From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-6.8 required=3.0 tests=DKIM_SIGNED,DKIM_VALID, DKIM_VALID_AU,HEADER_FROM_DIFFERENT_DOMAINS,INCLUDES_PATCH,MAILING_LIST_MULTI, SIGNED_OFF_BY,SPF_HELO_NONE,SPF_PASS autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 6569EC432C3 for ; Fri, 29 Nov 2019 21:25:33 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id 28F74217BA for ; Fri, 29 Nov 2019 21:25:33 +0000 (UTC) Authentication-Results: mail.kernel.org; dkim=pass (2048-bit key) header.d=linaro.org header.i=@linaro.org header.b="X2QQphmk" Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1727111AbfK2VZc (ORCPT ); Fri, 29 Nov 2019 16:25:32 -0500 Received: from mail-io1-f67.google.com ([209.85.166.67]:37232 "EHLO mail-io1-f67.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1727073AbfK2VZc (ORCPT ); Fri, 29 Nov 2019 16:25:32 -0500 Received: by mail-io1-f67.google.com with SMTP id k24so23120773ioc.4 for ; Fri, 29 Nov 2019 13:25:32 -0800 (PST) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linaro.org; s=google; h=mime-version:references:in-reply-to:from:date:message-id:subject:to :cc; bh=rQWXhLEHvXY0fOacTxBkCCu+SBMW2v/Wmdf1V/A/YNU=; b=X2QQphmk/ovbpXP2p/ACcYml+DNYCC0qh1VutHnuDWP+ygHjABDgNfFgfYysTlOyDV Ilh3I76sMi/N8IyACRcMSsJRoCtFf7ZEAlrYT0g7SfbsMCGHOqOG2jxR7Tmn1akOXa// vIerJI53FMkX1JugGdN+KFy2QF4ToSAARSwXC8oYshbETfbKip89dK5KTsrdUkmyjbuh 9oZII4kNa5Lp9Nx7R2y6vo7NsrlqmZA6j9Aa+50XgqX0ZbquZKq2VAvmqr7N4rzrs/wl QHdMysjCpAlGDOjJKGYpzofOyt4VOa6RaL/Nd7GsqHt5/iTGAQ6eNgUuX/HHLykNYYnR J9qw== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:mime-version:references:in-reply-to:from:date :message-id:subject:to:cc; bh=rQWXhLEHvXY0fOacTxBkCCu+SBMW2v/Wmdf1V/A/YNU=; b=QMiHxAf3J+oZBnVUaUr/BXxShMBVzUfVXnr5uYWlLhVTkbrD/uRsmpdFnuufkkEjoV FBStNf4RHhWZattBBHytOLnF+UC0RuCNcCGoT/OnQuEzOF1OdsdDv1PHLiCFLJmuVVlX qzUHwhfGZtm/tzsc1ykMPBrlFs9SY4oS6UAhkbig+SVlwqg0PGF34wivL/fyIhG8Xa4S 5sP7qx8SAEpikPHfBTYyOCa7+EU99EEr5xNUuk7eMG9qwMtlNU5G+A7o+CPzYHDqWWeI nVaT8C+7DZyCqeAKPGV7DYVlkg4aJEmAOYzSDqosDkWU3rfPBlWYN8DvmjJMjrwWdVM2 OP/w== X-Gm-Message-State: APjAAAVQ8wWjKyA223otf1tNdql5U6KYeQNqvgUid4fHbIGglOmDSAd3 kYUXW3RFU3waPLpZwbQOV+6ukDVlix401PWum1dySw== X-Google-Smtp-Source: APXvYqwR23S5it+fulG0l1d5hrTAFUe5YcVwda2R5nEcSJ/skggaAfJd21lD8TqoD9tVJ+OFQNyWFoB9joOxh6kdNug= X-Received: by 2002:a02:a08:: with SMTP id 8mr12145440jaw.98.1575062731367; Fri, 29 Nov 2019 13:25:31 -0800 (PST) MIME-Version: 1.0 References: <20191119231912.12768-1-mike.leach@linaro.org> <20191119231912.12768-9-mike.leach@linaro.org> In-Reply-To: From: Mathieu Poirier Date: Fri, 29 Nov 2019 14:25:20 -0700 Message-ID: Subject: Re: [PATCH v5 08/14] coresight: cti: Enable CTI associated with devices. To: Suzuki Kuruppassery Poulose Cc: Mike Leach , Coresight ML , linux-arm-kernel , devicetree@vger.kernel.org, "open list:DOCUMENTATION" Content-Type: text/plain; charset="UTF-8" Sender: devicetree-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: devicetree@vger.kernel.org On Fri, 29 Nov 2019 at 11:28, Suzuki Kuruppassery Poulose wrote: > > On 19/11/2019 23:19, Mike Leach wrote: > > The CoreSight subsystem enables a path of devices from source to sink. > > Any CTI devices associated with the path devices must be enabled at the > > same time. > > > > This patch adds an associated coresight_device element to the main > > coresight device structure, and uses this to create associations between > > the CTI and other devices based on the device tree data. The associated > > device element is used to enable CTI in conjunction with the path elements. > > > > CTI devices are reference counted so where a single CTI is associated with > > multiple elements on the path, it will be enabled on the first associated > > device enable, and disabled with the last associated device disable. > > > > Signed-off-by: Mike Leach > > --- > > drivers/hwtracing/coresight/coresight-cti.c | 87 +++++++++++++++++++ > > .../hwtracing/coresight/coresight-platform.c | 23 +++++ > > drivers/hwtracing/coresight/coresight-priv.h | 6 ++ > > drivers/hwtracing/coresight/coresight.c | 58 +++++++++++-- > > include/linux/coresight.h | 5 ++ > > 5 files changed, 173 insertions(+), 6 deletions(-) > > > > diff --git a/drivers/hwtracing/coresight/coresight-cti.c b/drivers/hwtracing/coresight/coresight-cti.c > > index 369488dd7b8e..cf116463149a 100644 > > --- a/drivers/hwtracing/coresight/coresight-cti.c > > +++ b/drivers/hwtracing/coresight/coresight-cti.c > > @@ -440,6 +440,90 @@ int cti_channel_setop(struct device *dev, enum cti_chan_set_op op, > > return err; > > } > > > > +/* > > + * Look for a matching connection device name in the list of > > + * connections. If found then swap in the csdev name and return > > + * found. > > + */ > > +static bool > > +cti_match_con_name(struct cti_device *ctidev, const char *node_name, > > + const char *csdev_name) > > Here we actually fixup the name of the connection, rather than simply > matching it. So it may be apt to rename this to cti_match_fixup_name() > > > +{ > > + struct cti_trig_con *trig_con; > > + > > + list_for_each_entry(trig_con, &ctidev->trig_cons, node) { > > + if (trig_con->con_dev_name) { > > + if (!strcmp(node_name, trig_con->con_dev_name)) { > > Can there be duplicate node_name's ? Does it make sense to store the > fwhandle along with the "temporary node_name" to match it later while > fixing up ? I think we're fine here - CS devices have bidirectional connections that are checked by the DTC. You get a bitter complaint if two CS nodes have the same value. > > > + /* match: so swap in csdev name */ > > + kfree(trig_con->con_dev_name); > > + trig_con->con_dev_name = > > + kstrdup(csdev_name, GFP_KERNEL); > > + return true; > > + } > > + } > > + } > > + return false; > > +} > > > > +/* > > + * Search the cti list to add an associated CTI into the supplied CS device > > + * This will set the association if CTI declared before the CS device > > + */ > > +void cti_add_assoc_to_csdev(struct coresight_device *csdev) > > +{ > > .. > > > + struct cti_drvdata *ect_item; > > + struct cti_device *ctidev; > > + const char *node_name = NULL, *csdev_name; > > + > > + /* protect the list */ > > + mutex_lock(&ect_mutex); > > + > > + /* exit if current is an ECT device.*/ > > + if ((csdev->type == CORESIGHT_DEV_TYPE_ECT) || list_empty(&ect_net)) > > + goto cti_add_done; > > + > > + /* if we didn't find the csdev previously we used the fwnode name */ > > + node_name = coresight_get_fwnode_name(csdev->dev.parent); > > We used "cti_plat_get_node_name()" when we added the name in the > absence of csdev in patch 7, could we not reuse the function here ? > > > + > > + if (!node_name) > > + goto cti_add_done; > > + > > + /* this is the name we want to use for the association */ > > + csdev_name = dev_name(&csdev->dev); > > > > + > > + /* for each CTI in list... */ > > + list_for_each_entry(ect_item, &ect_net, node) { > > + ctidev = &ect_item->ctidev; > > + if (cti_match_con_name(ctidev, node_name, csdev_name)) { > > + /* > > + * if we found a matching name then update the > > + * association pointers. > > + */ > > + csdev->ect_dev = ect_item->csdev; > > + goto cti_add_done; > > break; instead ? > > > + } > > + } > > +cti_add_done: > > + mutex_unlock(&ect_mutex); > > +} > > +EXPORT_SYMBOL_GPL(cti_add_assoc_to_csdev); > > + > > +/* > > + * Update the cross references where the associated device was found > > + * while we were building the connection info. This will occur if the > > + * assoc device was registered before the CTI. > > + */ > > +static void cti_update_conn_xrefs(struct cti_drvdata *drvdata) > > +{ > > + struct cti_trig_con *tc; > > + struct cti_device *ctidev = &drvdata->ctidev; > > + > > + list_for_each_entry(tc, &ctidev->trig_cons, node) { > > + if (tc->con_dev) > > + tc->con_dev->ect_dev = drvdata->csdev; > > + } > > Does this need to take the coresight_mutex to avoid racing against > a coresight_enable_path() ? Though this may be fine as long as the > CTI driver detects that that device was not enabled. > > Also, it looks like we have a potential issue with perf vs sysfs mode. > The perf mode doesn't seem to take the coresight_mutex, for > build_path/enable_path operations. This is outside the scope of this > series though. > > > +} > > + > > /** cti ect operations **/ > > int cti_enable(struct coresight_device *csdev) > > { > > @@ -574,6 +658,9 @@ static int cti_probe(struct amba_device *adev, const struct amba_id *id) > > drvdata->csdev_release = drvdata->csdev->dev.release; > > drvdata->csdev->dev.release = cti_device_release; > > > > + /* set any cross references */ > > + cti_update_conn_xrefs(drvdata); > > + > /* all done - dec pm refcount */ > > pm_runtime_put(&adev->dev); > > dev_info(&drvdata->csdev->dev, "CTI initialized\n"); > > diff --git a/drivers/hwtracing/coresight/coresight-platform.c b/drivers/hwtracing/coresight/coresight-platform.c > > index 3c5bee429105..6721cb1af5fe 100644 > > --- a/drivers/hwtracing/coresight/coresight-platform.c > > +++ b/drivers/hwtracing/coresight/coresight-platform.c > > @@ -293,6 +293,12 @@ static int of_get_coresight_platform_data(struct device *dev, > > > > return 0; > > } > > + > > +static inline const char *of_coresight_get_node_name(struct device *dev) > > +{ > > + return dev->of_node->full_name; > > +} > > + > > #else > > static inline int > > of_get_coresight_platform_data(struct device *dev, > > @@ -305,6 +311,11 @@ static inline int of_coresight_get_cpu(struct device *dev) > > { > > return -ENODEV; > > } > > + > > +static inline const char *of_coresight_get_node_name(struct device *dev) > > +{ > > + return NULL; > > +} > > #endif > > > > #ifdef CONFIG_ACPI > > @@ -766,6 +777,18 @@ static inline int acpi_coresight_get_cpu(struct device *dev) > > } > > #endif > > > > +const char *coresight_get_fwnode_name(struct device *dev) > > As mentioned above, please could we reuse the name helper we used > during the insertion rather than introducing a new wrapper which > effectively does the same thing ? > > > +{ > > + const char *node_name = NULL; > > + struct fwnode_handle *fwnode = dev_fwnode(dev); > > + > > + if (is_of_node(fwnode)) > > + node_name = of_coresight_get_node_name(dev); > > + > > + return node_name; > > +} > > +EXPORT_SYMBOL_GPL(coresight_get_fwnode_name); > > Why does this get exported ? If a following patch needs it, you may > always do that when you need it. > > > Cheers > Suzuki