From e7f1d65f56c43a747b9fd6cc3d2eea72210fa077 Mon Sep 17 00:00:00 2001 From: Srinivas Pandruvada Date: Tue, 11 Jan 2022 18:25:21 -0800 Subject: [PATCH 1/3] platform/x86: ISST: Fix possible circular locking dependency detected commit 17da2d5f93692086dd096a975225ffd5622d0bf8 upstream. As reported: [ 256.104522] ====================================================== [ 256.113783] WARNING: possible circular locking dependency detected [ 256.120093] 5.16.0-rc6-yocto-standard+ #99 Not tainted [ 256.125362] ------------------------------------------------------ [ 256.131673] intel-speed-sel/844 is trying to acquire lock: [ 256.137290] ffffffffc036f0d0 (punit_misc_dev_lock){+.+.}-{3:3}, at: isst_if_open+0x18/0x90 [isst_if_common] [ 256.147171] [ 256.147171] but task is already holding lock: [ 256.153135] ffffffff8ee7cb50 (misc_mtx){+.+.}-{3:3}, at: misc_open+0x2a/0x170 [ 256.160407] [ 256.160407] which lock already depends on the new lock. [ 256.160407] [ 256.168712] [ 256.168712] the existing dependency chain (in reverse order) is: [ 256.176327] [ 256.176327] -> #1 (misc_mtx){+.+.}-{3:3}: [ 256.181946] lock_acquire+0x1e6/0x330 [ 256.186265] __mutex_lock+0x9b/0x9b0 [ 256.190497] mutex_lock_nested+0x1b/0x20 [ 256.195075] misc_register+0x32/0x1a0 [ 256.199390] isst_if_cdev_register+0x65/0x180 [isst_if_common] [ 256.205878] isst_if_probe+0x144/0x16e [isst_if_mmio] ... [ 256.241976] [ 256.241976] -> #0 (punit_misc_dev_lock){+.+.}-{3:3}: [ 256.248552] validate_chain+0xbc6/0x1750 [ 256.253131] __lock_acquire+0x88c/0xc10 [ 256.257618] lock_acquire+0x1e6/0x330 [ 256.261933] __mutex_lock+0x9b/0x9b0 [ 256.266165] mutex_lock_nested+0x1b/0x20 [ 256.270739] isst_if_open+0x18/0x90 [isst_if_common] [ 256.276356] misc_open+0x100/0x170 [ 256.280409] chrdev_open+0xa5/0x1e0 ... The call sequence suggested that misc_device /dev file can be opened before misc device is yet to be registered, which is done only once. Here punit_misc_dev_lock was used as common lock, to protect the registration by multiple ISST HW drivers, one time setup, prevent duplicate registry of misc device and prevent load/unload when device is open. We can split into locks: - One which just prevent duplicate call to misc_register() and one time setup. Also never call again if the misc_register() failed or required one time setup is failed. This lock is not shared with any misc device callbacks. - The other lock protects registry, load and unload of HW drivers. Sequence in isst_if_cdev_register() - Register callbacks under punit_misc_dev_open_lock - Call isst_misc_reg() which registers misc_device on the first registry which is under punit_misc_dev_reg_lock, which is not shared with callbacks. Sequence in isst_if_cdev_unregister Just opposite of isst_if_cdev_register Intel-SIG: commit 17da2d5f9369 platform/x86: ISST: Fix possible circular locking dependency detected. Backport for ISST driver fix Reported-and-tested-by: Liwei Song Signed-off-by: Srinivas Pandruvada Link: https://lore.kernel.org/r/20220112022521.54669-1-srinivas.pandruvada@linux.intel.com Reviewed-by: Hans de Goede Signed-off-by: Hans de Goede [ yingbao jia: amend commit log ] Signed-off-by: yingbao jia --- .../intel_speed_select_if/isst_if_common.c | 97 ++++++++++++------- 1 file changed, 63 insertions(+), 34 deletions(-) diff --git a/drivers/platform/x86/intel_speed_select_if/isst_if_common.c b/drivers/platform/x86/intel_speed_select_if/isst_if_common.c index 61ddf9f769af..6a399a6fd881 100644 --- a/drivers/platform/x86/intel_speed_select_if/isst_if_common.c +++ b/drivers/platform/x86/intel_speed_select_if/isst_if_common.c @@ -595,7 +595,10 @@ static long isst_if_def_ioctl(struct file *file, unsigned int cmd, return ret; } -static DEFINE_MUTEX(punit_misc_dev_lock); +/* Lock to prevent module registration when already opened by user space */ +static DEFINE_MUTEX(punit_misc_dev_open_lock); +/* Lock to allow one share misc device for all ISST interace */ +static DEFINE_MUTEX(punit_misc_dev_reg_lock); static int misc_usage_count; static int misc_device_ret; static int misc_device_open; @@ -605,7 +608,7 @@ static int isst_if_open(struct inode *inode, struct file *file) int i, ret = 0; /* Fail open, if a module is going away */ - mutex_lock(&punit_misc_dev_lock); + mutex_lock(&punit_misc_dev_open_lock); for (i = 0; i < ISST_IF_DEV_MAX; ++i) { struct isst_if_cmd_cb *cb = &punit_callbacks[i]; @@ -627,7 +630,7 @@ static int isst_if_open(struct inode *inode, struct file *file) } else { misc_device_open++; } - mutex_unlock(&punit_misc_dev_lock); + mutex_unlock(&punit_misc_dev_open_lock); return ret; } @@ -636,7 +639,7 @@ static int isst_if_relase(struct inode *inode, struct file *f) { int i; - mutex_lock(&punit_misc_dev_lock); + mutex_lock(&punit_misc_dev_open_lock); misc_device_open--; for (i = 0; i < ISST_IF_DEV_MAX; ++i) { struct isst_if_cmd_cb *cb = &punit_callbacks[i]; @@ -644,7 +647,7 @@ static int isst_if_relase(struct inode *inode, struct file *f) if (cb->registered) module_put(cb->owner); } - mutex_unlock(&punit_misc_dev_lock); + mutex_unlock(&punit_misc_dev_open_lock); return 0; } @@ -661,6 +664,43 @@ static struct miscdevice isst_if_char_driver = { .fops = &isst_if_char_driver_ops, }; +static int isst_misc_reg(void) +{ + mutex_lock(&punit_misc_dev_reg_lock); + if (misc_device_ret) + goto unlock_exit; + + if (!misc_usage_count) { + misc_device_ret = isst_if_cpu_info_init(); + if (misc_device_ret) + goto unlock_exit; + + misc_device_ret = misc_register(&isst_if_char_driver); + if (misc_device_ret) { + isst_if_cpu_info_exit(); + goto unlock_exit; + } + } + misc_usage_count++; + +unlock_exit: + mutex_unlock(&punit_misc_dev_reg_lock); + + return misc_device_ret; +} + +static void isst_misc_unreg(void) +{ + mutex_lock(&punit_misc_dev_reg_lock); + if (misc_usage_count) + misc_usage_count--; + if (!misc_usage_count && !misc_device_ret) { + misc_deregister(&isst_if_char_driver); + isst_if_cpu_info_exit(); + } + mutex_unlock(&punit_misc_dev_reg_lock); +} + /** * isst_if_cdev_register() - Register callback for IOCTL * @device_type: The device type this callback handling. @@ -678,38 +718,31 @@ static struct miscdevice isst_if_char_driver = { */ int isst_if_cdev_register(int device_type, struct isst_if_cmd_cb *cb) { - if (misc_device_ret) - return misc_device_ret; + int ret; if (device_type >= ISST_IF_DEV_MAX) return -EINVAL; - mutex_lock(&punit_misc_dev_lock); + mutex_lock(&punit_misc_dev_open_lock); + /* Device is already open, we don't want to add new callbacks */ if (misc_device_open) { - mutex_unlock(&punit_misc_dev_lock); + mutex_unlock(&punit_misc_dev_open_lock); return -EAGAIN; } - if (!misc_usage_count) { - int ret; - - misc_device_ret = misc_register(&isst_if_char_driver); - if (misc_device_ret) - goto unlock_exit; - - ret = isst_if_cpu_info_init(); - if (ret) { - misc_deregister(&isst_if_char_driver); - misc_device_ret = ret; - goto unlock_exit; - } - } memcpy(&punit_callbacks[device_type], cb, sizeof(*cb)); punit_callbacks[device_type].registered = 1; - misc_usage_count++; -unlock_exit: - mutex_unlock(&punit_misc_dev_lock); + mutex_unlock(&punit_misc_dev_open_lock); - return misc_device_ret; + ret = isst_misc_reg(); + if (ret) { + /* + * No need of mutex as the misc device register failed + * as no one can open device yet. Hence no contention. + */ + punit_callbacks[device_type].registered = 0; + return ret; + } + return 0; } EXPORT_SYMBOL_GPL(isst_if_cdev_register); @@ -724,16 +757,12 @@ EXPORT_SYMBOL_GPL(isst_if_cdev_register); */ void isst_if_cdev_unregister(int device_type) { - mutex_lock(&punit_misc_dev_lock); - misc_usage_count--; + isst_misc_unreg(); + mutex_lock(&punit_misc_dev_open_lock); punit_callbacks[device_type].registered = 0; if (device_type == ISST_IF_DEV_MBOX) isst_delete_hash(); - if (!misc_usage_count && !misc_device_ret) { - misc_deregister(&isst_if_char_driver); - isst_if_cpu_info_exit(); - } - mutex_unlock(&punit_misc_dev_lock); + mutex_unlock(&punit_misc_dev_open_lock); } EXPORT_SYMBOL_GPL(isst_if_cdev_unregister); -- Gitee From 73df13a6faa4dc09c0e0a3e09179a1b4ce935b40 Mon Sep 17 00:00:00 2001 From: Srinivas Pandruvada Date: Wed, 29 Jun 2022 12:48:17 -0700 Subject: [PATCH 2/3] platform/x86: ISST: PUNIT device mapping with Sub-NUMA clustering commit 9a1aac8a96dc014bec49806a7a964bf2fdbd315f upstream. On a multiple package system using Sub-NUMA clustering, there is an issue in mapping Linux CPU number to PUNIT PCI device when manufacturer decided to reuse the PCI bus number across packages. Bus number can be reused as long as they are in different domain or segment. In this case some CPU will fail to find a PCI device to issue SST requests. When bus numbers are reused across CPU packages, we are using proximity information by matching CPU numa node id to PUNIT PCI device numa node id. But on a package there can be only one PUNIT PCI device, but multiple numa nodes (one for each sub cluster). So, the numa node ID of the PUNIT PCI device can only match with one numa node id of CPUs in a sub cluster in the package. Since there can be only one PUNIT PCI device per package, if we match with numa node id of any sub cluster in that package, we can use that mapping for any CPU in that package. So, store the match information in a per package data structure and return the information when there is no match. While here, use defines for max bus number instead of hardcoding. Intel-SIG: commit 9a1aac8a96dc platform/x86: ISST: PUNIT device mapping with Sub-NUMA clustering. Backport for ISST driver fix Signed-off-by: Srinivas Pandruvada Link: https://lore.kernel.org/r/20220629194817.2418240-1-srinivas.pandruvada@linux.intel.com Reviewed-by: Hans de Goede Signed-off-by: Hans de Goede [ yingbao jia: amend commit log ] Signed-off-by: yingbao jia --- .../intel_speed_select_if/isst_if_common.c | 39 +++++++++++++++---- 1 file changed, 32 insertions(+), 7 deletions(-) diff --git a/drivers/platform/x86/intel_speed_select_if/isst_if_common.c b/drivers/platform/x86/intel_speed_select_if/isst_if_common.c index 6a399a6fd881..2ea930517914 100644 --- a/drivers/platform/x86/intel_speed_select_if/isst_if_common.c +++ b/drivers/platform/x86/intel_speed_select_if/isst_if_common.c @@ -277,29 +277,38 @@ static int isst_if_get_platform_info(void __user *argp) return 0; } +#define ISST_MAX_BUS_NUMBER 2 struct isst_if_cpu_info { /* For BUS 0 and BUS 1 only, which we need for PUNIT interface */ - int bus_info[2]; - struct pci_dev *pci_dev[2]; + int bus_info[ISST_MAX_BUS_NUMBER]; + struct pci_dev *pci_dev[ISST_MAX_BUS_NUMBER]; int punit_cpu_id; int numa_node; }; +struct isst_if_pkg_info { + struct pci_dev *pci_dev[ISST_MAX_BUS_NUMBER]; +}; + static struct isst_if_cpu_info *isst_cpu_info; +static struct isst_if_pkg_info *isst_pkg_info; + #define ISST_MAX_PCI_DOMAINS 8 static struct pci_dev *_isst_if_get_pci_dev(int cpu, int bus_no, int dev, int fn) { struct pci_dev *matched_pci_dev = NULL; struct pci_dev *pci_dev = NULL; - int no_matches = 0; + int no_matches = 0, pkg_id; int i, bus_number; - if (bus_no < 0 || bus_no > 1 || cpu < 0 || cpu >= nr_cpu_ids || - cpu >= num_possible_cpus()) + if (bus_no < 0 || bus_no >= ISST_MAX_BUS_NUMBER || cpu < 0 || + cpu >= nr_cpu_ids || cpu >= num_possible_cpus()) return NULL; + pkg_id = topology_physical_package_id(cpu); + bus_number = isst_cpu_info[cpu].bus_info[bus_no]; if (bus_number < 0) return NULL; @@ -324,6 +333,8 @@ static struct pci_dev *_isst_if_get_pci_dev(int cpu, int bus_no, int dev, int fn } if (node == isst_cpu_info[cpu].numa_node) { + isst_pkg_info[pkg_id].pci_dev[bus_no] = _pci_dev; + pci_dev = _pci_dev; break; } @@ -342,6 +353,10 @@ static struct pci_dev *_isst_if_get_pci_dev(int cpu, int bus_no, int dev, int fn if (!pci_dev && no_matches == 1) pci_dev = matched_pci_dev; + /* Return pci_dev pointer for any matched CPU in the package */ + if (!pci_dev) + pci_dev = isst_pkg_info[pkg_id].pci_dev[bus_no]; + return pci_dev; } @@ -361,8 +376,8 @@ struct pci_dev *isst_if_get_pci_dev(int cpu, int bus_no, int dev, int fn) { struct pci_dev *pci_dev; - if (bus_no < 0 || bus_no > 1 || cpu < 0 || cpu >= nr_cpu_ids || - cpu >= num_possible_cpus()) + if (bus_no < 0 || bus_no >= ISST_MAX_BUS_NUMBER || cpu < 0 || + cpu >= nr_cpu_ids || cpu >= num_possible_cpus()) return NULL; pci_dev = isst_cpu_info[cpu].pci_dev[bus_no]; @@ -416,10 +431,19 @@ static int isst_if_cpu_info_init(void) if (!isst_cpu_info) return -ENOMEM; + isst_pkg_info = kcalloc(topology_max_packages(), + sizeof(*isst_pkg_info), + GFP_KERNEL); + if (!isst_pkg_info) { + kfree(isst_cpu_info); + return -ENOMEM; + } + ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "platform/x86/isst-if:online", isst_if_cpu_online, NULL); if (ret < 0) { + kfree(isst_pkg_info); kfree(isst_cpu_info); return ret; } @@ -432,6 +456,7 @@ static int isst_if_cpu_info_init(void) static void isst_if_cpu_info_exit(void) { cpuhp_remove_state(isst_if_online_id); + kfree(isst_pkg_info); kfree(isst_cpu_info); }; -- Gitee From 3a7624d733ace5ca49cd328492b7babf86cb2269 Mon Sep 17 00:00:00 2001 From: Srinivas Pandruvada Date: Tue, 27 Jul 2021 09:50:52 -0700 Subject: [PATCH 3/3] platform/x86: ISST: Fix optimization with use of numa commit d36d4a1d75d2a8bd14ec00d5cb0ce166f6886146 upstream. When numa is used to map CPU to PCI device, the optimized path to read from cached data is not working and still calls _isst_if_get_pci_dev(). The reason is that when caching the mapping, numa information is not available as it is read later. So move the assignment of isst_cpu_info[cpu].numa_node before calling _isst_if_get_pci_dev(). Intel-SIG: commit d36d4a1d75d2 platform/x86: ISST: Fix optimization with use of numa. Backport for ISST driver fix Fixes: aa2ddd242572 ("platform/x86: ISST: Use numa node id for cpu pci dev mapping") Signed-off-by: Srinivas Pandruvada Link: https://lore.kernel.org/r/20210727165052.427238-1-srinivas.pandruvada@linux.intel.com Signed-off-by: Hans de Goede [ yingbao jia: amend commit log ] Signed-off-by: yingbao jia --- drivers/platform/x86/intel_speed_select_if/isst_if_common.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/drivers/platform/x86/intel_speed_select_if/isst_if_common.c b/drivers/platform/x86/intel_speed_select_if/isst_if_common.c index 2ea930517914..fd102678c75f 100644 --- a/drivers/platform/x86/intel_speed_select_if/isst_if_common.c +++ b/drivers/platform/x86/intel_speed_select_if/isst_if_common.c @@ -394,6 +394,8 @@ static int isst_if_cpu_online(unsigned int cpu) u64 data; int ret; + isst_cpu_info[cpu].numa_node = cpu_to_node(cpu); + ret = rdmsrl_safe(MSR_CPU_BUS_NUMBER, &data); if (ret) { /* This is not a fatal error on MSR mailbox only I/F */ @@ -412,7 +414,6 @@ static int isst_if_cpu_online(unsigned int cpu) return ret; } isst_cpu_info[cpu].punit_cpu_id = data; - isst_cpu_info[cpu].numa_node = cpu_to_node(cpu); isst_restore_msr_local(cpu); -- Gitee