From patchwork Thu Aug 30 09:00:02 2012 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: MyungJoo Ham X-Patchwork-Id: 1387081 Return-Path: X-Original-To: patchwork-linux-pm@patchwork.kernel.org Delivered-To: patchwork-process-083081@patchwork1.kernel.org Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by patchwork1.kernel.org (Postfix) with ESMTP id 1CBB33FC33 for ; Thu, 30 Aug 2012 09:01:03 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753702Ab2H3JAp (ORCPT ); Thu, 30 Aug 2012 05:00:45 -0400 Received: from mailout3.samsung.com ([203.254.224.33]:36985 "EHLO mailout3.samsung.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752516Ab2H3I7w (ORCPT ); Thu, 30 Aug 2012 04:59:52 -0400 Received: from epcpsbgm2.samsung.com (epcpsbgm2 [203.254.230.27]) by mailout3.samsung.com (Oracle Communications Messaging Server 7u4-24.01(7.0.4.24.0) 64bit (built Nov 17 2011)) with ESMTP id <0M9K00MX78ZP75V0@mailout3.samsung.com>; Thu, 30 Aug 2012 17:59:50 +0900 (KST) X-AuditID: cbfee61b-b7faf6d00000476a-43-503f2b860644 Received: from epmmp2 ( [203.254.227.17]) by epcpsbgm2.samsung.com (EPCPMTA) with SMTP id AA.79.18282.68B2F305; Thu, 30 Aug 2012 17:59:50 +0900 (KST) Received: from localhost.localdomain ([10.90.51.65]) by mmp2.samsung.com (Oracle Communications Messaging Server 7u4-24.01 (7.0.4.24.0) 64bit (built Nov 17 2011)) with ESMTPA id <0M9K00KPQ8ZOWL10@mmp2.samsung.com>; Thu, 30 Aug 2012 17:59:50 +0900 (KST) From: MyungJoo Ham To: linux-pm@vger.kernel.org, linux-kernel@vger.kernel.org Cc: myungjoo.ham@gmail.com, rjw@sisk.pl, kyungmin.park@samsung.com, khilman@ti.com, markgross@thegnar.org, jean.pihet@newoldbits.com, mturquette@ti.com, jonghwa3.lee@samsung.com Subject: [PATCH v3 1/2] PM / devfreq: add global PM QoS support Date: Thu, 30 Aug 2012 18:00:02 +0900 Message-id: <1346317203-7357-2-git-send-email-myungjoo.ham@samsung.com> X-Mailer: git-send-email 1.7.4.1 In-reply-to: <1346317203-7357-1-git-send-email-myungjoo.ham@samsung.com> References: <1346317203-7357-1-git-send-email-myungjoo.ham@samsung.com> X-Brightmail-Tracker: H4sIAAAAAAAAA+NgFnrBJMWRmVeSWpSXmKPExsVy+t9jQd02bfsAg39TDC0u75rDZvG59wij A5PH501yAYxRXDYpqTmZZalF+nYJXBnvVnxmLFifUHFv726mBsYlvl2MnBwSAiYST2+tZ4ew xSQu3FvP1sXIxSEkMJ1RYvOTg0wQTguTxLnfN5lBqtgEdCW2brjLAmKLCFhJnP7fwQxSxCyw lVHi2okGVpCEsIC9xO67z8GKWARUJdpvTGcEsXkF3CWetHUyQqxTkFhw7y0biM0p4CHx//YG sF4hoJr/d88zTWDkXcDIsIpRNLUguaA4KT3XSK84Mbe4NC9dLzk/dxMj2P/PpHcwrmqwOMQo wMGoxMM7capdgBBrYllxZe4hRgkOZiURXnZ++wAh3pTEyqrUovz4otKc1OJDjNIcLErivE7n gKoF0hNLUrNTUwtSi2CyTBycUg2M9sLzV7BPkRMx+20hdbg2iFVEgS2ZT/WA237Jj6u0Hgvk xukyOC5sn+FuVCK3bvH5xZ9fn+pOPC5wzm7tmdn6nEKJCU/zC5JTjycv1uFz3tNznC374NKV x9dznd60SPILy91TVwo6n56UsS24Hf1FPXjFjdhTq0SbdXQueETs0kxYNO3s7wUuSizFGYmG WsxFxYkAaEOfwfsBAAA= Sender: linux-pm-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-pm@vger.kernel.org Even if the performance of a device is controlled properly with devfreq, sometimes, we still need to get PM-QoS inputs in order to meet the required performance. In our testbed of Exynos4412, which has on-chip various DMA devices, the memory interface and system bus are controlled according to their utilization by devfreq. However, in some multimedia applications including video-playing with MFC (multi-function codec) H/W and photo/video-capturing with FIMC H/W, we have observed issues due to insufficient DMA throughput or latency. In such applications, there are deadlines: less than 16.6ms with 60Hz. With shorter polling intervals (5 ~ 15ms), the frequencies fluctuate within a frame and we get missing frames and distorted pictures. With longer polling intervals (20 ~ 100ms), the response time is not sufficient and we get distorted or broken images. In other words, regardless of polling interval, we get poor results with hard-deadline H/Ws. They, in fact, have a preset requirement on DMA throughput. Thus, we need PM-QoS capabilities in devfreq. (Note that for general user applications, devfreq for bus/memory works fine. They are not so sensitive to tens of ms in performance increasing responses in general. In order to express how to handle QoS requests in devfreq devices, the devfreq device drivers only need to express the mappings of QoS value and frequency pairs with QoS class along with devfreq_add_device() call. As a side effect of the implementation, which happens to be positive, min/max freq is now enforced regardless of governor implementation. Tested on Exynos4412 machines with memory/bus frequencies and multimedia H/W blocks. (camera, video decoding, and video encoding) Signed-off-by: MyungJoo Ham Signed-off-by: Kyungmin Park --- Changed from V2-resend - Removed dependencies on global pm-qos class definitions - Revised data structure handling pm-qos (being ready for dev-pm-qos) Changes from V2 - Rebased Changes from V1 - Error handling at devfreq_add_device() - Handling pm_qos_max information - Styly update --- drivers/devfreq/devfreq.c | 127 +++++++++++++++++++++++++++++++++++++++++++-- include/linux/devfreq.h | 41 ++++++++++++++ 2 files changed, 164 insertions(+), 4 deletions(-) diff --git a/drivers/devfreq/devfreq.c b/drivers/devfreq/devfreq.c index 00e326c..d74b382 100644 --- a/drivers/devfreq/devfreq.c +++ b/drivers/devfreq/devfreq.c @@ -25,6 +25,7 @@ #include #include #include +#include #include "governor.h" static struct class *devfreq_class; @@ -136,8 +137,13 @@ int update_devfreq(struct devfreq *devfreq) * List from the highest proiority * max_freq (probably called by thermal when it's too hot) * min_freq + * qos_min_freq */ + if (devfreq->qos_min_freq && freq < devfreq->qos_min_freq) { + freq = devfreq->qos_min_freq; + flags &= ~DEVFREQ_FLAG_LEAST_UPPER_BOUND; /* Use GLB */ + } if (devfreq->min_freq && freq < devfreq->min_freq) { freq = devfreq->min_freq; flags &= ~DEVFREQ_FLAG_LEAST_UPPER_BOUND; /* Use GLB */ @@ -164,12 +170,12 @@ int update_devfreq(struct devfreq *devfreq) * devfreq_notifier_call() - Notify that the device frequency requirements * has been changed out of devfreq framework. * @nb the notifier_block (supposed to be devfreq->nb) - * @type not used + * @val not used. * @devp not used * * Called by a notifier that uses devfreq->nb. */ -static int devfreq_notifier_call(struct notifier_block *nb, unsigned long type, +static int devfreq_notifier_call(struct notifier_block *nb, unsigned long val, void *devp) { struct devfreq *devfreq = container_of(nb, struct devfreq, nb); @@ -183,6 +189,49 @@ static int devfreq_notifier_call(struct notifier_block *nb, unsigned long type, } /** + * devfreq_qos_notifier_call() - + */ +static int devfreq_qos_notifier_call(struct notifier_block *nb, + unsigned long value, void *devp) +{ + struct devfreq *devfreq = container_of(nb, struct devfreq, qos_nb); + int ret; + int i; + s32 default_value = PM_QOS_DEFAULT_VALUE; + struct devfreq_pm_qos_table *qos_list = devfreq->profile->qos_list; + bool qos_use_max = devfreq->profile->qos_use_max; + + if (!qos_list) + return NOTIFY_DONE; + + mutex_lock(&devfreq->lock); + + if (value == default_value) { + devfreq->qos_min_freq = 0; + goto update; + } + + for (i = 0; qos_list[i].freq; i++) { + /* QoS Met */ + if ((qos_use_max && qos_list[i].qos_value >= value) || + (!qos_use_max && qos_list[i].qos_value <= value)) { + devfreq->qos_min_freq = qos_list[i].freq; + goto update; + } + } + + /* Use the highest QoS freq */ + if (i > 0) + devfreq->qos_min_freq = qos_list[i - 1].freq; + +update: + ret = update_devfreq(devfreq); + mutex_unlock(&devfreq->lock); + + return ret; +} + +/** * _remove_devfreq() - Remove devfreq from the device. * @devfreq: the devfreq struct * @skip: skip calling device_unregister(). @@ -219,6 +268,10 @@ static void _remove_devfreq(struct devfreq *devfreq, bool skip) devfreq->being_removed = true; + if (devfreq->profile->qos_type) + pm_qos_remove_notifier(devfreq->profile->qos_type, + &devfreq->qos_nb); + if (devfreq->profile->exit) devfreq->profile->exit(devfreq->dev.parent); @@ -390,7 +443,7 @@ struct devfreq *devfreq_add_device(struct device *dev, void *data) { struct devfreq *devfreq; - int err = 0; + int err = 0, i; if (!dev || !profile || !governor) { dev_err(dev, "%s: Invalid parameters.\n", __func__); @@ -429,6 +482,61 @@ struct devfreq *devfreq_add_device(struct device *dev, devfreq->next_polling = devfreq->polling_jiffies = msecs_to_jiffies(devfreq->profile->polling_ms); devfreq->nb.notifier_call = devfreq_notifier_call; + devfreq->qos_nb.notifier_call = devfreq_qos_notifier_call; + + /* Check the sanity of qos_list/qos_type */ + if (profile->qos_type || profile->qos_list) { + + if (WARN(!profile->qos_type || !profile->qos_list, + "QoS requirement partially omitted for %s.\n", + dev_name(dev))) { + + err = -EINVAL; + goto err_dev; + } + + if (WARN(!profile->qos_list[0].freq, + "The first QoS requirement is the end of list for %s.\n", + dev_name(dev))) { + + err = -EINVAL; + goto err_dev; + } + + for (i = 1; profile->qos_list[i].freq; i++) { + if (WARN(profile->qos_list[i].freq <= + profile->qos_list[i - 1].freq, + "%s's qos_list[].freq not sorted in the ascending order. ([%d]=%lu, [%d]=%lu)\n", + dev_name(dev), i - 1, + profile->qos_list[i - 1].freq, i, + profile->qos_list[i].freq)) { + + err = -EINVAL; + goto err_dev; + } + + /* + * If QoS type is throughput(PM_QOS_MAX)-like, qos_value + * should be sorted in the ascending order. + * If QoS type is latency(PM_QOS_MIN)-like, qos_value + * should be sorted in the descending order. + */ + if (WARN((profile->qos_use_max && + profile->qos_list[i - 1].qos_value > + profile->qos_list[i].qos_value) || + (!profile->qos_use_max && + profile->qos_list[i - 1].qos_value < + profile->qos_list[i].qos_value), + "%s's qos_list[].qos_value is not sorted according to its QoS class.\n", + dev_name(dev))) { + + err = -EINVAL; + goto err_dev; + } + } + + pm_qos_add_notifier(profile->qos_type, &devfreq->qos_nb); + } devfreq->trans_table = devm_kzalloc(dev, sizeof(unsigned int) * devfreq->profile->max_state * @@ -443,7 +551,7 @@ struct devfreq *devfreq_add_device(struct device *dev, err = device_register(&devfreq->dev); if (err) { put_device(&devfreq->dev); - goto err_dev; + goto err_qos_add; } if (governor->init) @@ -471,6 +579,9 @@ out: err_init: device_unregister(&devfreq->dev); +err_qos_add: + if (profile->qos_type || profile->qos_list) + pm_qos_remove_notifier(profile->qos_type, &devfreq->qos_nb); err_dev: mutex_unlock(&devfreq->lock); kfree(devfreq); @@ -568,6 +679,13 @@ static ssize_t show_central_polling(struct device *dev, !to_devfreq(dev)->governor->no_central_polling); } +static ssize_t show_qos_min_freq(struct device *dev, + struct device_attribute *attr, + char *buf) +{ + return sprintf(buf, "%lu\n", to_devfreq(dev)->qos_min_freq); +} + static ssize_t store_min_freq(struct device *dev, struct device_attribute *attr, const char *buf, size_t count) { @@ -685,6 +803,7 @@ static struct device_attribute devfreq_attrs[] = { store_polling_interval), __ATTR(min_freq, S_IRUGO | S_IWUSR, show_min_freq, store_min_freq), __ATTR(max_freq, S_IRUGO | S_IWUSR, show_max_freq, store_max_freq), + __ATTR(qos_min_freq, S_IRUGO, show_qos_min_freq, NULL), __ATTR(trans_stat, S_IRUGO, show_trans_table, NULL), { }, }; diff --git a/include/linux/devfreq.h b/include/linux/devfreq.h index 30dc0d8..d11db17 100644 --- a/include/linux/devfreq.h +++ b/include/linux/devfreq.h @@ -53,6 +53,21 @@ struct devfreq_dev_status { #define DEVFREQ_FLAG_LEAST_UPPER_BOUND 0x1 /** + * struct devfreq_pm_qos_table - An PM QoS requiement entry for devfreq dev. + * @freq Lowest frequency to meet the QoS requirement + * represented by qos_value. If freq=0, it means that + * this element is the last in the array. + * @qos_value The qos value defined in pm_qos_params.h + * + * Note that the array of devfreq_pm_qos_table should be sorted by freq + * in the ascending order except for the last element, which should be 0. + */ +struct devfreq_pm_qos_table { + unsigned long freq; /* 0 if this is the last element */ + s32 qos_value; +}; + +/** * struct devfreq_dev_profile - Devfreq's user device profile * @initial_freq The operating frequency when devfreq_add_device() is * called. @@ -71,11 +86,33 @@ struct devfreq_dev_status { * from devfreq_remove_device() call. If the user * has registered devfreq->nb at a notifier-head, * this is the time to unregister it. + * @qos_type QoS Type (defined in pm_qos_params.h) + * 0 (PM_QOS_RESERVED) if not used. + * @qos_use_max True: the highest QoS value is used (for QoS + * requirement of throughput, bandwidth, or similar) + * False: the lowest QoS value is used (for QoS + * requirement of latency, delay, or similar) + * @enable_dev_pm_qos dev_pm_qos is enabled using the qos_list. + * @qos_list Array of QoS requirements ending with .freq = 0 + * NULL if not used. It should be either NULL or + * have a length > 1 with a first element effective. + * This QoS specification is shared by the global + * PM QoS (qos_type) and the per-dev PM QoS + * (enable_dev_pm_qos). + * + * Note that the array of qos_list should be sorted by freq + * in the ascending order. */ struct devfreq_dev_profile { unsigned long initial_freq; unsigned int polling_ms; + /* Optional QoS Handling Specification */ + int qos_type; /* 0: No global PM-QoS support */ + bool qos_use_max; /* true if "bandwidth"/"throughput"-like values */ + bool enable_dev_pm_qos; /* False: No per-dev PM-QoS support */ + struct devfreq_pm_qos_table *qos_list; /* QoS handling specification */ + int (*target)(struct device *dev, unsigned long *freq, u32 flags); int (*get_dev_status)(struct device *dev, struct devfreq_dev_status *stat); @@ -139,6 +176,8 @@ struct devfreq_governor { * order to prevent trying to remove the object multiple times. * @min_freq Limit minimum frequency requested by user (0: none) * @max_freq Limit maximum frequency requested by user (0: none) + * @qos_nb notifier block used to notify pm qos requests + * @qos_min_freq Limit minimum frequency requested by QoS * * This structure stores the devfreq information for a give device. * @@ -167,6 +206,8 @@ struct devfreq { unsigned long min_freq; unsigned long max_freq; + struct notifier_block qos_nb; + unsigned long qos_min_freq; /* information for device freqeuncy transition */ unsigned int total_trans;