- 22 Jul, 2022 2 commits
-
-
Sudeep Holla authored
Currently, everytime an information needs to be fetched from the PPTT, the table is mapped via acpi_get_table() and unmapped after the use via acpi_put_table() which is fine. However we do this at runtime especially when the CPU is hotplugged out and plugged in back since we re-populate the cache topology and other information. However, with the support to fetch LLC information from the PPTT in the cpuhotplug path which is executed in the atomic context, it is preferred to avoid mapping and unmapping of the PPTT for every single use as the acpi_get_table() might sleep waiting for a mutex. In order to avoid the same, the table is needs to just mapped once on the boot CPU and is never unmapped allowing it to be used at runtime with out the hassle of mapping and unmapping the table. Reported-by:
Guenter Roeck <linux@roeck-us.net> Cc: Rafael J. Wysocki <rafael@kernel.org> Signed-off-by:
Sudeep Holla <sudeep.holla@arm.com> -- Hi Rafael, Sorry to bother you again on this PPTT changes. Guenter reported an issue with lockdep enabled in -next that include my cacheinfo/arch_topology changes to utilise LLC from PPTT in the CPU hotplug path. Please ack the change once you are happy so that I can get it merged with other fixes via Greg's tree. Regards, Sudeep Acked-by:
Rafael J. Wysocki <rafael.j.wysocki@intel.com> Link: https://lore.kernel.org/r/20220720-arch_topo_fixes-v3-2-43d696288e84@arm.comSigned-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
Sudeep Holla authored
On couple of architectures like RISC-V and ARM64, we need to detect cache attribues quite early during the boot when the secondary CPUs start. So we will call detect_cache_attributes in the atomic context and since use of normal allocation can sleep, we will end up getting "sleeping in the atomic context" bug splat. In order avoid that, move the allocation to use atomic version in preparation to move the actual detection of cache attributes in the CPU hotplug path which is atomic. Cc: Ionela Voinescu <ionela.voinescu@arm.com> Tested-by:
Conor Dooley <conor.dooley@microchip.com> Signed-off-by:
-
- 15 Jul, 2022 1 commit
-
-
Phil Auld authored
Using bin_attributes with a 0 size causes fstat and friends to return that 0 size. This breaks userspace code that retrieves the size before reading the file. Rather than reverting 75bd50fa ("drivers/base/node.c: use bin_attribute to break the size limitation of cpumap ABI") let's put in a size value at compile time. For cpulist the maximum size is on the order of NR_CPUS * (ceil(log10(NR_CPUS)) + 1)/2 which for 8192 is 20480 (8192 * 5)/2. In order to get near that you'd need a system with every other CPU on one node. For example: (0,2,4,8, ... ). To simplify the math and support larger NR_CPUS in the future we are using (NR_CPUS * 7)/2. We also set it to a min of PAGE_SIZE to retain the older behavior for smaller NR_CPUS. The cpumap file the size works out to be NR_CPUS/4 + NR_CPUS/32 - 1 (or NR_CPUS * 9/32 - 1) including the ","s. Add a set of macros for these values to cpumask.h so they can be used in multiple places. Apply these to the handful of such files in drivers/base/topology.c as well as node.c. As an example, on an 80 cpu 4-node system (NR_CPUS == 8192): before: -r--r--r--. 1 root root 0 Jul 12 14:08 system/node/node0/cpulist -r--r--r--. 1 root root 0 Jul 11 17:25 system/node/node0/cpumap after: -r--r--r--. 1 root root 28672 Jul 13 11:32 system/node/node0/cpulist -r--r--r--. 1 root root 4096 Jul 13 11:31 system/node/node0/cpumap CONFIG_NR_CPUS = 16384 -r--r--r--. 1 root root 57344 Jul 13 14:03 system/node/node0/cpulist -r--r--r--. 1 root root 4607 Jul 13 14:02 system/node/node0/cpumap The actual number of cpus doesn't matter for the reported size since they are based on NR_CPUS. Fixes: 75bd50fa ("drivers/base/node.c: use bin_attribute to break the size limitation of cpumap ABI") Fixes: bb9ec13d ("topology: use bin_attribute to break the size limitation of cpumap ABI") Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Rafael J. Wysocki" <rafael@kernel.org> Cc: Yury Norov <yury.norov@gmail.com> Cc: stable@vger.kernel.org Acked-by: Yury Norov <yury.norov@gmail.com> (for include/linux/cpumask.h) Signed-off-by:
Phil Auld <pauld@redhat.com> Link: https://lore.kernel.org/r/20220715134924.3466194-1-pauld@redhat.comSigned-off-by:
-
- 14 Jul, 2022 3 commits
-
-
Florian Fainelli authored
Following this mercurial changeset: https://www.mercurial-scm.org/repo/hg-stable/rev/d4ba4d51f85f update the MAINTAINERS entry to replace the now obsolete identity. Signed-off-by:
Florian Fainelli <f.fainelli@gmail.com> Link: https://lore.kernel.org/r/20220712185419.45487-1-f.fainelli@gmail.comSigned-off-by:
-
Lee Jones authored
Going forward, I'll be using my kernel.org for upstream work. Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mauro Carvalho Chehab <mchehab+huawei@kernel.org> Signed-off-by:
Lee Jones <lee.jones@linaro.org> Signed-off-by:
Lee Jones <lee@kernel.org> Link: https://lore.kernel.org/r/20220714112533.539910-6-lee@kernel.orgSigned-off-by:
-
Lee Jones authored
Going forward, I'll be using my kernel.org for upstream work. Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mauro Carvalho Chehab <mchehab+huawei@kernel.org> Signed-off-by:
-
- 12 Jul, 2022 1 commit
-
-
Nick Desaulniers authored
Should the need for toolchain mitigations ever be necessary, add a group for toolchain ambassadors. Add Nick Desaulniers as LLVM's ambassador for the embargoed hardware issues process. Signed-off-by:
Nick Desaulniers <ndesaulniers@google.com> Link: https://lore.kernel.org/r/20220711181101.1559558-1-ndesaulniers@google.comSigned-off-by:
-
- 08 Jul, 2022 1 commit
-
-
Greg Kroah-Hartman authored
Merge tag 'arch-cache-topo-5.20' of git://git.kernel.org/pub/scm/linux/kernel/git/sudeep.holla/linux into driver-core-next Sudeep writes: cacheinfo and arch_topology updates for v5.20 These are updates to fix some discrepancies we have in the CPU topology parsing from the device tree /cpu-map node and the divergence from the behaviour on a ACPI enabled platform. The expectation is that both DT and ACPI enabled systems must present consistent view of the CPU topology. The current assignment of generated cluster count as the physical package identifier for each CPU is wrong. The device tree bindings for CPU topology supports sockets to infer the socket or physical package identifier for a given CPU. It is now being made use of you address the issue. These updates also assigns the cluster identifier as parsed from the device tree cluster nodes within /cpu-map without support for nesting of the clusters as there are no such reported/known platforms. In order to be on par with ACPI PPTT physical package/socket support, these updates also include support for socket nodes in /cpu-map. The only exception is that the last level cache id information can be inferred from the same ACPI PPTT while we need to parse CPU cache nodes in the device tree. The cacheinfo changes here is to enable the re-use of the cacheinfo to detect the cache attributes for all the CPU quite early even before the scondardaries are booted so that the information can be used to build the schedular domains especially the last level cache(LLC). * tag 'arch-cache-topo-5.20' of git://git.kernel.org/pub/scm/linux/kernel/git/sudeep.holla/linux: (21 commits) ACPI: Remove the unused find_acpi_cpu_cache_topology() arch_topology: Warn that topology for nested clusters is not supported arch_topology: Add support for parsing sockets in /cpu-map arch_topology: Set cluster identifier in each core/thread from /cpu-map arch_topology: Limit span of cpu_clustergroup_mask() arch_topology: Don't set cluster identifier as physical package identifier arch_topology: Avoid parsing through all the CPUs once a outlier CPU is found arch_topology: Check for non-negative value rather than -1 for IDs validity arch_topology: Set thread sibling cpumask only within the cluster arch_topology: Drop LLC identifier stash from the CPU topology arm64: topology: Remove redundant setting of llc_id in CPU topology arch_topology: Use the last level cache information from the cacheinfo arch_topology: Add support to parse and detect cache attributes cacheinfo: Align checks in cache_shared_cpu_map_{setup,remove} for readability cacheinfo: Use cache identifiers to check if the caches are shared if available cacheinfo: Allow early detection and population of cache attributes cacheinfo: Add support to check if last level cache(LLC) is valid or shared cacheinfo: Move cache_leaves_are_shared out of CONFIG_OF cacheinfo: Add helper to access any cache index for a given CPU cacheinfo: Use of_cpu_device_node_get instead cpu_dev->of_node ...
-
- 06 Jul, 2022 1 commit
-
-
Imran Khan authored
This reverts commit b8f35fa1. This is causing regression due to same kernfs_node getting added multiple times in kernfs_notify_list so revert it until safe way of using llist in this context is found. Reported-by:
Nathan Chancellor <nathan@kernel.org> Reported-by:
Michael Walle <michael@walle.cc> Reported-by:
Marek Szyprowski <m.szyprowski@samsung.com> Signed-off-by:
Imran Khan <imran.f.khan@oracle.com> Cc: Tejun Heo <tj@kernel.org> Link: https://lore.kernel.org/r/20220705201026.2487665-1-imran.f.khan@oracle.comSigned-off-by:
-
- 04 Jul, 2022 21 commits
-
-
Sudeep Holla authored
The sole user of this find_acpi_cpu_cache_topology() was arm64 topology which is now consolidated into the generic arch_topology without the need of this function. Drop the unused function find_acpi_cpu_cache_topology(). Link: https://lore.kernel.org/r/20220704101605.1318280-22-sudeep.holla@arm.com Cc: Rafael J. Wysocki <rafael@kernel.org> Reported-by:
Ionela Voinescu <ionela.voinescu@arm.com> Tested-by:
-
Sudeep Holla authored
We don't support the topology for clusters of CPU clusters while the DT and ACPI bindings theoritcally support the same. Just warn about the same so that it is clear to the users of arch_topology that the nested clusters are not yet supported. Link: https://lore.kernel.org/r/20220704101605.1318280-21-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
Finally let us add support for socket nodes in /cpu-map in the device tree. Since this may not be present in all the old platforms and even most of the existing platforms, we need to assume absence of the socket node indicates that it is a single socket system and handle appropriately. Also it is likely that most single socket systems skip to as the node since it is optional. Link: https://lore.kernel.org/r/20220704101605.1318280-20-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
Let us set the cluster identifier as parsed from the device tree cluster nodes within /cpu-map. We don't support nesting of clusters yet as there are no real hardware to support clusters of clusters. Link: https://lore.kernel.org/r/20220704101605.1318280-19-sudeep.holla@arm.comTested-by:
-
Ionela Voinescu authored
Currently the cluster identifier is not set on DT based platforms. The reset or default value is -1 for all the CPUs. Once we assign the cluster identifier values correctly, the cluster_sibling mask will be populated and returned by cpu_clustergroup_mask() to contribute in the creation of the CLS scheduling domain level, if SCHED_CLUSTER is enabled. To avoid topologies that will result in questionable or incorrect scheduling domains, impose restrictions regarding the span of clusters, as presented to scheduling domains building code: cluster_sibling should not span more or the same CPUs as cpu_coregroup_mask(). This is needed in order to obtain a strict separation between the MC and CLS levels, and maintain the same domains for existing platforms in the presence of CONFIG_SCHED_CLUSTER, where the new cluster information is redundant and irrelevant for the scheduler. While previously the scheduling domain builder code would have removed MC as redundant and kept CLS if SCHED_CLUSTER was enabled and the cpu_coregroup_mask() and cpu_clustergroup_mask() spanned the same CPUs, now CLS will be removed and MC kept. Link: https://lore.kernel.org/r/20220704101605.1318280-18-sudeep.holla@arm.com Cc: Darren Hart <darren@os.amperecomputing.com> Tested-by:
Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by:
-
Sudeep Holla authored
Currently as we parse the CPU topology from /cpu-map node from the device tree, we assign generated cluster count as the physical package identifier for each CPU which is wrong. The device tree bindings for CPU topology supports sockets to infer the socket or physical package identifier for a given CPU. Since it is fairly new and not supported on most of the old and existing systems, we can assume all such systems have single socket/physical package. Fix the physical package identifier to 0 by removing the assignment of cluster identifier to the same. Link: https://lore.kernel.org/r/20220704101605.1318280-17-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
There is no point in looping through all the CPU's physical package identifier to check if it is valid or not once a CPU which is outside the topology(i.e. outlier CPU) is found. Let us just break out of the loop early in such case. Link: https://lore.kernel.org/r/20220704101605.1318280-16-sudeep.holla@arm.comTested-by:
Gavin Shan <gshan@redhat.com> Signed-off-by:
-
Sudeep Holla authored
Instead of just comparing the cpu topology IDs with -1 to check their validity, improve that by checking for a valid non-negative value. Link: https://lore.kernel.org/r/20220704101605.1318280-15-sudeep.holla@arm.comSuggested-by:
Andy Shevchenko <andy.shevchenko@gmail.com> Tested-by:
-
Sudeep Holla authored
Currently the cluster identifier is not set on the DT based platforms. The reset or default value is -1 for all the CPUs. Once we assign the cluster identifier values correctly that may result in getting the thread siblings wrong as the core identifiers can be same for 2 different CPUs belonging to 2 different cluster. So, in order to get the thread sibling cpumasks correct, we need to update them only if the cores they belong are in the same cluster within the socket. Let us skip updation of the thread sibling cpumaks if the cluster identifier doesn't match. This change won't affect even if the cluster identifiers are not set currently but will avoid any breakage once we set the same correctly. Link: https://lore.kernel.org/r/20220704101605.1318280-14-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
Since the cacheinfo LLC information is used directly in arch_topology, there is no need to parse and store the LLC ID information only for ACPI systems in the CPU topology. Remove the redundant LLC ID from the generic CPU arch_topology information. Link: https://lore.kernel.org/r/20220704101605.1318280-13-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
Since the cacheinfo LLC information is used directly in arch_topology, there is no need to parse and fetch the LLC ID information only for ACPI systems. Just drop the redundant parsing and setting of llc_id in CPU topology from ACPI PPTT. Link: https://lore.kernel.org/r/20220704101605.1318280-12-sudeep.holla@arm.com Cc: Will Deacon <will@kernel.org> Cc: Catalin Marinas <catalin.marinas@arm.com> Tested-by:
Catalin Marinas <catalin.marinas@arm.com> Signed-off-by:
-
Sudeep Holla authored
The cacheinfo is now initialised early along with the CPU topology initialisation. Instead of relying on the LLC ID information parsed separately only with ACPI PPTT elsewhere, migrate to use the similar information from the cacheinfo. This is generic for both DT and ACPI systems. The ACPI LLC ID information parsed separately can now be removed from arch specific code. Link: https://lore.kernel.org/r/20220704101605.1318280-11-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
Currently ACPI populates just the minimum information about the last level cache from PPTT in order to feed the same to build sched_domains. Similar support for DT platforms is not present. In order to enable the same, the entire cache hierarchy information can be built as part of CPU topoplogy parsing both on ACPI and DT platforms. Note that this change builds the cacheinfo early even on ACPI systems, but the current mechanism of building llc_sibling mask remains unchanged. Link: https://lore.kernel.org/r/20220704101605.1318280-10-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
The checks to skip the CPU itself or no cacheinfo case are implemented bit differently though the effect is exactly same. Just align the implementation in both cache_shared_cpu_map_{setup,remove} just for improved readability. No functional change. Link: https://lore.kernel.org/r/20220704101605.1318280-9-sudeep.holla@arm.comTested-by: -
Sudeep Holla authored
The cache identifiers is an optional property on most of the platforms. The presence of one must be indicated by the CACHE_ID valid bit in the attributes. We can use the cache identifiers provided by the firmware to check if any two cpus share the same cache instead of relying on the fw_token generated and set in the OS. Link: https://lore.kernel.org/r/20220704101605.1318280-8-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
Some architecture/platforms may need to setup cache properties very early in the boot along with other cpu topologies so that all these information can be used to build sched_domains which is used by the scheduler. Allow detect_cache_attributes to be called quite early during the boot. Link: https://lore.kernel.org/r/20220704101605.1318280-7-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
It is useful to have helper to check if the given two CPUs share last level cache. We can do that check by comparing fw_token or by comparing the cache ID. Currently we check just for fw_token as the cache ID is optional. This helper can be used to build the llc_sibling during arch specific topology parsing and feeding information to the sched_domains. This also helps to get rid of llc_id in the CPU topology as it is sort of duplicate information. Also add helper to check if the llc information in cacheinfo is valid or not. Link: https://lore.kernel.org/r/20220704101605.1318280-6-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
cache_leaves_are_shared is already used even with ACPI and PPTT. It checks if the cache leaves are the shared based on fw_token pointer. However it is defined conditionally only if CONFIG_OF is enabled which is wrong. Move the function cache_leaves_are_shared out of CONFIG_OF and keep it generic. It also handles the case where both OF and ACPI is not defined. Link: https://lore.kernel.org/r/20220704101605.1318280-5-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
The cacheinfo for a given CPU at a given index is used at quite a few places by fetching the base point for index 0 using the helper per_cpu_cacheinfo(cpu) and offsetting it by the required index. Instead, add another helper to fetch the required pointer directly and use it to simplify and improve readability. Link: https://lore.kernel.org/r/20220704101605.1318280-4-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
The of_cpu_device_node_get takes care of fetching the CPU'd device node either from cached cpu_dev->of_node if cpu_dev is initialised or uses of_get_cpu_node to parse and fetch node if cpu_dev isn't available yet. Just use of_cpu_device_node_get instead of getting the cpu device first and then using cpu_dev->of_node for two reasons: 1. There is no other use of cpu_dev and can be simplified 2. It enabled the use detect_cache_attributes and hence cache_setup_of_node much earlier before the CPUs are registered as devices. Link: https://lore.kernel.org/r/20220704101605.1318280-3-sudeep.holla@arm.comTested-by:
-
Sudeep Holla authored
There is need to use the cache sharing information quite early during the boot before the secondary cores are up and running. The permanent memory map for all the ACPI tables(via acpi_permanent_mmap) is turned on in acpi_early_init() which is quite late for the above requirement. As a result there is possibility that the ACPI PPTT gets mapped to different virtual addresses. In such scenarios, using virtual address as fw_token before the acpi_permanent_mmap is enabled results in different fw_token for the same cache entity and hence wrong cache sharing information will be deduced based on the same. Instead of using virtual address, just use the table offset as the unique firmware token for the caches. The same offset is used as ACPI identifiers if the firmware has not set a valid one for other entries in the ACPI PPTT. Link: https://lore.kernel.org/r/20220704101605.1318280-2-sudeep.holla@arm.com Cc: linux-acpi@vger.kernel.org Tested-by:
Rafael J. Wysocki <rafael@kernel.org> Signed-off-by:
-
- 01 Jul, 2022 2 commits
-
-
Yushan Zhou authored
When lockdep is enabled, lockdep_assert_held_write would cause potential NULL pointer dereference. Fix the following smatch warnings: fs/kernfs/dir.c:1353 __kernfs_remove() warn: variable dereferenced before check 'kn' (see line 1346) Fixes: 393c3714 ("kernfs: switch global kernfs_rwsem lock to per-fs lock") Signed-off-by:
Yushan Zhou <katrinzhou@tencent.com> Link: https://lore.kernel.org/r/20220630082512.3482581-1-zys.zljxml@gmail.comSigned-off-by:
-
Liang He authored
In of_register_trusted_foundations(), we need to hold the reference returned by of_find_compatible_node() and then use it to call of_node_put() for refcount balance. Signed-off-by:
Liang He <windhl@126.com> Link: https://lore.kernel.org/r/20220628021640.4015-1-windhl@126.comSigned-off-by:
-
- 27 Jun, 2022 8 commits
-
-
Saravana Kannan authored
Commit 71066545 ("driver core: Set fw_devlink.strict=1 by default") enabled iommus and dmas dependency enforcement by default. On some systems, this caused the console device's probe to get delayed until the deferred_probe_timeout expires. We need consoles to work as soon as possible, so mark the console device node with FWNODE_FLAG_BEST_EFFORT so that fw_delink knows not to delay the probe of the console device for suppliers without drivers. The driver can then make the decision on where it can probe without those suppliers or defer its probe. Fixes: 71066545 ("driver core: Set fw_devlink.strict=1 by default") Reported-by:
Sascha Hauer <sha@pengutronix.de> Reported-by:
Peng Fan <peng.fan@nxp.com> Tested-by:
Saravana Kannan <saravanak@google.com> Link: https://lore.kernel.org/r/20220623080344.783549-3-saravanak@google.comSigned-off-by:
-
Saravana Kannan authored
When firmware sets the FWNODE_FLAG_BEST_EFFORT flag for a fwnode, fw_devlink will do a best effort ordering for that device where it'll only enforce the probe/suspend/resume ordering of that device with suppliers that have drivers. The driver of that device can then decide if it wants to defer probe or probe without the suppliers. This will be useful for avoid probe delays of the console device that were caused by commit 71066545 ("driver core: Set fw_devlink.strict=1 by default"). Fixes: 71066545 ("driver core: Set fw_devlink.strict=1 by default") Reported-by:
-
Imran Khan authored
In current kernfs design a single mutex, kernfs_open_file_mutex, protects the list of kernfs_open_file instances corresponding to a sysfs attribute. So even if different tasks are opening or closing different sysfs files they can contend on osq_lock of this mutex. The contention is more apparent in large scale systems with few hundred CPUs where most of the CPUs have running tasks that are opening, accessing or closing sysfs files at any point of time. Using hashed mutexes in place of a single global mutex, can significantly reduce contention around global mutex and hence can provide better scalability. Moreover as these hashed mutexes are not part of kernfs_node objects we will not see any singnificant change in memory utilization of kernfs based file systems like sysfs, cgroupfs etc. Modify interface introduced in previous patch to make use of hashed mutexes. Use kernfs_node address as hashing key. Acked-by:
Tejun Heo <tj@kernel.org> Signed-off-by:
-
Imran Khan authored
This allows to change underlying mutex locking, without needing to change the users of the lock. For example next patch modifies this interface to use hashed mutexes in place of a single global kernfs_open_file_mutex. Acked-by:
-
Imran Khan authored
At present kernfs_notify_list is implemented as a singly linked list of kernfs_node(s), where last element points to itself and value of ->attr.next tells if node is present on the list or not. Both addition and deletion to list happen under kernfs_notify_lock. Change kernfs_notify_list to llist so that addition to list can heppen locklessly. Suggested by: Al Viro <viro@zeniv.linux.org.uk> Acked-by:
-
Imran Khan authored
After removal of kernfs_open_node->refcnt in the previous patch, kernfs_open_node_lock can be removed as well by making ->attr.open RCU protected. kernfs_put_open_node can delegate freeing to ->attr.open to RCU and other readers of ->attr.open can do so under rcu_read_(un)lock. Suggested by: Al Viro <viro@zeniv.linux.org.uk> Acked-by:
-
Lin Feng authored
Since previous 'rc = -EINVAL;', rc value doesn't change, so not necessary to re-assign it again. Signed-off-by:
Lin Feng <linf@wangsu.com> Link: https://lore.kernel.org/r/20220617091746.206515-1-linf@wangsu.comSigned-off-by:
-
Zhang Wensheng authored
In __driver_attach function, There are also AA deadlock problem, like the commit b232b02b ("driver core: fix deadlock in __device_attach"). stack like commit b232b02b ("driver core: fix deadlock in __device_attach"). list below: In __driver_attach function, The lock holding logic is as follows: ... __driver_attach if (driver_allows_async_probing(drv)) device_lock(dev) // get lock dev async_schedule_dev(__driver_attach_async_helper, dev); // func async_schedule_node async_schedule_node_domain(func) entry = kzalloc(sizeof(struct async_entry), GFP_ATOMIC); /* when fail or work limit, sync to execute func, but __driver_attach_async_helper will get lock dev as will, which will lead to A-A deadlock. */ if (!entry || atomic_read(&entry_count) > MAX_WORK) { func; else queue_work_node(node, system_unbound_wq, &entry->work) device_unlock(dev) As above show, when it is allowed to do async probes, because of out of memory or work limit, async work is not be allowed, to do sync execute instead. it will lead to A-A deadlock because of __driver_attach_async_helper getting lock dev. Reproduce: and it can be reproduce by make the condition (if (!entry || atomic_read(&entry_count) > MAX_WORK)) untenable, like below: [ 370.785650] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 370.787154] task:swapper/0 state:D stack: 0 pid: 1 ppid: 0 flags:0x00004000 [ 370.788865] Call Trace: [ 370.789374] <TASK> [ 370.789841] __schedule+0x482/0x1050 [ 370.790613] schedule+0x92/0x1a0 [ 370.791290] schedule_preempt_disabled+0x2c/0x50 [ 370.792256] __mutex_lock.isra.0+0x757/0xec0 [ 370.793158] __mutex_lock_slowpath+0x1f/0x30 [ 370.794079] mutex_lock+0x50/0x60 [ 370.794795] __device_driver_lock+0x2f/0x70 [ 370.795677] ? driver_probe_device+0xd0/0xd0 [ 370.796576] __driver_attach_async_helper+0x1d/0xd0 [ 370.797318] ? driver_probe_device+0xd0/0xd0 [ 370.797957] async_schedule_node_domain+0xa5/0xc0 [ 370.798652] async_schedule_node+0x19/0x30 [ 370.799243] __driver_attach+0x246/0x290 [ 370.799828] ? driver_allows_async_probing+0xa0/0xa0 [ 370.800548] bus_for_each_dev+0x9d/0x130 [ 370.801132] driver_attach+0x22/0x30 [ 370.801666] bus_add_driver+0x290/0x340 [ 370.802246] driver_register+0x88/0x140 [ 370.802817] ? virtio_scsi_init+0x116/0x116 [ 370.803425] scsi_register_driver+0x1a/0x30 [ 370.804057] init_sd+0x184/0x226 [ 370.804533] do_one_initcall+0x71/0x3a0 [ 370.805107] kernel_init_freeable+0x39a/0x43a [ 370.805759] ? rest_init+0x150/0x150 [ 370.806283] kernel_init+0x26/0x230 [ 370.806799] ret_from_fork+0x1f/0x30 To fix the deadlock, move the async_schedule_dev outside device_lock, as we can see, in async_schedule_node_domain, the parameter of queue_work_node is system_unbound_wq, so it can accept concurrent operations. which will also not change the code logic, and will not lead to deadlock. Fixes: ef0ff683 ("driver core: Probe devices asynchronously instead of the driver") Signed-off-by:
Zhang Wensheng <zhangwensheng5@huawei.com> Link: https://lore.kernel.org/r/20220622074327.497102-1-zhangwensheng5@huawei.comSigned-off-by:
-
