On 08/01/2024 14:42, Suzuki K Poulose wrote:
Hi James
+Cc: Tao Zhang quic_taozha@quicinc.com +Cc: Mao Jinlong quic_jinlmao@quicinc.com
On 12/12/2023 15:54, James Clark wrote:
'enable', which probably should have been 'enabled', is only ever read in the core code in relation to controlling sources, and specifically only sources in sysfs mode. Confusingly it's not labelled as such and relying on it can be a source of bugs like the one fixed by commit 078dbba3f0c9 ("coresight: Fix crash when Perf and sysfs modes are used concurrently").
Most importantly, it can only be used when the coresight_mutex is held which is only done when enabling and disabling paths in sysfs mode, and not Perf mode.
I think we may be able to relax this a bit for the syfs. The sole reason for holding the mutex is for the "build_path" (and get_enabled_sink) where we need to make sure that no devices are removed/added. We hold necessary refcount on the device and the module (via coresight_grab_device()). After which, we should be able to release the mutex and perform the rest without it in coresight_enable()
After looking at the per-sink trace ID maps a bit more, I'm not sure if it will be worth the mental effort and risk to relax the sysfs locking right now.
We also currently have other things like writing to the global tracer_path which are outside of build_path/get_enabled_sink. But for the per-sink maps change we'll also have some tracking for sysfs mode about which sink map was used for each source and sink. And this state will be accessed across multiple sources, and after building the path, so it makes sense to leave the locking as-is for now IMO.
I also can't see a realistic gain from doing it, most sysfs use cases would be done from a single threaded script. Maybe in the future we could do the change to move the per-device locks into struct coresight_device, and then the core code can use them for things that need to be locked, but don't need the full coresight_mutex. And then that would also work for the per-sink case. But at the moment each device has its own lock so that's difficult.
So to prevent its usage spreading and leaking out to other devices, remove it.
It's use is equivalent to checking if the mode is currently sysfs, as due to the coresight_mutex lock, mode == CS_MODE_SYSFS can only become true or untrue when that lock is held, and when mode == CS_MODE_SYSFS the device is both enabled and in sysfs mode.
All of the above makes sense and looks good to me.
The one place it was used outside of the core code is in TPDA, but that pattern is more appropriately represented using refcounts inside the device's own spinlock.
But, I think we can clean up this code a bit more better. See below.
Signed-off-by: James Clark james.clark@arm.com
drivers/hwtracing/coresight/coresight-core.c | 86 +++++++------------- drivers/hwtracing/coresight/coresight-tpda.c | 12 ++- include/linux/coresight.h | 2 - 3 files changed, 38 insertions(+), 62 deletions(-)
diff --git a/drivers/hwtracing/coresight/coresight-core.c b/drivers/hwtracing/coresight/coresight-core.c index ab226441e5f4..1d0bd1586590 100644 --- a/drivers/hwtracing/coresight/coresight-core.c +++ b/drivers/hwtracing/coresight/coresight-core.c @@ -279,29 +279,18 @@ EXPORT_SYMBOL_GPL(coresight_add_helper); static int coresight_enable_sink(struct coresight_device *csdev, enum cs_mode mode, void *data) { - int ret = sink_ops(csdev)->enable(csdev, mode, data);
- if (ret) - return ret;
- csdev->enable = true;
- return 0; + return sink_ops(csdev)->enable(csdev, mode, data); } static void coresight_disable_sink(struct coresight_device *csdev) { - int ret = sink_ops(csdev)->disable(csdev); - if (ret) - return; - csdev->enable = false; + sink_ops(csdev)->disable(csdev); } static int coresight_enable_link(struct coresight_device *csdev, struct coresight_device *parent, struct coresight_device *child) { - int ret = 0; int link_subtype; struct coresight_connection *inconn, *outconn; @@ -317,19 +306,13 @@ static int coresight_enable_link(struct coresight_device *csdev, if (link_subtype == CORESIGHT_DEV_SUBTYPE_LINK_SPLIT && IS_ERR(outconn)) return PTR_ERR(outconn); - ret = link_ops(csdev)->enable(csdev, inconn, outconn); - if (!ret) - csdev->enable = true;
- return ret; + return link_ops(csdev)->enable(csdev, inconn, outconn); } static void coresight_disable_link(struct coresight_device *csdev, struct coresight_device *parent, struct coresight_device *child) { - int i; - int link_subtype; struct coresight_connection *inconn, *outconn; if (!parent || !child) @@ -337,26 +320,8 @@ static void coresight_disable_link(struct coresight_device *csdev, inconn = coresight_find_out_connection(parent, csdev); outconn = coresight_find_out_connection(csdev, child); - link_subtype = csdev->subtype.link_subtype; link_ops(csdev)->disable(csdev, inconn, outconn);
- if (link_subtype == CORESIGHT_DEV_SUBTYPE_LINK_MERG) { - for (i = 0; i < csdev->pdata->nr_inconns; i++) - if (atomic_read(&csdev->pdata->in_conns[i]->dest_refcnt) != - 0) - return; - } else if (link_subtype == CORESIGHT_DEV_SUBTYPE_LINK_SPLIT) { - for (i = 0; i < csdev->pdata->nr_outconns; i++) - if (atomic_read(&csdev->pdata->out_conns[i]->src_refcnt) != - 0) - return; - } else { - if (atomic_read(&csdev->refcnt) != 0) - return; - }
- csdev->enable = false; } int coresight_enable_source(struct coresight_device *csdev, enum cs_mode mode, @@ -364,11 +329,16 @@ int coresight_enable_source(struct coresight_device *csdev, enum cs_mode mode, { int ret; - if (!csdev->enable) { + /* + * Comparison with CS_MODE_SYSFS works without taking any device + * specific spinlock because the truthyness of that comparison can only + * change with coresight_mutex held, which we already have here. + */ + lockdep_assert_held(&coresight_mutex); + if (local_read(&csdev->mode) != CS_MODE_SYSFS) { ret = source_ops(csdev)->enable(csdev, data, mode); if (ret) return ret; - csdev->enable = true; } atomic_inc(&csdev->refcnt); @@ -385,22 +355,12 @@ static bool coresight_is_helper(struct coresight_device *csdev) static int coresight_enable_helper(struct coresight_device *csdev, enum cs_mode mode, void *data) { - int ret = helper_ops(csdev)->enable(csdev, mode, data);
- if (ret) - return ret;
- csdev->enable = true; - return 0; + return helper_ops(csdev)->enable(csdev, mode, data); } static void coresight_disable_helper(struct coresight_device *csdev) { - int ret = helper_ops(csdev)->disable(csdev, NULL);
- if (ret) - return; - csdev->enable = false; + helper_ops(csdev)->disable(csdev, NULL); } static void coresight_disable_helpers(struct coresight_device *csdev) @@ -445,11 +405,15 @@ EXPORT_SYMBOL_GPL(coresight_disable_source); static bool coresight_disable_source_sysfs(struct coresight_device *csdev, void *data) { + lockdep_assert_held(&coresight_mutex); + if (local_read(&csdev->mode) != CS_MODE_SYSFS) + return false;
if (atomic_dec_return(&csdev->refcnt) == 0) { coresight_disable_source(csdev, data); - csdev->enable = false; + return true; } - return !csdev->enable; + return false; } /* @@ -1097,7 +1061,13 @@ int coresight_enable(struct coresight_device *csdev) if (ret) goto out; - if (csdev->enable) { + /* + * mode == SYSFS implies that it's already enabled. Don't look at the + * refcount to determine this because we don't claim the source until + * coresight_enable_source() so can still race with Perf mode which + * doesn't hold coresight_mutex. + */ + if (local_read(&csdev->mode) == CS_MODE_SYSFS) { /* * There could be multiple applications driving the software * source. So keep the refcount for each such user when the @@ -1183,7 +1153,7 @@ void coresight_disable(struct coresight_device *csdev) if (ret) goto out; - if (!csdev->enable || !coresight_disable_source_sysfs(csdev, NULL)) + if (!coresight_disable_source_sysfs(csdev, NULL)) goto out; switch (csdev->subtype.source_subtype) { @@ -1249,7 +1219,9 @@ static ssize_t enable_source_show(struct device *dev, { struct coresight_device *csdev = to_coresight_device(dev); - return scnprintf(buf, PAGE_SIZE, "%u\n", csdev->enable); + guard(mutex)(&coresight_mutex); + return scnprintf(buf, PAGE_SIZE, "%u\n", + local_read(&csdev->mode) == CS_MODE_SYSFS); } static ssize_t enable_source_store(struct device *dev, diff --git a/drivers/hwtracing/coresight/coresight-tpda.c b/drivers/hwtracing/coresight/coresight-tpda.c index 5f82737c37bb..65c70995ab00 100644 --- a/drivers/hwtracing/coresight/coresight-tpda.c +++ b/drivers/hwtracing/coresight/coresight-tpda.c @@ -148,7 +148,11 @@ static int __tpda_enable(struct tpda_drvdata *drvdata, int port) CS_UNLOCK(drvdata->base); - if (!drvdata->csdev->enable) + /* + * Only do pre-port enable for first port that calls enable when the + * device's main refcount is still 0 + */ + if (!atomic_read(&drvdata->csdev->refcnt)) tpda_enable_pre_port(drvdata);
Relying on the "csdev->enable" to do pre-port configuration looks like a complete hack to me. This could have been performed once and for all during the probe time, at say, tpda_init_default_data(). That value is (drvdata->atid) never updated and need not be re-written unless the value is lost during a power idle.
Mao, Tao, are you able to confirm this ? If that is the case, we don't need this csdev->refcnt.
"unless the value is lost during a power idle."
That one might be a reason to do it. I suppose it makes sense to write all the registers on enable, but not to keep re-writing the trace ID again for every port that's used.
ret = tpda_enable_port(drvdata, port); @@ -169,6 +173,7 @@ static int tpda_enable(struct coresight_device *csdev, ret = __tpda_enable(drvdata, in->dest_port); if (!ret) { atomic_inc(&in->dest_refcnt); + atomic_inc(&csdev->refcnt); dev_dbg(drvdata->dev, "TPDA inport %d enabled.\n", in->dest_port); } } @@ -197,9 +202,10 @@ static void tpda_disable(struct coresight_device *csdev, struct tpda_drvdata *drvdata = dev_get_drvdata(csdev->dev.parent); spin_lock(&drvdata->spinlock); - if (atomic_dec_return(&in->dest_refcnt) == 0) + if (atomic_dec_return(&in->dest_refcnt) == 0) { __tpda_disable(drvdata, in->dest_port);
+ atomic_dec(&csdev->refcnt);
If we need this, then: This should be performed outside the if () isn't it ?
e.g.,
Operation: in_conn->refcnt csdev->refcnt port_enable : port0 0 - > 1 0 - > 1 port_enable : port0 1 - > 2 1 - > 2
I didn't get 1 -> 2 here, because the in->dest_refcnt is only incremented once if it was already 0 (although it is once for each port):
if (atomic_read(&in->dest_refcnt) == 0) { ret = __tpda_enable(drvdata, in->dest_port); if (!ret) { atomic_inc(&in->dest_refcnt);
That basically makes csdev->refcnt count how many ports were used, but each port's refcount can only be used once. And then when each port is disabled csdev->refcnt returns to zero on the last port.
port_disable: port0 2 - > 1 2 (no change) port_disable: port0 1 - > 0 2 - > 1
As you can see the csdev->refcnt skipped a dec.
Suzuki
+ } spin_unlock(&drvdata->spinlock); dev_dbg(drvdata->dev, "TPDA inport %d disabled\n", in->dest_port); diff --git a/include/linux/coresight.h b/include/linux/coresight.h index ba817f563ff7..46e6667f72ce 100644 --- a/include/linux/coresight.h +++ b/include/linux/coresight.h @@ -233,7 +233,6 @@ struct coresight_sysfs_link { * a non-atomic read would also work. * @refcnt: keep track of what is in use. * @orphan: true if the component has connections that haven't been linked.
- @enable: 'true' if component is currently part of an active path.
* @sysfs_sink_activated: 'true' when a sink has been selected for use via sysfs * by writing a 1 to the 'enable_sink' file. A sink can be * activated but not yet enabled. Enabling for a _sink_ happens @@ -260,7 +259,6 @@ struct coresight_device { local_t mode; atomic_t refcnt; bool orphan; - bool enable; /* sink specific fields */ bool sysfs_sink_activated; struct dev_ext_attribute *ea;