > +     struct clk_core *tmp, *parent;
> > +     LIST_HEAD(head);
> >
> >       lockdep_assert_held(&enable_lock);
> >
> > -     if (!core)
> > -             return 0;
> > -
> > -     if (WARN(core->prepare_count == 0,
> > -         "Enabling unprepared %s\n", core->name))
> > -             return -ESHUTDOWN;
> > +     while (core) {
> > +             list_add(&core->enable_list, &head);
> > +             /* Stop once we see a clk that is already enabled */
> > +             if (core->enable_count)
> > +                     break;
> > +             core = core->parent;
> > +     }
> >
> > -     if (core->enable_count == 0) {
> > -             ret = clk_core_enable(core->parent);
> > +     list_for_each_entry_safe(core, tmp, &head, enable_list) {
> > +             list_del_init(&core->enable_list);
> >
> > -             if (ret)
> > -                     return ret;
> > +             if (WARN_ON(core->prepare_count == 0)) {
> > +                     ret = -ESHUTDOWN;
> > +                     goto err;
> > +             }
> >
> > -             trace_clk_enable_rcuidle(core);
> > +             if (core->enable_count == 0) {
> > +                     trace_clk_enable_rcuidle(core);
> >
> > -             if (core->ops->enable)
> > -                     ret = core->ops->enable(core->hw);
> > +                     if (core->ops->enable)
> > +                             ret = core->ops->enable(core->hw);
> >
> > -             trace_clk_enable_complete_rcuidle(core);
> > +                     trace_clk_enable_complete_rcuidle(core);
> >
> > -             if (ret) {
> > -                     clk_core_disable(core->parent);
> > -                     return ret;
> > +                     if (ret)
> > +                             goto err;
> >               }
> > +
> > +             core->enable_count++;
> >       }
> >
> > -     core->enable_count++;
> >       return 0;
> > +err:
> > +     parent = core->parent;
> > +     list_for_each_entry_safe_continue(core, tmp, &head, enable_list)
> > +             list_del_init(&core->enable_list);
> > +     clk_core_disable(parent);
> > +     return ret;
> >  }
> >
> >  static int clk_core_enable_lock(struct clk_core *core)
> > @@ -3281,6 +3294,8 @@ struct clk *clk_register(struct device *dev, struct clk_hw *hw)
> >       core->num_parents = hw->init->num_parents;
> >       core->min_rate = 0;
> >       core->max_rate = ULONG_MAX;
> > +     INIT_LIST_HEAD(&core->prepare_list);
> > +     INIT_LIST_HEAD(&core->enable_list);
> >       hw->core = core;
> >
> >       /* allocate local copy in case parent_names is __initdata */
>
>

Powered by blists - more mailing lists