CoreSight

coresight@lists.linaro.org

7 participants
2442 discussions

[PATCH v3 0/6] Arm CoreSight: Support AUX pause and resume

by Leo Yan

This series is to enable AUX pause and resume on Arm CoreSight. The first patch extracts the trace unit controlling operations to two functions. These two functions will be used by AUX pause and resume. Patches 02 and 03 change the ETMv4 driver to prepare callback functions for AUX pause and resume. Patch 04 changes the ETM perf layer to support AUX pause and resume in a perf session. The patches 05 updates buffer on AUX pause occasion, which can mitigate the trace data lose issue. Patch 07 documents the AUX pause usages with Arm CoreSight. This patch set has been verified on the Hikey960 board and TC platform. The previous one uses ETR and the later uses TRBE as sink. It is suggested to disable CPUIdle (add `nohlt` option in Linux command line) when verifying this series. ETM and funnel drivers are found issues during CPU suspend and resume which will be addressed separately. Changes from v2: - Rebased on CoreSight next branch. - Dropped the uAPI 'update_buf_on_pause' and updated document respectively (Suzuki). - Renamed ETM callbacks to .pause_perf() and .resume_perf() (Suzuki). - Minor improvement for error handling in the AUX resume flow. Changes from v1: - Added validation function pointers in pause and resume APIs (Mike). Leo Yan (6): coresight: etm4x: Extract the trace unit controlling coresight: Introduce pause and resume APIs for source coresight: etm4x: Hook pause and resume callbacks coresight: perf: Support AUX trace pause and resume coresight: perf: Update buffer on AUX pause Documentation: coresight: Document AUX pause and resume .../trace/coresight/coresight-perf.rst | 31 ++++ drivers/hwtracing/coresight/coresight-core.c | 22 +++ .../hwtracing/coresight/coresight-etm-perf.c | 84 +++++++++- .../coresight/coresight-etm4x-core.c | 143 +++++++++++++----- drivers/hwtracing/coresight/coresight-etm4x.h | 2 + drivers/hwtracing/coresight/coresight-priv.h | 2 + include/linux/coresight.h | 4 + 7 files changed, 246 insertions(+), 42 deletions(-) -- 2.34.1

4 months, 3 weeks

Re: [bug report] coresight: Dynamically add connections

by James Clark

On 02/04/2025 12:00 pm, Dan Carpenter wrote: > Hello James Clark, > > Commit 3d4ff657e454 ("coresight: Dynamically add connections") from > Apr 25, 2023 (linux-next), leads to the following Smatch static > checker warning: > > drivers/hwtracing/coresight/coresight-platform.c:279 of_coresight_parse_endpoint() > warn: inconsistent refcounting 'rep->kobj.kref.refcount.refs.counter': > > drivers/hwtracing/coresight/coresight-platform.c > 196 static int of_coresight_parse_endpoint(struct device *dev, > 197 struct device_node *ep, > 198 struct coresight_platform_data *pdata) > 199 { > 200 int ret = 0; > 201 struct of_endpoint endpoint, rendpoint; > 202 struct device_node *rparent = NULL; > 203 struct device_node *rep = NULL; > 204 struct device *rdev = NULL; > 205 struct fwnode_handle *rdev_fwnode; > 206 struct coresight_connection conn = {}; > 207 struct coresight_connection *new_conn; > 208 > 209 do { > 210 /* Parse the local port details */ > 211 if (of_graph_parse_endpoint(ep, &endpoint)) > 212 break; > 213 /* > 214 * Get a handle on the remote endpoint and the device it is > 215 * attached to. > 216 */ > 217 rep = of_graph_get_remote_endpoint(ep); > 218 if (!rep) > 219 break; > 220 rparent = of_coresight_get_port_parent(rep); > 221 if (!rparent) > 222 break; > 223 if (of_graph_parse_endpoint(rep, &rendpoint)) > 224 break; > 225 > 226 rdev_fwnode = of_fwnode_handle(rparent); > 227 /* If the remote device is not available, defer probing */ > 228 rdev = coresight_find_device_by_fwnode(rdev_fwnode); > 229 if (!rdev) { > 230 ret = -EPROBE_DEFER; > 231 break; > 232 } > 233 > 234 conn.src_port = endpoint.port; > 235 /* > 236 * Hold the refcount to the target device. This could be > 237 * released via: > 238 * 1) coresight_release_platform_data() if the probe fails or > 239 * this device is unregistered. > 240 * 2) While removing the target device via > 241 * coresight_remove_match() > 242 */ > 243 conn.dest_fwnode = fwnode_handle_get(rdev_fwnode); > 244 conn.dest_port = rendpoint.port; > 245 > 246 /* > 247 * Get the firmware node of the filter source through the > 248 * reference. This could be used to filter the source in > 249 * building path. > 250 */ > 251 conn.filter_src_fwnode = > 252 fwnode_find_reference(&ep->fwnode, "filter-source", 0); > 253 if (IS_ERR(conn.filter_src_fwnode)) { > 254 conn.filter_src_fwnode = NULL; > 255 } else { > 256 conn.filter_src_dev = > 257 coresight_find_csdev_by_fwnode(conn.filter_src_fwnode); > 258 if (conn.filter_src_dev && > 259 !coresight_is_device_source(conn.filter_src_dev)) { > 260 dev_warn(dev, "port %d: Filter handle is not a trace source : %s\n", > 261 conn.src_port, dev_name(&conn.filter_src_dev->dev)); > 262 conn.filter_src_dev = NULL; > 263 conn.filter_src_fwnode = NULL; > 264 } > 265 } > 266 > 267 new_conn = coresight_add_out_conn(dev, pdata, &conn); > 268 if (IS_ERR_VALUE(new_conn)) { > 269 fwnode_handle_put(conn.dest_fwnode); > 270 return PTR_ERR(new_conn); > > Should this be a break statement instead of return statement? > Hi Dan, Yes it probably should, I'll send a patch. Thanks James

4 months, 3 weeks

Re: [PATCH v1 9/9] coresight: Consolidate clock enabling

by Leo Yan

Hi Jie, [ + Rob ] On Wed, Apr 02, 2025 at 08:55:51AM +0800, Jie Gan wrote: [...] > > > { > > > - struct clk *pclk = NULL; > > > + WARN_ON(!pclk); > > > if (!dev_is_amba(dev)) { > > > - pclk = devm_clk_get_enabled(dev, "apb_pclk"); > > > - if (IS_ERR(pclk)) > > > - pclk = devm_clk_get_enabled(dev, "apb"); > > > + *pclk = devm_clk_get_enabled(dev, "apb_pclk"); > > > + if (IS_ERR(*pclk)) > > > + *pclk = devm_clk_get_enabled(dev, "apb"); > > > + if (IS_ERR(*pclk)) > > > + return PTR_ERR(*pclk); > > > + } else { > > > + /* Don't enable pclk for an AMBA device */ > > > + *pclk = NULL; > > > > Now the "apb" clock won't be enabled for amba devices. I'm assuming > > that's fine if the clock was always called "apb_pclk" for them, but the > > commit that added the new clock name didn't specify any special casing > > either. > > > > Can we have a comment that says it's deliberate? But the more I think > > about it the more I'm confused why CTCU needed a different clock name to > > be defined, when all the other Coresight devices use "apb_pclk". > > Hi James, > > The original clock-name for CTCU is apb_pclk, but the dt-binding maintainer > request me to change it to apb, that's why the clock name is different from > others. > > I am not why we need apb instead of apb_pclk in dt-binding. Maybe some rules > have changed for dt-binding requirement. My conclusion is that if a device is an Arm Primecell peripheral, it should use the clock name "apb_pclk" (See the DT binding doc [1]). CTCU is not an Arm Primecell peripheral, so it does not need to strictly follow up the clock naming for Primecell peripheral. In Arm CoreSight framework, for code consistency, I would suggest using the clock naming "apb_pclk" for the APB clock for a newly added device that even it is not a Primecell peripheral. (We don't need to make any change to the CTCU driver, as we need to remain compatible with existed DTB blobs). Cc'ing Rob in case he has any suggestions. Thanks, Leo [1] Documentation/devicetree/bindings/arm/primecell.yaml

4 months, 3 weeks

Re: [PATCH] coresight: fix the wrong type of the trace_id in coresight_path

by Mike Leach

Hi, On Wed, 2 Apr 2025 at 01:50, Jie Gan <quic_jiegan(a)quicinc.com> wrote: > > > > On 4/1/2025 5:56 PM, Mike Leach wrote: > > Hi, > > > > On Tue, 1 Apr 2025 at 07:11, Anshuman Khandual > > <anshuman.khandual(a)arm.com> wrote: > >> > >> On 4/1/25 07:12, Jie Gan wrote: > >>> The trace_id in coresight_path may contain an error number which means a > >>> negative integer, but the current type of the trace_id is u8. Change the > >>> type to int to fix it. > >>> > >>> Reported-by: Dan Carpenter <dan.carpenter(a)linaro.org> > >>> Fixes: 3c03c49b2fa5 ("Coresight: Introduce a new struct coresight_path") > >>> Signed-off-by: Jie Gan <jie.gan(a)oss.qualcomm.com> > >> > >> LGTM > >> > >> Reviewed-by: Anshuman Khandual <anshuman.khandual(a)arm.com> > >> > >>> --- > >>> include/linux/coresight.h | 2 +- > >>> 1 file changed, 1 insertion(+), 1 deletion(-) > >>> > >>> diff --git a/include/linux/coresight.h b/include/linux/coresight.h > >>> index d79a242b271d..c2bf10c43e7c 100644 > >>> --- a/include/linux/coresight.h > >>> +++ b/include/linux/coresight.h > >>> @@ -337,7 +337,7 @@ static struct coresight_dev_list (var) = { \ > >>> */ > >>> struct coresight_path { > >>> struct list_head path_list; > >>> - u8 trace_id; > >>> + int trace_id; > >>> }; > >>> > >>> enum cs_mode { > > > > There are many places in the Coresight drivers that assign a u8 > > traceid from the path trace ID. > > > > e.g. > > In coresight-etm4x-core.c : etm4_enable_perf() > > > > drvdata->trcid = path->trace_id; > > > > drvdata->trcid is defined as a u8 - the reason being trace IDs are > > 128 bits wide with some reserved values. > > > > Will this not just trigger the same issue if path->trace_id is changed > > to an int? Even if not it is inconsistent handling of the trace ID > > values. > > > > Trace ID errors should be handled by returning an invalid trace ID > > value - were the trace ID value will fail the macro > > IS_VALID_CS_TRACE_ID(), or separate the return of a trace ID from an > > error return in a function. > > > > Hi Mike, > > The path->trace_id is verified after it has been assigned with the logic > you mentioned: > > if (!IS_VALID_CS_TRACE_ID(path->trace_id)) > goto err_path; > > So it should be safe to assign to another u8 parameter, like you mentioned: > > In coresight-etm4x-core.c : etm4_enable_perf() > > drvdata->trcid = path->trace_id; > It is safe but will it not trigger a warning just like the one you are trying to fix as the types are mismatched? Mike > Thanks, > Jie > > > > Regards > > > > Mike > > > > > > > > -- > > Mike Leach > > Principal Engineer, ARM Ltd. > > Manchester Design Centre. UK > -- Mike Leach Principal Engineer, ARM Ltd. Manchester Design Centre. UK

4 months, 3 weeks

Re: [PATCH] coresight: catu: Introduce refcount and spinlock for enabling/disabling

by James Clark

On 28/03/2025 10:38 pm, Yabin Cui wrote: > When tracing ETM data on multiple CPUs concurrently via the > perf interface, the CATU device is shared across different CPU > paths. This can lead to race conditions when multiple CPUs attempt > to enable or disable the CATU device simultaneously. > > To address these race conditions, this patch introduces the > following changes: > > 1. The enable and disable operations for the CATU device are not > reentrant. Therefore, a spinlock is added to ensure that only > one CPU can enable or disable a given CATU device at any point > in time. > > 2. A reference counter is used to manage the enable/disable state > of the CATU device. The device is enabled when the first CPU > requires it and is only disabled when the last CPU finishes > using it. This ensures the device remains active as long as at > least one CPU needs it. > > Signed-off-by: Yabin Cui <yabinc(a)google.com> > --- > drivers/hwtracing/coresight/coresight-catu.c | 29 ++++++++++++++------ > drivers/hwtracing/coresight/coresight-catu.h | 1 + > 2 files changed, 22 insertions(+), 8 deletions(-) > > diff --git a/drivers/hwtracing/coresight/coresight-catu.c b/drivers/hwtracing/coresight/coresight-catu.c > index 275cc0d9f505..834a7ffbbdbc 100644 > --- a/drivers/hwtracing/coresight/coresight-catu.c > +++ b/drivers/hwtracing/coresight/coresight-catu.c > @@ -458,12 +458,19 @@ static int catu_enable_hw(struct catu_drvdata *drvdata, enum cs_mode cs_mode, > static int catu_enable(struct coresight_device *csdev, enum cs_mode mode, > void *data) > { > - int rc; > + int rc = 0; > + unsigned long flags; > struct catu_drvdata *catu_drvdata = csdev_to_catu_drvdata(csdev); > > - CS_UNLOCK(catu_drvdata->base); > - rc = catu_enable_hw(catu_drvdata, mode, data); > - CS_LOCK(catu_drvdata->base); > + spin_lock_irqsave(&catu_drvdata->spinlock, flags); Hi Yabin, This needs to be a raw_spinlock since [1]. Also you might as well use the new guard() thing to save someone find-and-replacing it later. But I'm wondering if this is accurate. The ETR's refcount is dependent on the pid of the owner of the trace session: /* Do not proceed if this device is associated with another session */ if (drvdata->pid != -1 && drvdata->pid != pid) { rc = -EBUSY; goto unlock_out; } /* * No HW configuration is needed if the sink is already in * use for this session. */ if (drvdata->pid == pid) { csdev->refcnt++; goto unlock_out; } If the helpers get enabled first, could this mean that CATU gets associated with a different session than the ETR? Maybe not, but it would be easier to understand if the core code handled the refcounting and locking for linked devices. [1]: https://lore.kernel.org/all/20250306121110.1647948-3-yeoreum.yun@arm.com/ Thanks James > + if (csdev->refcnt == 0) { > + CS_UNLOCK(catu_drvdata->base); > + rc = catu_enable_hw(catu_drvdata, mode, data); > + CS_LOCK(catu_drvdata->base); > + } > + if (!rc) > + csdev->refcnt++; > + spin_unlock_irqrestore(&catu_drvdata->spinlock, flags); > return rc; > } > > @@ -486,12 +493,17 @@ static int catu_disable_hw(struct catu_drvdata *drvdata) > > static int catu_disable(struct coresight_device *csdev, void *__unused) > { > - int rc; > + int rc = 0; > + unsigned long flags; > struct catu_drvdata *catu_drvdata = csdev_to_catu_drvdata(csdev); > > - CS_UNLOCK(catu_drvdata->base); > - rc = catu_disable_hw(catu_drvdata); > - CS_LOCK(catu_drvdata->base); > + spin_lock_irqsave(&catu_drvdata->spinlock, flags); > + if (--csdev->refcnt == 0) { > + CS_UNLOCK(catu_drvdata->base); > + rc = catu_disable_hw(catu_drvdata); > + CS_LOCK(catu_drvdata->base); > + } > + spin_unlock_irqrestore(&catu_drvdata->spinlock, flags); > return rc; > } > > @@ -550,6 +562,7 @@ static int __catu_probe(struct device *dev, struct resource *res) > dev->platform_data = pdata; > > drvdata->base = base; > + spin_lock_init(&drvdata->spinlock); > catu_desc.access = CSDEV_ACCESS_IOMEM(base); > catu_desc.pdata = pdata; > catu_desc.dev = dev; > diff --git a/drivers/hwtracing/coresight/coresight-catu.h b/drivers/hwtracing/coresight/coresight-catu.h > index 141feac1c14b..663282ec6381 100644 > --- a/drivers/hwtracing/coresight/coresight-catu.h > +++ b/drivers/hwtracing/coresight/coresight-catu.h > @@ -65,6 +65,7 @@ struct catu_drvdata { > void __iomem *base; > struct coresight_device *csdev; > int irq; > + spinlock_t spinlock; > }; > > #define CATU_REG32(name, offset) \

4 months, 3 weeks

Re: [PATCH] coresight: fix the wrong type of the trace_id in coresight_path

by Anshuman Khandual

On 4/1/25 07:12, Jie Gan wrote: > The trace_id in coresight_path may contain an error number which means a > negative integer, but the current type of the trace_id is u8. Change the > type to int to fix it. > > Reported-by: Dan Carpenter <dan.carpenter(a)linaro.org> > Fixes: 3c03c49b2fa5 ("Coresight: Introduce a new struct coresight_path") > Signed-off-by: Jie Gan <jie.gan(a)oss.qualcomm.com> LGTM Reviewed-by: Anshuman Khandual <anshuman.khandual(a)arm.com> > --- > include/linux/coresight.h | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/include/linux/coresight.h b/include/linux/coresight.h > index d79a242b271d..c2bf10c43e7c 100644 > --- a/include/linux/coresight.h > +++ b/include/linux/coresight.h > @@ -337,7 +337,7 @@ static struct coresight_dev_list (var) = { \ > */ > struct coresight_path { > struct list_head path_list; > - u8 trace_id; > + int trace_id; > }; > > enum cs_mode {

4 months, 3 weeks

Re: [PATCH v1 1/1] coresight: cti: Replace inclusion by struct fwnode_handle forward declaration

by James Clark

On 31/03/2025 8:14 am, Andy Shevchenko wrote: > The fwnode.h is not supposed to be used by the drivers as it > has the definitions for the core parts for different device > property provider implementations. Drop it. > > Since the code wants to use the pointer to the struct fwnode_handle > the forward declaration is provided. > > Signed-off-by: Andy Shevchenko <andriy.shevchenko(a)linux.intel.com> > --- > drivers/hwtracing/coresight/coresight-cti.h | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/drivers/hwtracing/coresight/coresight-cti.h b/drivers/hwtracing/coresight/coresight-cti.h > index 16e310e7e9d4..8362a47c939c 100644 > --- a/drivers/hwtracing/coresight/coresight-cti.h > +++ b/drivers/hwtracing/coresight/coresight-cti.h > @@ -9,7 +9,6 @@ > > #include <linux/coresight.h> > #include <linux/device.h> > -#include <linux/fwnode.h> > #include <linux/list.h> > #include <linux/spinlock.h> > #include <linux/sysfs.h> > @@ -17,6 +16,8 @@ > > #include "coresight-priv.h" > > +struct fwnode_handle; > + > /* > * Device registers > * 0x000 - 0x144: CTI programming and status Reviewed-by: James Clark <james.clark(a)linaro.org>

4 months, 3 weeks

[PATCH V3 0/2] Fix Coresight instruction synthesis logic

by Tanmay Jagdale

When we use perf to catpure Coresight trace and generate instruction trace using 'perf script', we get the following output: # perf record -e cs_etm/@tmc_etr0/ -C 9 taskset -c 9 sleep 1 # perf script --itrace=i1ns --ns -Fcomm,tid,pid,time,cpu,event,ip,sym,addr,symoff,flags,callindent .. perf 9024/9024 [009] 2690.650470551: instructions: call 0 ffffb305591aed54 coresight_timeout+0x28 perf 9024/9024 [009] 2690.650470551: instructions: call 0 ffffb305591aed58 coresight_timeout+0x2c perf 9024/9024 [009] 2690.650470551: instructions: call 0 ffffb305591aed5c coresight_timeout+0x30 perf 9024/9024 [009] 2690.650470551: instructions: call 0 ffffb305591aed60 coresight_timeout+0x34 perf 9024/9024 [009] 2690.650470551: instructions: jmp 0 ffffb305591aed7c coresight_timeout+0x50 perf 9024/9024 [009] 2690.650470551: instructions: jmp 0 ffffb305591aed80 coresight_timeout+0x54 perf 9024/9024 [009] 2690.650470551: instructions: jmp 0 ffffb305591aed84 coresight_timeout+0x58 perf 9024/9024 [009] 2690.650470552: instructions: jcc 0 ffffb305591aede4 coresight_timeout+0xb8 perf 9024/9024 [009] 2690.650470552: instructions: jcc 0 ffffb305591aede8 coresight_timeout+0xbc perf 9024/9024 [009] 2690.650470552: instructions: jcc 0 ffffb305591aedec coresight_timeout+0xc0 perf 9024/9024 [009] 2690.650470552: instructions: jcc 0 ffffb305591aedf0 coresight_timeout+0xc4 perf 9024/9024 [009] 2690.650470557: instructions: call 0 ffffb305591bccec ete_sysreg_read+0x0 perf 9024/9024 [009] 2690.650470557: instructions: call 0 ffffb305591bccf0 ete_sysreg_read+0x4 perf 9024/9024 [009] 2690.650470557: instructions: call 0 ffffb305591bccf4 ete_sysreg_read+0x8 perf 9024/9024 [009] 2690.650470557: instructions: call 0 ffffb305591bccf8 ete_sysreg_read+0xc perf 9024/9024 [009] 2690.650470557: instructions: call 0 ffffb305591bccfc ete_sysreg_read+0x10 perf 9024/9024 [009] 2690.650470557: instructions: call 0 ffffb305591bcd00 ete_sysreg_read+0x14 This output has the following issues: 1. Non-branch instructions have mnemonics of branch instructions (Column 6) 2. Branch target address is missing (Column 7) This patch fixes these issues by changing the logic of instruction syntehsis for the Coresight trace queues. Output after applying the patch: ... perf 6111/6111 [008] 457.332794461: instructions: 0 ffffb305591aed54 coresight_timeout+0x28 perf 6111/6111 [008] 457.332794461: instructions: 0 ffffb305591aed58 coresight_timeout+0x2c perf 6111/6111 [008] 457.332794461: instructions: 0 ffffb305591aed5c coresight_timeout+0x30 perf 6111/6111 [008] 457.332794461: instructions: jmp ffffb305591aed7c ffffb305591aed60 coresight_timeout+0x34 perf 6111/6111 [008] 457.332794461: instructions: 0 ffffb305591aed7c coresight_timeout+0x50 perf 6111/6111 [008] 457.332794461: instructions: 0 ffffb305591aed80 coresight_timeout+0x54 perf 6111/6111 [008] 457.332794461: instructions: jcc ffffb305591aede4 ffffb305591aed84 coresight_timeout+0x58 perf 6111/6111 [008] 457.332794462: instructions: 0 ffffb305591aede4 coresight_timeout+0xb8 perf 6111/6111 [008] 457.332794462: instructions: 0 ffffb305591aede8 coresight_timeout+0xbc perf 6111/6111 [008] 457.332794462: instructions: 0 ffffb305591aedec coresight_timeout+0xc0 perf 6111/6111 [008] 457.332794462: instructions: call ffffb305591bccec ffffb305591aedf0 coresight_timeout+0xc4 perf 6111/6111 [008] 457.332794462: instructions: 0 ffffb305591bccec ete_sysreg_read+0x0 perf 6111/6111 [008] 457.332794462: instructions: 0 ffffb305591bccf0 ete_sysreg_read+0x4 perf 6111/6111 [008] 457.332794462: instructions: 0 ffffb305591bccf4 ete_sysreg_read+0x8 perf 6111/6111 [008] 457.332794462: instructions: 0 ffffb305591bccf8 ete_sysreg_read+0xc perf 6111/6111 [008] 457.332794462: instructions: 0 ffffb305591bccfc ete_sysreg_read+0x10 perf 6111/6111 [008] 457.332794462: instructions: 0 ffffb305591bcd00 ete_sysreg_read+0x14 Changes in V3 - Rebased to latest coresight-next branch - Added Reviewed-by tags from V2 [2] Changes in V2 - Updated commit message of Patch 1 - As discussed in the previous version [1], there were differences in instruction trace output before and after the patch. The timestamps for the instructions were not in sync. Added a patch 2 which fixes this issue. Changes in V1 - https://lkml.org/lkml/2023/6/23/912 [1] https://lkml.org/lkml/2023/6/28/506 [2] https://lore.kernel.org/lkml/b2c02eb9-0940-4770-a4b7-22d2af8078db@arm.com/ Tanmay Jagdale (2): perf: cs-etm: Fixes in instruction sample synthesis perf: cs-etm: Store previous timestamp in packet queue tools/perf/util/cs-etm.c | 49 +++++++++++++++++++++++++++++++--------- tools/perf/util/cs-etm.h | 1 + 2 files changed, 39 insertions(+), 11 deletions(-) -- 2.43.0

5 months

Re: [PATCH v4] coresight: prevent deactivate active config while enabling the config

by Leo Yan

Hi Levi, On Mon, Mar 24, 2025 at 07:17:40PM +0000, Yeoreum Yun wrote: > While enable active config via cscfg_csdev_enable_active_config(), > active config could be deactivated via configfs' sysfs interface. > This could make UAF issue in below scenario: > > CPU0 CPU1 > (sysfs enable) load module > cscfg_load_config_sets() > activate config. // sysfs > (sys_active_cnt == 1) > ... > cscfg_csdev_enable_active_config() > lock(csdev->cscfg_csdev_lock) > // here load config activate by CPU1 > unlock(csdev->cscfg_csdev_lock) > > deactivate config // sysfs > (sys_activec_cnt == 0) > cscfg_unload_config_sets() > unload module > > // access to config_desc which freed > // while unloading module. > cfs_csdev_enable_config I am not sure if this flow can happen. CoreSight configfs feature is integrated into the CoreSight core layer, and the other CoreSight modules are dependent on it. For example, if the ETM4x module is not removed, the kernel module management will natually prevent the CoreSight core module from being removed. > To address this, use cscfg_config_desc's active_cnt as a reference count > which will be holded when > - activate the config. > - enable the activated config. > and put the module reference when config_active_cnt == 0. > > Signed-off-by: Yeoreum Yun <yeoreum.yun(a)arm.com> > --- > Since v3: > - Remove enable arguments in cscfg_config_desc_get() (from Mike). > - https://lore.kernel.org/all/20250109171956.3535294-1-yeoreum.yun@arm.com/ > --- > .../hwtracing/coresight/coresight-config.h | 2 +- > .../coresight/coresight-etm4x-core.c | 3 ++ > .../hwtracing/coresight/coresight-syscfg.c | 52 +++++++++++++------ > 3 files changed, 41 insertions(+), 16 deletions(-) > > diff --git a/drivers/hwtracing/coresight/coresight-config.h b/drivers/hwtracing/coresight/coresight-config.h > index b9ebc9fcfb7f..90fd937d3bd8 100644 > --- a/drivers/hwtracing/coresight/coresight-config.h > +++ b/drivers/hwtracing/coresight/coresight-config.h > @@ -228,7 +228,7 @@ struct cscfg_feature_csdev { > * @feats_csdev:references to the device features to enable. > */ > struct cscfg_config_csdev { > - const struct cscfg_config_desc *config_desc; > + struct cscfg_config_desc *config_desc; > struct coresight_device *csdev; > bool enabled; > struct list_head node; > diff --git a/drivers/hwtracing/coresight/coresight-etm4x-core.c b/drivers/hwtracing/coresight/coresight-etm4x-core.c > index e5972f16abff..ef96028fa56b 100644 > --- a/drivers/hwtracing/coresight/coresight-etm4x-core.c > +++ b/drivers/hwtracing/coresight/coresight-etm4x-core.c > @@ -1020,6 +1020,9 @@ static void etm4_disable_sysfs(struct coresight_device *csdev) > smp_call_function_single(drvdata->cpu, etm4_disable_hw, drvdata, 1); > > raw_spin_unlock(&drvdata->spinlock); > + > + cscfg_csdev_disable_active_config(csdev); > + In general, we need to split changes into several patches if each addresses a different issue. From my understanding, the change above is to fix missing to disable config when disable Sysfs mode. If so, could we use a seperate patch for this change? > cpus_read_unlock(); > > /* > diff --git a/drivers/hwtracing/coresight/coresight-syscfg.c b/drivers/hwtracing/coresight/coresight-syscfg.c > index a70c1454b410..6d8c212ad434 100644 > --- a/drivers/hwtracing/coresight/coresight-syscfg.c > +++ b/drivers/hwtracing/coresight/coresight-syscfg.c > @@ -391,14 +391,17 @@ static void cscfg_owner_put(struct cscfg_load_owner_info *owner_info) > static void cscfg_remove_owned_csdev_configs(struct coresight_device *csdev, void *load_owner) > { > struct cscfg_config_csdev *config_csdev, *tmp; > + unsigned long flags; > > if (list_empty(&csdev->config_csdev_list)) > return; > > + raw_spin_lock_irqsave(&csdev->cscfg_csdev_lock, flags); I think we should use spinlock to guard the condition checking list_empty(). Here the race condition is the 'config_csdev_list' list and configurations on the list. For atomicity, we should use lock to protect any operations on the list (read, add, delete, etc). A side topic, as here it adds locks for protecting 'config_csdev_list', I am wandering why we do not do the same thing for 'feature_csdev_list' (See cscfg_remove_owned_csdev_features() and cscfg_get_feat_csdev()). > list_for_each_entry_safe(config_csdev, tmp, &csdev->config_csdev_list, node) { > if (config_csdev->config_desc->load_owner == load_owner) > list_del(&config_csdev->node); > } > + raw_spin_unlock_irqrestore(&csdev->cscfg_csdev_lock, flags); > } > > static void cscfg_remove_owned_csdev_features(struct coresight_device *csdev, void *load_owner) > @@ -867,6 +870,25 @@ void cscfg_csdev_reset_feats(struct coresight_device *csdev) > } > EXPORT_SYMBOL_GPL(cscfg_csdev_reset_feats); > > +static bool cscfg_config_desc_get(struct cscfg_config_desc *config_desc) > +{ > + if (!atomic_fetch_inc(&config_desc->active_cnt)) { > + /* must ensure that config cannot be unloaded in use */ > + if (unlikely(cscfg_owner_get(config_desc->load_owner))) { > + atomic_dec(&config_desc->active_cnt); > + return false; > + } > + } > + > + return true; > +} > + > +static void cscfg_config_desc_put(struct cscfg_config_desc *config_desc) > +{ > + if (!atomic_dec_return(&config_desc->active_cnt)) > + cscfg_owner_put(config_desc->load_owner); > +} > + > /* > * This activate configuration for either perf or sysfs. Perf can have multiple > * active configs, selected per event, sysfs is limited to one. > @@ -890,22 +912,17 @@ static int _cscfg_activate_config(unsigned long cfg_hash) > if (config_desc->available == false) > return -EBUSY; > > - /* must ensure that config cannot be unloaded in use */ > - err = cscfg_owner_get(config_desc->load_owner); > - if (err) > + if (!cscfg_config_desc_get(config_desc)) { > + err = -EINVAL; > break; > + } > + > /* > * increment the global active count - control changes to > * active configurations > */ > atomic_inc(&cscfg_mgr->sys_active_cnt); Seems to me, it is more reasonable to use 'sys_active_cnt' to acquire the module reference instead of 'config_desc->active_cnt'. The reason is 'sys_active_cnt' is a global counter. > - /* > - * mark the descriptor as active so enable config on a > - * device instance will use it > - */ > - atomic_inc(&config_desc->active_cnt); > - > err = 0; > dev_dbg(cscfg_device(), "Activate config %s.\n", config_desc->name); > break; > @@ -920,9 +937,8 @@ static void _cscfg_deactivate_config(unsigned long cfg_hash) > > list_for_each_entry(config_desc, &cscfg_mgr->config_desc_list, item) { > if ((unsigned long)config_desc->event_ea->var == cfg_hash) { > - atomic_dec(&config_desc->active_cnt); > atomic_dec(&cscfg_mgr->sys_active_cnt); > - cscfg_owner_put(config_desc->load_owner); > + cscfg_config_desc_put(config_desc); > dev_dbg(cscfg_device(), "Deactivate config %s.\n", config_desc->name); > break; > } > @@ -1047,7 +1063,7 @@ int cscfg_csdev_enable_active_config(struct coresight_device *csdev, > unsigned long cfg_hash, int preset) > { > struct cscfg_config_csdev *config_csdev_active = NULL, *config_csdev_item; > - const struct cscfg_config_desc *config_desc; > + struct cscfg_config_desc *config_desc; > unsigned long flags; > int err = 0; > > @@ -1062,8 +1078,8 @@ int cscfg_csdev_enable_active_config(struct coresight_device *csdev, > raw_spin_lock_irqsave(&csdev->cscfg_csdev_lock, flags); > list_for_each_entry(config_csdev_item, &csdev->config_csdev_list, node) { > config_desc = config_csdev_item->config_desc; > - if ((atomic_read(&config_desc->active_cnt)) && > - ((unsigned long)config_desc->event_ea->var == cfg_hash)) { > + if (((unsigned long)config_desc->event_ea->var == cfg_hash) && > + cscfg_config_desc_get(config_desc)) { This seems to me not right. Why a config descriptor is get in multiple places? One time getting a config descriptor is in _cscfg_activate_config(), another is at here. To be honest, I am not clear what is the difference between 'activate' config and 'enable active' config. Literally, I think we only need to acquire the config at its creating phase (maybe match to activate config?). > config_csdev_active = config_csdev_item; > csdev->active_cscfg_ctxt = (void *)config_csdev_active; > break; > @@ -1097,7 +1113,11 @@ int cscfg_csdev_enable_active_config(struct coresight_device *csdev, > err = -EBUSY; > raw_spin_unlock_irqrestore(&csdev->cscfg_csdev_lock, flags); > } > + > + if (err) > + cscfg_config_desc_put(config_desc); > } > + > return err; > } > EXPORT_SYMBOL_GPL(cscfg_csdev_enable_active_config); > @@ -1136,8 +1156,10 @@ void cscfg_csdev_disable_active_config(struct coresight_device *csdev) > raw_spin_unlock_irqrestore(&csdev->cscfg_csdev_lock, flags); > > /* true if there was an enabled active config */ > - if (config_csdev) > + if (config_csdev) { > cscfg_csdev_disable_config(config_csdev); > + cscfg_config_desc_put(config_csdev->config_desc); > + } > } > EXPORT_SYMBOL_GPL(cscfg_csdev_disable_active_config); > > -- > LEVI:{C3F47F37-75D8-414A-A8BA-3980EC8A46D7} >

5 months

[PATCH v3 0/7] coresight: Clear self hosted claim tag on probe

by James Clark

I've gotten stuck a few times with unusable Coresight after a warm boot due to lingering claim tags, especially when testing the Coresight panic patchsets. This change does some tidy ups, adds some debug messages and clears the self hosted claim tag on probe. The last two commits are unrelated tidyups but they touch some of the same functions so to avoid extra conflicts I'm including them here. This gets as far as fixing the claim tag issue, but there is some other state not being cleared on probe that results in the following error. This can be fixed up as a later change: coresight tmc_etf0: timeout while waiting for TMC to be Ready coresight tmc_etf0: Failed to enable : TMC is not ready Changes in v3: - Collapse rename and locked/unlocked addition commits of coresight_clear_self_claim_tag() so we don't change the name twice. - Make coresight_clear_self_claim_tag() a bit more generic by only doing UNLOCK for MMIO devices (although there is no use of this right now) - Link to v2: https://lore.kernel.org/r/20250318-james-coresight-claim-tags-v2-0-e9c8a9cd… Changes in v2: * Revert most of the interface changes, just call coresight_clear_self_claim_tag() directly. This is possible because we're not doing the read first, so it has fewer knock on effects. * Split out the change to add struct cs_access to etm3x * Add another warning for racing with external debugger -- 2.34.1 --- James Clark (7): coresight: Convert tag clear function to take a struct cs_access coresight: Only check bottom two claim bits coresight: Add claim tag warnings and debug messages coresight: etm3x: Convert raw base pointer to struct coresight access coresight: Clear self hosted claim tag on probe coresight: Remove inlines from static function definitions coresight: Remove extern from function declarations drivers/hwtracing/coresight/coresight-catu.c | 12 +-- drivers/hwtracing/coresight/coresight-core.c | 87 ++++++++++++++-------- drivers/hwtracing/coresight/coresight-cti-core.c | 2 + drivers/hwtracing/coresight/coresight-etb10.c | 4 +- drivers/hwtracing/coresight/coresight-etm.h | 6 +- drivers/hwtracing/coresight/coresight-etm3x-core.c | 28 +++---- .../hwtracing/coresight/coresight-etm3x-sysfs.c | 8 +- drivers/hwtracing/coresight/coresight-etm4x-core.c | 10 ++- .../hwtracing/coresight/coresight-etm4x-sysfs.c | 4 +- drivers/hwtracing/coresight/coresight-funnel.c | 1 + drivers/hwtracing/coresight/coresight-platform.c | 26 +++---- drivers/hwtracing/coresight/coresight-priv.h | 20 ++--- drivers/hwtracing/coresight/coresight-replicator.c | 3 +- drivers/hwtracing/coresight/coresight-stm.c | 6 +- .../coresight/coresight-syscfg-configfs.c | 2 +- drivers/hwtracing/coresight/coresight-tmc-core.c | 9 ++- drivers/hwtracing/coresight/coresight-tmc-etr.c | 16 ++-- drivers/hwtracing/coresight/coresight-trbe.c | 18 ++--- include/linux/coresight.h | 40 +++++----- 19 files changed, 168 insertions(+), 134 deletions(-) --- base-commit: 5442d22da7dbff3ba8c6720fc6f23ea4934d402d change-id: 20250317-james-coresight-claim-tags-ae1461f1f5e0 Best regards, -- James Clark <james.clark(a)linaro.org>

5 months

Jump to page:

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

CoreSight