Re: [PATCH v4 1/7] interconnect: Add generic on-chip interconnect API

[Evan Green]

Hi Georgi,

On Fri, Mar 9, 2018 at 1:12 PM Georgi Djakov <georgi.djakov@xxxxxxxxxx>
wrote:

> This patch introduce a new API to get requirements and configure the
> interconnect buses across the entire chipset to fit with the current
> demand.

> The API is using a consumer/provider-based model, where the providers are
> the interconnect buses and the consumers could be various drivers.
> The consumers request interconnect resources (path) between endpoints and
> set the desired constraints on this data flow path. The providers receive
> requests from consumers and aggregate these requests for all master-slave
> pairs on that path. Then the providers configure each participating in the
> topology node according to the requested data flow path, physical links
and
> constraints. The topology could be complicated and multi-tiered and is SoC
> specific.

> Signed-off-by: Georgi Djakov <georgi.djakov@xxxxxxxxxx>
> ---
>   Documentation/interconnect/interconnect.rst |  96 ++++++
>   drivers/Kconfig                             |   2 +
>   drivers/Makefile                            |   1 +
>   drivers/interconnect/Kconfig                |  10 +
>   drivers/interconnect/Makefile               |   1 +
>   drivers/interconnect/core.c                 | 489
++++++++++++++++++++++++++++
>   include/linux/interconnect-provider.h       | 109 +++++++
>   include/linux/interconnect.h                |  40 +++
>   8 files changed, 748 insertions(+)
>   create mode 100644 Documentation/interconnect/interconnect.rst
>   create mode 100644 drivers/interconnect/Kconfig
>   create mode 100644 drivers/interconnect/Makefile
>   create mode 100644 drivers/interconnect/core.c
>   create mode 100644 include/linux/interconnect-provider.h
>   create mode 100644 include/linux/interconnect.h

...
> diff --git a/drivers/interconnect/core.c b/drivers/interconnect/core.c
> new file mode 100644
> index 000000000000..6306e258b9b9
> --- /dev/null
> +++ b/drivers/interconnect/core.c
> @@ -0,0 +1,489 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * Interconnect framework core driver
> + *
> + * Copyright (c) 2018, Linaro Ltd.
> + * Author: Georgi Djakov <georgi.djakov@xxxxxxxxxx>
> + */
> +
> +#include <linux/device.h>
> +#include <linux/idr.h>
> +#include <linux/init.h>
> +#include <linux/interconnect.h>
> +#include <linux/interconnect-provider.h>
> +#include <linux/list.h>
> +#include <linux/module.h>
> +#include <linux/mutex.h>
> +#include <linux/slab.h>
> +
> +static DEFINE_IDR(icc_idr);
> +static LIST_HEAD(icc_provider_list);
> +static DEFINE_MUTEX(icc_provider_list_mutex);
> +static DEFINE_MUTEX(icc_path_mutex);
> +
> +/**
> + * struct icc_req - constraints that are attached to each node
> + *
> + * @req_node: entry in list of requests for the particular @node
> + * @node: the interconnect node to which this constraint applies
> + * @avg_bw: an integer describing the average bandwidth in kbps
> + * @peak_bw: an integer describing the peak bandwidth in kbps
> + */
> +struct icc_req {
> +       struct hlist_node req_node;
> +       struct icc_node *node;
> +       u32 avg_bw;
> +       u32 peak_bw;
> +};
> +
> +/**
> + * struct icc_path - interconnect path structure
> + * @num_nodes: number of hops (nodes)
> + * @reqs: array of the requests applicable to this path of nodes
> + */
> +struct icc_path {
> +       size_t num_nodes;
> +       struct icc_req reqs[0];
> +};
> +
> +static struct icc_node *node_find(const int id)
> +{
> +       struct icc_node *node;
> +
> +       node = idr_find(&icc_idr, id);
> +
> +       return node;
> +}
> +
> +static struct icc_path *path_allocate(struct icc_node *node, ssize_t
num_nodes)
> +{

So node is really the destination, correct? Then we use ->reverse to walk
backwards num_nodes steps towards the source. It might increase readability
to call the parameter dest, then assign that to a local called node for
traversal.

> +       struct icc_path *path;
> +       size_t i;
> +
> +       path = kzalloc(sizeof(*path) + num_nodes * sizeof(*path->reqs),
> +                      GFP_KERNEL);
> +       if (!path)
> +               return ERR_PTR(-ENOMEM);
> +
> +       path->num_nodes = num_nodes;
> +
> +       for (i = 0; i < num_nodes; i++) {
> +               hlist_add_head(&path->reqs[i].req_node, &node->req_list);
> +
> +               path->reqs[i].node = node;
> +               /* reference to previous node was saved during path
traversal */
> +               node = node->reverse;
> +       }
> +
> +       return path;
> +}
> +
> +static struct icc_path *path_find(struct icc_node *src, struct icc_node
*dst)
> +{
> +       struct icc_node *node = NULL;
> +       struct list_head traverse_list;
> +       struct list_head edge_list;
> +       struct list_head tmp_list;
> +       size_t i, number = 0;
> +       bool found = false;
> +
> +       INIT_LIST_HEAD(&traverse_list);
> +       INIT_LIST_HEAD(&edge_list);
> +       INIT_LIST_HEAD(&tmp_list);

tmp_list is really the list of nodes you've already visited and need to
remember to reset is_traversed for. Maybe calling this done_list or
visited_list would be more descriptive.

> +
> +       list_add_tail(&src->search_list, &traverse_list);

For added paranoia, you could set src->reverse to NULL so that somebody
elsewhere who had a bug in their back-traversal would fall off the end,
rather than into some previous scrapped path.

> +
> +       do {
> +               list_for_each_entry(node, &traverse_list, search_list) {
> +                       if (node == dst) {
> +                               found = true;
> +                               list_add(&node->search_list, &tmp_list);
> +                               break;
> +                       }
> +                       for (i = 0; i < node->num_links; i++) {
> +                               struct icc_node *tmp = node->links[i];
> +
> +                               if (!tmp)
> +                                       return ERR_PTR(-ENOENT);

You just bail out here, but never clean up the nodes is_traversed, which
will ruin later searches. Maybe a goto towards the common cleanup path?

> +
> +                               if (tmp->is_traversed)
> +                                       continue;
> +
> +                               tmp->is_traversed = true;
> +                               tmp->reverse = node;
> +                               list_add_tail(&tmp->search_list,
&edge_list);
> +                       }
> +               }
> +               if (found)
> +                       break;
> +
> +               list_splice_init(&traverse_list, &tmp_list);
> +               list_splice_init(&edge_list, &traverse_list);
> +
> +               /* count the number of nodes */
> +               number++;

Depth might be a better name for this, since this really counts the hops
away from the source, rather than the number of nodes you've processed.

> +
> +       } while (!list_empty(&traverse_list));
> +
> +       /* reset the traversed state */
> +       list_for_each_entry(node, &tmp_list, search_list)
> +               node->is_traversed = false;
> +
> +       if (found)
> +               return path_allocate(dst, number);
> +
> +       return ERR_PTR(-EPROBE_DEFER);
> +}
> +
> +static int path_init(struct icc_path *path)
> +{
> +       struct icc_node *node;
> +       size_t i;
> +
> +       for (i = 0; i < path->num_nodes; i++) {
> +               node = path->reqs[i].node;
> +
> +               mutex_lock(&node->provider->lock);
> +               node->provider->users++;
> +               mutex_unlock(&node->provider->lock);
> +       }
> +
> +       return 0;
> +}

This function cannot fail, nor do you check its return value, so you should
change the return type to void.

I'm wondering if the locking here is a little sketchy. I was in the process
of typing a suggestion that you call path_init from within path_find, since
it seemed weird to have this gray zone of a path without its reference
counts, when I noticed the locks.

I can't evaluate fully, since the implementation seems to be missing
icc_node_remove, a critical function in terms of evaluating the locks. You
have an icc_del_provider, but its warning of if (provider->users) is pretty
weak, since without node removal provider->users could easily be
incremented after the provider lock is released. It also leaks all of its
nodes, since there's no way to remove them.

Here's my suggestion as far as the locking goes:
* To add or remove links/nodes from the graph, you're going to need to hold
a global lock to avoid colliding with traversals. You've already got an
icc_path_mutex, so that would work.
* icc_link_create needs to hold the global icc_path_mutex, since it's
messing with arrays and connections used in path traversal, and doesn't
need to hold the provider lock, since it's not changing anything there.
* The presumably upcoming icc_link_destroy, or its parent icc_node_destroy,
also needs to hold the global lock. node_destroy may also need the provider
lock in symmetry with icc_node_add.
* Provider->users will be protected under the global icc_path_mutex, rather
than the provider lock. Then move path_init into path_find, or inline it
into path_allocate.
* Once you do that, provider->lock is now only protecting its node list.
For now, it's probably more efficient to roll the protection of
provider->nodes under the global lock as well, and remove the lock from the
provider altogether. If you anticipate other functions in the future that
will require a lock in the provider, then it might make sense to keep the
lock, or maybe just add it later with that new functionality.

> +
> +static void node_aggregate(struct icc_node *node)
> +{
> +       struct icc_req *r;
> +       u32 agg_avg = 0;
> +       u32 agg_peak = 0;
> +
> +       hlist_for_each_entry(r, &node->req_list, req_node) {
> +               /* sum(averages) and max(peaks) */
> +               agg_avg += r->avg_bw;
> +               agg_peak = max(agg_peak, r->peak_bw);
> +       }
> +
> +       node->avg_bw = agg_avg;
> +       node->peak_bw = agg_peak;
> +}
> +
> +static void provider_aggregate(struct icc_provider *provider, u32
*avg_bw,
> +                              u32 *peak_bw)
> +{
> +       struct icc_node *n;
> +       u32 agg_avg = 0;
> +       u32 agg_peak = 0;
> +
> +       /* aggregate for the interconnect provider */
> +       list_for_each_entry(n, &provider->nodes, node_list) {
> +               /* sum the average and max the peak */
> +               agg_avg += n->avg_bw;
> +               agg_peak = max(agg_peak, n->peak_bw);
> +       }
> +
> +       *avg_bw = agg_avg;
> +       *peak_bw = agg_peak;
> +}
> +
> +static int constraints_apply(struct icc_path *path)
> +{

Nit: maybe name it apply_constraints, since constraints_apply sounds like a
query (do the constraints apply?).

> +       struct icc_node *next, *prev = NULL;
> +       int i;
> +
> +       for (i = 0; i < path->num_nodes; i++, prev = next) {
> +               struct icc_provider *provider;
> +               u32 avg_bw = 0;
> +               u32 peak_bw = 0;
> +               int ret;
> +
> +               next = path->reqs[i].node;
> +               /*
> +                * Both endpoints should be valid master-slave pairs of
the
> +                * same interconnect provider that will be configured.
> +                */
> +               if (!next || !prev)
> +                       continue;
> +
> +               if (next->provider != prev->provider)
> +                       continue;

next should never be null, right? So you could shorten this to if (!prev ||
(next->provider != prev->provider))

> +
> +               provider = next->provider;
> +               mutex_lock(&provider->lock);
> +
> +               /* aggregate requests for the provider */
> +               provider_aggregate(provider, &avg_bw, &peak_bw);
> +
> +               if (provider->set) {
> +                       /* set the constraints */
> +                       ret = provider->set(prev, next, avg_bw, peak_bw);
> +               }
> +
> +               mutex_unlock(&provider->lock);
> +
> +               if (ret)
> +                       return ret;
> +       }
> +
> +       return 0;
> +}
> +
> +/**
> + * icc_set() - set constraints on an interconnect path between two
endpoints
> + * @path: reference to the path returned by icc_get()
> + * @avg_bw: average bandwidth in kbps
> + * @peak_bw: peak bandwidth in kbps
> + *
> + * This function is used by an interconnect consumer to express its own
needs
> + * in term of bandwidth and QoS for a previously requested path between
two

"in terms of" rather than "in term of", and not really QoS yet, right?

> + * endpoints. The requests are aggregated and each node is updated
accordingly.
> + *
> + * Returns 0 on success, or an approproate error code otherwise.

appropriate

> + */
> +int icc_set(struct icc_path *path, u32 avg_bw, u32 peak_bw)
> +{
> +       struct icc_node *node;
> +       size_t i;
> +       int ret;
> +
> +       if (!path)
> +               return 0;

Can we ditch this null check? My understanding is it's generally preferred
to skip this if it's only there to avoid developer errors.

> +
> +       for (i = 0; i < path->num_nodes; i++) {
> +               node = path->reqs[i].node;
> +
> +               mutex_lock(&icc_path_mutex);
> +
> +               /* update the consumer request for this path */
> +               path->reqs[i].avg_bw = avg_bw;
> +               path->reqs[i].peak_bw = peak_bw;
> +
> +               /* aggregate requests for this node */
> +               node_aggregate(node);
> +
> +               mutex_unlock(&icc_path_mutex);
> +       }
> +
> +       ret = constraints_apply(path);
> +       if (ret)
> +               pr_err("interconnect: error applying constraints (%d)",
ret);
> +
> +       return ret;
> +}
> +EXPORT_SYMBOL_GPL(icc_set);
> +
> +/**
> + * icc_get() - return a handle for path between two endpoints
> + * @src_id: source device port id
> + * @dst_id: destination device port id
> + *
> + * This function will search for a path between two endpoints and return
an
> + * icc_path handle on success. Use icc_put() to release
> + * constraints when the they are not needed anymore.
> + *
> + * Return: icc_path pointer on success, or ERR_PTR() on error
> + */
> +struct icc_path *icc_get(const int src_id, const int dst_id)
> +{
> +       struct icc_node *src, *dst;
> +       struct icc_path *path = ERR_PTR(-EPROBE_DEFER);
> +
> +       src = node_find(src_id);
> +       if (!src)
> +               goto out;
> +
> +       dst = node_find(dst_id);
> +       if (!dst)
> +               goto out;
> +
> +       mutex_lock(&icc_path_mutex);
> +       path = path_find(src, dst);
> +       mutex_unlock(&icc_path_mutex);
> +       if (IS_ERR(path))
> +               goto out;
> +
> +       path_init(path);
> +
> +out:
> +       return path;
> +}
> +EXPORT_SYMBOL_GPL(icc_get);
> +
> +/**
> + * icc_put() - release the reference to the icc_path
> + * @path: interconnect path
> + *
> + * Use this function to release the constraints on a path when the path
is
> + * no longer needed. The constraints will be re-aggregated.
> + */
> +void icc_put(struct icc_path *path)
> +{
> +       struct icc_node *node;
> +       size_t i;
> +       int ret;
> +
> +       if (!path || WARN_ON_ONCE(IS_ERR(path)))
> +               return;
> +
> +       ret = icc_set(path, 0, 0);
> +       if (ret)
> +               pr_err("%s: error (%d)\n", __func__, ret);
> +
> +       for (i = 0; i < path->num_nodes; i++) {
> +               node = path->reqs[i].node;
> +               hlist_del(&path->reqs[i].req_node);
> +
> +               mutex_lock(&node->provider->lock);
> +               node->provider->users--;
> +               mutex_unlock(&node->provider->lock);
> +       }
> +
> +       kfree(path);
> +}
> +EXPORT_SYMBOL_GPL(icc_put);
> +
> +/**
> + * icc_node_create() - create a node
> + * @id: node id
> + *
> + * Return: icc_node pointer on success, or ERR_PTR() on error
> + */
> +struct icc_node *icc_node_create(int id)
> +{
> +       struct icc_node *node;
> +
> +       /* check if node already exists */
> +       node = node_find(id);
> +       if (node)
> +               return node;

This is probably going to do more harm than good once icc_node_delete comes
in, since it almost certainly indicates a programmer error or ID collision,
and will likely result in a double free. We should probably fail with
EEXIST instead.

> +
> +       node = kzalloc(sizeof(*node), GFP_KERNEL);
> +       if (!node)
> +               return ERR_PTR(-ENOMEM);
> +
> +       id = idr_alloc(&icc_idr, node, id, id + 1, GFP_KERNEL);
> +       if (WARN(id < 0, "couldn't get idr"))
> +               return ERR_PTR(id);
> +
> +       node->id = id;
> +
> +       return node;
> +}
> +EXPORT_SYMBOL_GPL(icc_node_create);
> +
> +/**
> + * icc_link_create() - create a link between two nodes
> + * @src_id: source node id
> + * @dst_id: destination node id
> + *
> + * Return: 0 on success, or an error code otherwise
> + */
> +int icc_link_create(struct icc_node *node, const int dst_id)
> +{
> +       struct icc_node *dst;
> +       struct icc_node **new;
> +       int ret = 0;
> +
> +       if (IS_ERR_OR_NULL(node))
> +               return PTR_ERR(node);

Remove this.

> +
> +       mutex_lock(&node->provider->lock);
> +
> +       dst = node_find(dst_id);
> +       if (!dst)
> +               dst = icc_node_create(dst_id);

icc_node_create can fail, you should fail here if it does.

> +
> +       new = krealloc(node->links,
> +                      (node->num_links + 1) * sizeof(*node->links),
> +                      GFP_KERNEL);
> +       if (!new) {
> +               ret = -ENOMEM;
> +               goto out;
> +       }
> +
> +       node->links = new;
> +       node->links[node->num_links++] = dst;
> +
> +out:
> +       mutex_unlock(&node->provider->lock);
> +
> +       return 0;
> +}
> +EXPORT_SYMBOL_GPL(icc_link_create);
> +
> +/**
> + * icc_add_node() - add an interconnect node to interconnect provider
> + * @node: pointer to the interconnect node
> + * @provider: pointer to the interconnect provider
> + *
> + * Return: 0 on success, or an error code otherwise
> + */
> +int icc_node_add(struct icc_node *node, struct icc_provider *provider)
> +{
> +       if (WARN_ON(!node))
> +               return -EINVAL;
> +
> +       if (WARN_ON(!provider))
> +               return -EINVAL;

Remove these.

> +
> +       node->provider = provider;
> +
> +       mutex_lock(&provider->lock);
> +       list_add_tail(&node->node_list, &provider->nodes);
> +       mutex_unlock(&provider->lock);
> +
> +       return 0;
> +}

icc_node_add should be exported, right? I see it being used in msm8916.c.
You should make sure that "make allmodconfig" still builds with your
changes.

I think you should add a safety check in icc_link_create to ensure that the
node has a provider before adding any links. If some consumer made a
mistake and added links before adding the node to the provider, path
traversal would use the uninitialized or NULL provider pointer. I was
thinking about this while noticing that you assign node->provider before
acquiring the lock.

> +
> +/**
> + * icc_add_provider() - add a new interconnect provider
> + * @icc_provider: the interconnect provider that will be added into
topology
> + *
> + * Return: 0 on success, or an error code otherwise
> + */
> +int icc_add_provider(struct icc_provider *provider)
> +{
> +       if (WARN_ON(!provider))
> +               return -EINVAL;
> +

Remove this one. The one below is okay.

> +       if (WARN_ON(!provider->set))
> +               return -EINVAL;
> +
> +       mutex_init(&provider->lock);
> +       INIT_LIST_HEAD(&provider->nodes);
> +
> +       mutex_lock(&icc_provider_list_mutex);
> +       list_add(&provider->provider_list, &icc_provider_list);
> +       mutex_unlock(&icc_provider_list_mutex);
> +
> +       dev_dbg(provider->dev, "interconnect provider added to
topology\n");
> +
> +       return 0;
> +}
> +EXPORT_SYMBOL_GPL(icc_add_provider);
> +
> +/**
> + * icc_del_provider() - delete previously added interconnect provider
> + * @icc_provider: the interconnect provider that will be removed from
topology
> + *
> + * Return: 0 on success, or an error code otherwise
> + */
> +int icc_del_provider(struct icc_provider *provider)
> +{
> +       mutex_lock(&provider->lock);
> +       if (provider->users) {
> +               pr_warn("interconnect provider still has %d users\n",
> +                       provider->users);
> +       }
> +       mutex_unlock(&provider->lock);
> +
> +       mutex_lock(&icc_provider_list_mutex);
> +       list_del(&provider->provider_list);
> +       mutex_unlock(&icc_provider_list_mutex);
> +
> +       return 0;
> +}
> +EXPORT_SYMBOL_GPL(icc_del_provider);
> +
> +MODULE_AUTHOR("Georgi Djakov <georgi.djakov@xxxxxxxxxx");
> +MODULE_DESCRIPTION("Interconnect Driver Core");
> +MODULE_LICENSE("GPL v2");
> diff --git a/include/linux/interconnect-provider.h
b/include/linux/interconnect-provider.h
> new file mode 100644
> index 000000000000..779b5b5b1306
> --- /dev/null
> +++ b/include/linux/interconnect-provider.h
> @@ -0,0 +1,109 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +/*
> + * Copyright (c) 2018, Linaro Ltd.
> + * Author: Georgi Djakov <georgi.djakov@xxxxxxxxxx>
> + */
> +
> +#ifndef _LINUX_INTERCONNECT_PROVIDER_H
> +#define _LINUX_INTERCONNECT_PROVIDER_H
> +
> +#include <linux/interconnect.h>
> +
> +struct icc_node;
> +
> +/**
> + * struct icc_provider - interconnect provider (controller) entity that
might
> + * provide multiple interconnect controls
> + *
> + * @provider_list: list of the registered interconnect providers
> + * @nodes: internal list of the interconnect provider nodes
> + * @set: pointer to device specific set operation function
> + * @dev: the device this interconnect provider belongs to
> + * @lock: lock to provide consistency during aggregation/update of
constraints
> + * @users: count of active users
> + * @data: pointer to private data
> + */
> +struct icc_provider {
> +       struct list_head        provider_list;
> +       struct list_head        nodes;
> +       int (*set)(struct icc_node *src, struct icc_node *dst,
> +                  u32 avg_bw, u32 peak_bw);
> +       struct device           *dev;
> +       struct mutex            lock;
> +       int                     users;
> +       void                    *data;
> +};
> +
> +/**
> + * struct icc_node - entity that is part of the interconnect topology
> + *
> + * @id: platform specific node id
> + * @name: node name used in debugfs
> + * @links: a list of targets where we can go next when traversing
> + * @num_links: number of links to other interconnect nodes
> + * @provider: points to the interconnect provider of this node
> + * @node_list: list of interconnect nodes associated with @provider
> + * @search_list: list used when walking the nodes graph
> + * @reverse: pointer to previous node when walking the nodes graph
> + * @is_traversed: flag that is used when walking the nodes graph
> + * @req_list: a list of QoS constraint requests associated with this node
> + * @avg_bw: aggregated value of average bandwidth
> + * @peak_bw: aggregated value of peak bandwidth
> + * @data: pointer to private data
> + */
> +struct icc_node {
> +       int                     id;

Why int here? Are you expecting negative numbers? Maybe u32 instead? Or
even better, maybe a typedef u32 icc_id? Ooh yeah, that way we know when
parameters and such are passed around that they refer to this.

> +       const char              *name;
> +       struct icc_node         **links;
> +       size_t                  num_links;
> +
> +       struct icc_provider     *provider;
> +       struct list_head        node_list;

So the difference between node_list and links is that node_list nodes live
inside this node, whereas links point at other peers?

Oh no, I get it now after reading the .c file: node_list is the list entry
in the parent provider's "nodes" list. The comment description could be
clearer about that.

-Evan