- 12 Jul, 2018 1 commit
-
-
Reinette Chatre authored
Lockdep is reporting a possible circular locking dependency: ====================================================== WARNING: possible circular locking dependency detected 4.18.0-rc1-test-test+ #4 Not tainted ------------------------------------------------------ user_example/766 is trying to acquire lock: 0000000073479a0f (rdtgroup_mutex){+.+.}, at: pseudo_lock_dev_mmap but task is already holding lock: 000000001ef7a35b (&mm->mmap_sem){++++}, at: vm_mmap_pgoff+0x9f/0x which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&mm->mmap_sem){++++}: _copy_to_user+0x1e/0x70 filldir+0x91/0x100 dcache_readdir+0x54/0x160 iterate_dir+0x142/0x190 __x64_sys_getdents+0xb9/0x170 do_syscall_64+0x86/0x200 entry_SYSCALL_64_after_hwframe+0x49/0xbe -> #1 (&sb->s_type->i_mutex_key#3){++++}: start_creating+0x60/0x100 debugfs_create_dir+0xc/0xc0 rdtgroup_pseudo_lock_create+0x217/0x4d0 rdtgroup_schemata_write+0x313/0x3d0 kernfs_fop_write+0xf0/0x1a0 __vfs_write+0x36/0x190 vfs_write+0xb7/0x190 ksys_write+0x52/0xc0 do_syscall_64+0x86/0x200 entry_SYSCALL_64_after_hwframe+0x49/0xbe -> #0 (rdtgroup_mutex){+.+.}: __mutex_lock+0x80/0x9b0 pseudo_lock_dev_mmap+0x2f/0x170 mmap_region+0x3d6/0x610 do_mmap+0x387/0x580 vm_mmap_pgoff+0xcf/0x110 ksys_mmap_pgoff+0x170/0x1f0 do_syscall_64+0x86/0x200 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: rdtgroup_mutex --> &sb->s_type->i_mutex_key#3 --> &mm->mmap_sem Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&mm->mmap_sem); lock(&sb->s_type->i_mutex_key#3); lock(&mm->mmap_sem); lock(rdtgroup_mutex); *** DEADLOCK *** 1 lock held by user_example/766: #0: 000000001ef7a35b (&mm->mmap_sem){++++}, at: vm_mmap_pgoff+0x9f/0x110 rdtgroup_mutex is already being released temporarily during pseudo-lock region creation to prevent the potential deadlock between rdtgroup_mutex and mm->mmap_sem that is obtained during device_create(). Move the debugfs creation into this area to avoid the same circular dependency. Signed-off-by:Reinette Chatre <reinette.chatre@intel.com> Signed-off-by:
Thomas Gleixner <tglx@linutronix.de> Cc: fenghua.yu@intel.com Cc: tony.luck@intel.com Cc: vikas.shivappa@linux.intel.com Cc: gavin.hindman@intel.com Cc: jithu.joseph@intel.com Cc: hpa@zytor.com Link: https://lkml.kernel.org/r/fffb57f9c6b8285904c9a60cc91ce21591af17fe.1531332480.git.reinette.chatre@intel.com
-
- 03 Jul, 2018 4 commits
-
-
Reinette Chatre authored
When a resource group enters pseudo-locksetup mode it reflects that the platform supports cache pseudo-locking and the resource group is unused, ready to be used for a pseudo-locked region. Until it is set up as a pseudo-locked region the resource group is "locked down" such that no new tasks or cpus can be assigned to it. This is accomplished in a user visible way by making the cpus, cpus_list, and tasks resctrl files inaccassible (user cannot read from or write to these files). When the resource group changes to pseudo-locked mode it represents a cache pseudo-locked region. While not appropriate to make any changes to the cpus assigned to this region it is useful to make it easy for the user to see which cpus are associated with the pseudo-locked region. Modify the permissions of the cpus/cpus_list file when the resource group changes to pseudo-locked mode to support reading (not writing). The information presented to the user when reading the file are the cpus associated with the pseudo-locked region. Signed-off-by:
-
Reinette Chatre authored
As the mode of a resource group changes, the operations it can support may also change. One way in which the supported operations are managed is to modify the permissions of the files within the resource group's resctrl directory. At the moment only two possible permissions are supported: the default permissions or no permissions in support for when the operation is "locked down". It is possible where an operation on a resource group may have more possibilities. For example, if by default changes can be made to the resource group by writing to a resctrl file while the current settings can be obtained by reading from the file, then it may be possible that in another mode it is only possible to read the current settings, and not change them. Make it possible to modify some of the permissions of a resctrl file in support of a more flexible way to manage the operations on a resource group. In this preparation work the original behavior is maintained where all permissions are restored. Signed-off-by:
-
Reinette Chatre authored
When a resource group enters pseudo-locksetup mode a pseudo_lock_region is associated with it. When the user writes to the resource group's schemata file the CBM of the requested pseudo-locked region is entered into the pseudo_lock_region struct. If any part of pseudo-lock region creation fails the resource group will remain in pseudo-locksetup mode with the pseudo_lock_region associated with it. In case of failure during pseudo-lock region creation care needs to be taken to ensure that the pseudo_lock_region struct associated with the resource group is cleared from any pseudo-locking data - especially the CBM. This is because the existence of a pseudo_lock_region struct with a CBM is significant in other areas of the code, for example, the display of bit_usage and initialization of a new resource group. Fix the error path of pseudo-lock region creation to ensure that the pseudo_lock_region struct is cleared at each error exit. Fixes: 018961ae ("x86/intel_rdt: Pseudo-lock region creation/removal core") Signed-off-by:
-
Reinette Chatre authored
The pseudo_lock_region_clear() function is moved to earlier in the file in preparation for its use in functions that currently appear before it. No functional change. Signed-off-by:
-
- 24 Jun, 2018 4 commits
-
-
Reinette Chatre authored
Deeper C-states impact cache content through shrinking of the cache or flushing entire cache to memory before reducing power to the cache. Deeper C-states will thus negatively impact the pseudo-locked regions. To avoid impacting pseudo-locked regions C-states are limited on pseudo-locked region creation so that cores associated with the pseudo-locked region are prevented from entering deeper C-states. This is accomplished by requesting a CPU latency target which will prevent the core from entering C6 across all supported platforms. Signed-off-by:
-
Reinette Chatre authored
Broadwell microarchitecture supports pseudo-locking. Add support for the L3 cache related performance events of these systems so that the success of pseudo-locking can be measured more accurately on these platforms. Signed-off-by:
-
Reinette Chatre authored
Intel Goldmont processors supports non-architectural precise events that can be used to give us more insight into the success of L2 cache pseudo-locking on these platforms. Introduce a new measurement trigger that will enable two precise events, MEM_LOAD_UOPS_RETIRED.L2_HIT and MEM_LOAD_UOPS_RETIRED.L2_MISS, while accessing pseudo-locked data. A new tracepoint, pseudo_lock_l2, is created to make these results visible to the user. Signed-off-by:
-
Reinette Chatre authored
After a pseudo-locked region is created it needs to be made available to user space for usage. A character device supporting mmap() is created for each pseudo-locked region. A user space application can now use mmap() system call to map pseudo-locked region into its virtual address space. Signed-off-by:
-
- 23 Jun, 2018 31 commits
-
-
Reinette Chatre authored
There is no simple yes/no test to determine if pseudo-locking was successful. In order to test pseudo-locking we expose a debugfs file for each pseudo-locked region that will record the latency of reading the pseudo-locked memory at a stride of 32 bytes (hardcoded). These numbers will give us an idea of locking was successful or not since they will reflect cache hits and cache misses (hardware prefetching is disabled during the test). The new debugfs file "pseudo_lock_measure" will, when the pseudo_lock_mem_latency tracepoint is enabled, record the latency of accessing each cache line twice. Kernel tracepoints offer us histograms (when CONFIG_HIST_TRIGGERS is enabled) that is a simple way to visualize the memory access latency and immediately see any cache misses. For example, the hist trigger below before trigger of the measurement will display the memory access latency and instances at each latency: echo 'hist:keys=latency' > /sys/kernel/debug/tracing/events/resctrl/\ pseudo_lock_mem_latency/trigger echo 1 > /sys/kernel/debug/tracing/events/resctrl/pseudo_lock_mem_latency/enable echo 1 > /sys/kernel/debug/resctrl/<newlock>/pseudo_lock_measure echo 0 > /sys/kernel/debug/tracing/events/resctrl/pseudo_lock_mem_latency/enable cat /sys/kernel/debug/tracing/events/resctrl/pseudo_lock_mem_latency/hist Signed-off-by: -
Reinette Chatre authored
In preparation for support of debugging of RDT sub features the user can now enable a RDT debugfs region. The debug area is always enabled when CONFIG_DEBUG_FS is set as advised in http://lkml.kernel.org/r/20180523080501.GA6822@kroah.com Also from same discussion in above linked email, no error checking on the debugfs creation return value since code should not behave differently when debugging passes or fails. Even on failure the returned value can be passed safely to other debugfs calls. Signed-off-by:
-
Reinette Chatre authored
The RDT system's initialization does not have the corresponding exit handling to ensure everything initialized on load is cleaned up also. Introduce the cleanup routines that complement all initialization. This includes the removal of a duplicate rdtgroup_init() declaration. Signed-off-by:
-
Reinette Chatre authored
Information about resources as well as resource groups are contained in a variety of resctrl files. Now that pseudo-locked regions can be created the files can be updated to present appropriate information to the user. Update the resource group's schemata file to show only the information of the pseudo-locked region. Update the resource group's size file to show the size in bytes of only the pseudo-locked region. Update the bit_usage file to use the letter 'P' for all pseudo-locked regions. Signed-off-by:
-
Reinette Chatre authored
The user triggers the creation of a pseudo-locked region when writing a valid schemata to the schemata file of a resource group in the pseudo-locksetup mode. A valid schemata is one that: (1) does not overlap with any other resource group, (2) does not involve a cache that already contains a pseudo-locked region within its hierarchy. After a valid schemata is parsed the system is programmed to associate the to be pseudo-lock bitmask with the closid associated with the resource group. With the system set up the pseudo-locked region can be created. Signed-off-by:
-
Reinette Chatre authored
The user requests a pseudo-locked region by providing a schemata to a resource group that is in the pseudo-locksetup mode. This is the functionality that consumes the parsed user data and creates the pseudo-locked region. First, required information is deduced from user provided data. This includes, how much memory does the requested bitmask represent, which CPU the requested region is associated with, and what is the cache line size of that cache (to learn the stride needed for locking). Second, a contiguous block of memory matching the requested bitmask is allocated. Finally, pseudo-locking is performed. The resource group already has the allocation that reflects the requested bitmask. With this class of service active and interference minimized, the allocated memory is loaded into the cache. Signed-off-by:
-
Reinette Chatre authored
Knowing the model specific prefetch disable bits is required to support cache pseudo-locking because the hardware prefetchers need to be disabled when the kernel memory is pseudo-locked to cache. We add these bits only for platforms known to support cache pseudo-locking. When the user requests locksetup mode to be entered it will fail if the prefetch disabling bits are not known for the platform. Signed-off-by:
-
Reinette Chatre authored
A pseudo-locked region does not have a class of service associated with it and thus not tracked in the array of control values maintained as part of the domain. Even so, when the user provides a new bitmask for another resource group it needs to be checked for interference with existing pseudo-locked regions. Additionally only one pseudo-locked region can be created in any cache hierarchy. Introduce two utilities in support of above scenarios: (1) a utility that can be used to test if a given capacity bitmask overlaps with any pseudo-locked regions associated with a particular cache instance, (2) a utility that can be used to test if a pseudo-locked region exists within a particular cache hierarchy. Signed-off-by:
-
Reinette Chatre authored
Resource groups used for pseudo-locking do not require the same work on removal as the other resource groups. The resource group removal is split in two in preparation for support of pseudo-locking resource groups. A single re-ordering occurs - the setting of the rdtgrp flag is moved to later. This flag is not used by any of the code between its original and new location. Signed-off-by:
-
Reinette Chatre authored
The user can request entering pseudo-locksetup mode by writing "pseudo-locksetup" to the mode file. Act on this request as well as support switching from a pseudo-locksetup mode (before pseudo-locked mode was entered). It is not supported to modify the mode once pseudo-locked mode has been entered. The schemata reflects the new mode by adding "uninitialized" to all resources. The size resctrl file reports zero for all cache domains in support of the uninitialized nature. Since there are no users of this class of service its allocations can be ignored when searching for appropriate default allocations for new resource groups. For the same reason resource groups in pseudo-locksetup mode are not considered when testing if new resource groups may overlap. Signed-off-by:
-
Reinette Chatre authored
The locksetup mode is the way in which the user communicates that the resource group will be used for a pseudo-locked region. Locksetup mode should thus ensure that all restrictions on a resource group are met before locksetup mode can be entered. The resource group should also be configured to ensure that it cannot be modified in unsupported ways when a pseudo-locked region. Introduce the support where the request for entering locksetup mode can be validated. This includes: CDP is not active, no cpus or tasks are assigned to the resource group, monitoring is not in progress on the resource group. Once the resource group is determined ready for a pseudo-locked region it is configured to not allow future changes to these properties. Signed-off-by:
-
Reinette Chatre authored
A pseudo-locked region is introduced representing an instance of a pseudo-locked cache region. Each cache instance (domain) can support one pseudo-locked region. Similarly a resource group can be used for one pseudo-locked region. Include a pointer to a pseudo-locked region from the domain and resource group structures. Signed-off-by:
-
Reinette Chatre authored
When a resource group is pseudo-locked it is orphaned without a class of service associated with it. We thus do not want any monitoring in progress on a resource group that will be used for pseudo-locking. Introduce a test that can be used to determine if pseudo-locking in progress on a resource group. Temporarily mark it as unused to avoid compile warnings until it is used. Signed-off-by:
-
Reinette Chatre authored
In support of Cache Pseudo-Locking we need to restrict access to specific resctrl files to protect the state of a resource group used for pseudo-locking from being changed in unsupported ways. Introduce two utilities that can be used to either restrict or restore the access to all files irrelevant to cache pseudo-locking when pseudo-locking in progress for the resource group. At this time introduce a new source file, intel_rdt_pseudo_lock.c, that will contain most of the code related to cache pseudo-locking. Temporarily mark these new functions as unused to silence compile warnings until they are used. Signed-off-by:
-
Reinette Chatre authored
We intend to modify file permissions to make the "tasks", "cpus", and "cpus_list" not accessible to the user when cache pseudo-locking in progress. Even so, it is still possible for the user to force the file permissions (using chmod) to make them writeable. Similarly, directory permissions will be modified to prevent future monitor group creation but the user can override these restrictions also. Add additional checks to the files we intend to restrict to ensure that no modifications from user space are attempted while setting up a pseudo-locking or after a pseudo-locked region is set up. Signed-off-by:
-
Reinette Chatre authored
When a resource group is used for Cache Pseudo-Locking then the region of cache ends up being orphaned with no class of service referring to it. The resctrl files intended to manage how the classes of services are utilized thus become irrelevant. The fact that a resctrl file is not relevant can be communicated to the user by setting all of its permissions to zero. That is, its read, write, and execute permissions are unset for all users. Introduce two utilities, rdtgroup_kn_mode_restrict() and rdtgroup_kn_mode_restore(), that can be used to restrict and restore the permissions of a file or directory belonging to a resource group. Signed-off-by:
-
Reinette Chatre authored
In considering changes to a resource group it becomes necessary to know whether tasks have been assigned to the resource group in question. Introduce a new utility that can be used to check if any tasks have been assigned to a particular resource group. Signed-off-by:
-
Reinette Chatre authored
By default, if the opener has CAP_DAC_OVERRIDE, a kernfs file can be opened regardless of RW permissions. Writing to a kernfs file will thus succeed even if permissions are 0000. It's required to restrict the actions that can be performed on a resource group from userspace based on the mode of the resource group. This restriction will be done through a modification of the file permissions. That is, for example, if a resource group is locked then the user cannot add tasks to the resource group. For this restriction through file permissions to work it has to be ensured that the permissions are always respected. To do so the resctrl filesystem is created with the KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK flag that will result in open(2) failing with -EACCESS regardless of CAP_DAC_OVERRIDE if the permission does not have the respective read or write access. Signed-off-by:
-
Reinette Chatre authored
The two modes used to manage Cache Pseudo-Locked regions are introduced. A resource group is assigned "pseudo-locksetup" mode when the user indicates that this resource group will be used for a Cache Pseudo-Locked region. When the Cache Pseudo-Locked region has been set up successfully after the user wrote the requested schemata to the "schemata" file, then the mode will automatically changed to "pseudo-locked". The user is not able to modify the mode to "pseudo-locked" by writing "pseudo-locked" to the "mode" file directly. Signed-off-by:
-
Reinette Chatre authored
Add description of Cache Pseudo-Locking feature, its interface, as well as an example of its usage. Signed-off-by:
-
Reinette Chatre authored
The schemata file displays the allocations associated with each domain of each resource. The syntax of this file reflects the capacity bitmask (CBM) of the actual allocation. In order to determine the actual size of an allocation the user needs to dig through three different files to query the variables needed to compute it (the cache size, the CBM length, and the schemata). Introduce a new file "size" associated with each resource group that will mirror the schemata file syntax and display the size in bytes of each allocation. Signed-off-by:
-
Reinette Chatre authored
With cache regions now explicitly marked as "shareable" or "exclusive" we would like to communicate to the user how portions of the cache are used. Introduce "bit_usage" that indicates for each resource how portions of the cache are configured to be used. To assist the user to distinguish whether the sharing is from software or hardware we add the following annotation: 0 - currently unused X - currently available for sharing and used by software and hardware H - currently used by hardware only but available for software use S - currently used and shareable by software only E - currently used exclusively by one resource group Signed-off-by:
-
Reinette Chatre authored
When the administrator requests a change in a resource group's schemata we have to ensure that the new schemata respects the current resource group as well as the other active resource groups' schemata. The new schemata is not allowed to overlap with the schemata of any exclusive resource groups. Similarly, if the resource group being changed is exclusive then its new schemata is not allowed to overlap with any schemata of any other active resource group. Signed-off-by:
-
Reinette Chatre authored
Each resource is associated with a configurable callback that should be used to parse the information provided for the particular resource from user space. In addition to the resource and domain pointers this callback is provided with just the character buffer being parsed. In support of flexible parsing the callback is modified to support a void pointer as argument. This enables resources that need more data than just the user provided data to pass its required data to the callback without affecting the signatures for the callbacks of all the other resources. Signed-off-by:
-
Reinette Chatre authored
cbm_validate() receives a pointer to the variable that will be initialized with a validated capacity bitmask. The pointer points to a variable of type unsigned long that is immediately assigned to a variable of type u32 by the caller on return from cbm_validate(). Let cbm_validate() initialize a variable of type u32 directly. At this time also change tha variable name "data" within parse_cbm() to a name more reflective of the content: "cbm_val". This frees up the generic "data" to be used later when it is indeed used for a collection of input. Signed-off-by:
-
Reinette Chatre authored
The new "mode" file now accepts "exclusive" that means that the allocations of this resource group cannot be shared. Enable users to modify a resource group's mode to "exclusive". To succeed it is required that there is no overlap between resource group's current schemata and that of all the other active resource groups as well as cache regions potentially used by other hardware entities. Signed-off-by:
-
Reinette Chatre authored
At the moment all allocations are shareable. There is no way for a user to designate that an allocation associated with a resource group cannot be shared by another. Introduce the new mode "exclusive". When a resource group is marked as such it implies that no overlap is allowed between its allocation and that of another resource group. Signed-off-by:
-
Reinette Chatre authored
Currently when a new resource group is created its allocations would be those that belonged to the resource group to which its closid belonged previously. That is, we can encounter a case like: mkdir newgroup cat newgroup/schemata L2:0=ff;1=ff echo 'L2:0=0xf0;1=0xf0' > newgroup/schemata cat newgroup/schemata L2:0=0xf0;1=0xf0 rmdir newgroup mkdir newnewgroup cat newnewgroup/schemata L2:0=0xf0;1=0xf0 When the new group is created it would be reasonable to expect its allocations to be initialized with all regions that it can possibly use. At this time these regions would be all that are shareable by other resource groups as well as regions that are not currently used. If the available cache region is found to be non-contiguous the available region is adjusted to enforce validity. When a new resource group is created the hardware is initialized with these new default allocations. Signed-off-by:
-
Reinette Chatre authored
In support of the work done to enable resource groups to have different modes some static functions need to be available for sharing amongst all RDT components. Signed-off-by:
-
Reinette Chatre authored
During CAT feature discovery the capacity bitmasks (CBMs) associated with all the classes of service are initialized to all ones, even if the class of service is not in use. Introduce a test that can be used to determine if a class of service is in use. This test enables code interested in parsing the CBMs to know if its values are meaningful or can be ignored. Temporarily mark the function as unused to silence compile warnings until it is used. Signed-off-by:
-
Reinette Chatre authored
A new resctrl file "mode" associated with each resource group is introduced. This file will display the resource group's current mode and an administrator can also use it to modify the resource group's mode. Only shareable mode is currently supported. Signed-off-by:
-
