- 11 Feb, 2014 4 commits
-
-
Tejun Heo authored
Currently cgroup uses combination of inode->i_mutex'es and cgroup_mutex for synchronization. With the scheduled kernfs conversion, i_mutex'es will be removed. Unfortunately, just using cgroup_mutex isn't possible. All kernfs file and syscall operations, most of which require grabbing cgroup_mutex, will be called with kernfs active ref held and, if we try to perform kernfs removals under cgroup_mutex, it can deadlock as kernfs_remove() tries to drain the target node. Let's introduce a new outer mutex, cgroup_tree_mutex, which protects stuff used during hierarchy changing operations - cftypes and all the operations which may affect the cgroupfs. It also covers css association and iteration. This allows cgroup_css(), for_each_css() and other css iterators to be called under cgroup_tree_mutex. The new mutex will nest above both kernfs's active ref protection and cgroup_mutex. By protecting tree modifications with a separate outer mutex, we can get rid of the forementioned deadlock condition. Actual file additions and removals now require cgroup_tree_mutex instead of cgroup_mutex. Currently, cgroup_tree_mutex is never used without cgroup_mutex; however, we'll soon add hierarchy modification sections which are only protected by cgroup_tree_mutex. In the future, we might want to make the locking more granular by better splitting the coverages of the two mutexes. For now, this should do. v2: Rebased on top of 0ab02ca8 ("cgroup: protect modifications to cgroup_idr with cgroup_mutex"). Signed-off-by:
Tejun Heo <tj@kernel.org> Acked-by:
Li Zefan <lizefan@huawei.com>
-
Tejun Heo authored
css_from_dir() returns the matching css (cgroup_subsys_state) given a dentry and subsystem. The function doesn't pin the css before returning and requires the caller to be holding RCU read lock or cgroup_mutex and handling pinning on the caller side. Given that users of the function are likely to want to pin the returned css (both existing users do) and that getting and putting css's are very cheap, there's no reason for the interface to be tricky like this. Rename css_from_dir() to css_tryget_from_dir() and make it try to pin the found css and return it only if pinning succeeded. The callers are updated so that they no longer do RCU locking and pinning around the function and just use the returned css. This will also ease converting cgroup to kernfs. Signed-off-by:
Michal Hocko <mhocko@suse.cz> Acked-by:
-
Tejun Heo authored
Pull for-3.14-fixes to receive 0ab02ca8 ("cgroup: protect modifications to cgroup_idr with cgroup_mutex") prior to kernfs conversion series to avoid non-trivial conflicts. Signed-off-by:
-
Li Zefan authored
Setup cgroupfs like this: # mount -t cgroup -o cpuacct xxx /cgroup # mkdir /cgroup/sub1 # mkdir /cgroup/sub2 Then run these two commands: # for ((; ;)) { mkdir /cgroup/sub1/tmp && rmdir /mnt/sub1/tmp; } & # for ((; ;)) { mkdir /cgroup/sub2/tmp && rmdir /mnt/sub2/tmp; } & After seconds you may see this warning: ------------[ cut here ]------------ WARNING: CPU: 1 PID: 25243 at lib/idr.c:527 sub_remove+0x87/0x1b0() idr_remove called for id=6 which is not allocated. ... Call Trace: [<ffffffff8156063c>] dump_stack+0x7a/0x96 [<ffffffff810591ac>] warn_slowpath_common+0x8c/0xc0 [<ffffffff81059296>] warn_slowpath_fmt+0x46/0x50 [<ffffffff81300aa7>] sub_remove+0x87/0x1b0 [<ffffffff810f3f02>] ? css_killed_work_fn+0x32/0x1b0 [<ffffffff81300bf5>] idr_remove+0x25/0xd0 [<ffffffff810f2bab>] cgroup_destroy_css_killed+0x5b/0xc0 [<ffffffff810f4000>] css_killed_work_fn+0x130/0x1b0 [<ffffffff8107cdbc>] process_one_work+0x26c/0x550 [<ffffffff8107eefe>] worker_thread+0x12e/0x3b0 [<ffffffff81085f96>] kthread+0xe6/0xf0 [<ffffffff81570bac>] ret_from_fork+0x7c/0xb0 ---[ end trace 2d1577ec10cf80d0 ]--- It's because allocating/removing cgroup ID is not properly synchronized. The bug was introduced when we converted cgroup_ida to cgroup_idr. While synchronization is already done inside ida_simple_{get,remove}(), users are responsible for concurrent calls to idr_{alloc,remove}(). tj: Refreshed on top of b58c8998 ("cgroup: fix error return from cgroup_create()"). Fixes: 4e96ee8e ("cgroup: convert cgroup_ida to cgroup_idr") Cc: <stable@vger.kernel.org> #3.12+ Reported-by:
-
- 08 Feb, 2014 16 commits
-
-
Tejun Heo authored
Pending kernfs conversion depends on kernfs improvements in driver-core-next. Pull it into for-3.15. Signed-off-by: -
Tejun Heo authored
Pending kernfs conversion depends on fixes in for-3.14-fixes. Pull it into for-3.15. Signed-off-by: -
Tejun Heo authored
cgroup_root_mutex was added to avoid deadlock involving namespace_sem via cgroup_show_options(). It added a lot of overhead for the small purpose of it and, because it's nested under cgroup_mutex, it has very limited usefulness. The previous patch made cgroup_show_options() not use cgroup_root_mutex, so nobody needs it anymore. Remove it. Signed-off-by:
-
Tejun Heo authored
cgroup_show_options() grabs cgroup_root_mutex to protect the options changing while printing; however, holding root_mutex or not doesn't really make much difference for the function. subsys_mask can be atomically tested and most of the options aren't allowed to change anyway once mounted. The only field which needs synchronization is ->release_agent_path. This patch introduces a dedicated spinlock to synchronize accesses to the field and drops cgroup_root_mutex locking from cgroup_show_options(). The next patch will remove cgroup_root_mutex. Signed-off-by:
-
Tejun Heo authored
It's no longer referenced outside cgroup core, so renaming is easy. Let's rename it for consistency & brevity. This patch is pure rename. Signed-off-by:
-
Tejun Heo authored
cgroup_subsys is a bit messier than it needs to be. * The name of a subsys can be different from its internal identifier defined in cgroup_subsys.h. Most subsystems use the matching name but three - cpu, memory and perf_event - use different ones. * cgroup_subsys_id enums are postfixed with _subsys_id and each cgroup_subsys is postfixed with _subsys. cgroup.h is widely included throughout various subsystems, it doesn't and shouldn't have claim on such generic names which don't have any qualifier indicating that they belong to cgroup. * cgroup_subsys->subsys_id should always equal the matching cgroup_subsys_id enum; however, we require each controller to initialize it and then BUG if they don't match, which is a bit silly. This patch cleans up cgroup_subsys names and initialization by doing the followings. * cgroup_subsys_id enums are now postfixed with _cgrp_id, and each cgroup_subsys with _cgrp_subsys. * With the above, renaming subsys identifiers to match the userland visible names doesn't cause any naming conflicts. All non-matching identifiers are renamed to match the official names. cpu_cgroup -> cpu mem_cgroup -> memory perf -> perf_event * controllers no longer need to initialize ->subsys_id and ->name. They're generated in cgroup core and set automatically during boot. * Redundant cgroup_subsys declarations removed. * While updating BUG_ON()s in cgroup_init_early(), convert them to WARN()s. BUGging that early during boot is stupid - the kernel can't print anything, even through serial console and the trap handler doesn't even link stack frame properly for back-tracing. This patch doesn't introduce any behavior changes. v2: Rebased on top of fe1217c4 ("net: net_cls: move cgroupfs classid handling into core"). Signed-off-by:
Neil Horman <nhorman@tuxdriver.com> Acked-by:
"David S. Miller" <davem@davemloft.net> Acked-by:
"Rafael J. Wysocki" <rjw@rjwysocki.net> Acked-by:
Peter Zijlstra <peterz@infradead.org> Acked-by:
Aristeu Rozanski <aris@redhat.com> Acked-by:
Ingo Molnar <mingo@redhat.com> Acked-by:
-
Tejun Heo authored
With module supported dropped from net_prio, no controller is using cgroup module support. None of actual resource controllers can be built as a module and we aren't gonna add new controllers which don't control resources. This patch drops module support from cgroup. * cgroup_[un]load_subsys() and cgroup_subsys->module removed. * As there's no point in distinguishing IS_BUILTIN() and IS_MODULE(), cgroup_subsys.h now uses IS_ENABLED() directly. * enum cgroup_subsys_id now exactly matches the list of enabled controllers as ordered in cgroup_subsys.h. * cgroup_subsys[] is now a contiguously occupied array. Size specification is no longer necessary and dropped. * for_each_builtin_subsys() is removed and for_each_subsys() is updated to not require any locking. * module ref handling is removed from rebind_subsystems(). * Module related comments dropped. v2: Rebased on top of fe1217c4 ("net: net_cls: move cgroupfs classid handling into core"). v3: Added {} around the if (need_forkexit_callback) block in cgroup_post_fork() for readability as suggested by Li. Signed-off-by:
-
Tejun Heo authored
net_prio is the only cgroup which is allowed to be built as a module. The savings from allowing one controller to be built as a module are tiny especially given that cgroup module support itself adds quite a bit of complexity. Given that none of other controllers has much chance of being made a module and that we're unlikely to add new modular controllers, the added complexity is simply not justifiable. As a first step to drop cgroup module support, this patch changes the config option to bool from tristate and drops module related code from it. Also, while an earlier commit fe1217c4 ("net: net_cls: move cgroupfs classid handling into core") dropped module support from net_cls cgroup, it retained a call to cgroup_load_subsys(), which is noop for built-in controllers. Drop it along with init_netclassid_cgroup(). v2: Removed modular version of task_netprioidx() in include/net/netprio_cgroup.h as suggested by Li Zefan. v3: Rebased on top of fe1217c4 ("net: net_cls: move cgroupfs classid handling into core"). net_cls cgroup part is mostly dropped except for removal of init_netclassid_cgroup(). Signed-off-by:
-
Tejun Heo authored
cgroup_cfts_commit() walks the cgroup hierarchy that the target subsystem is attached to and tries to apply the file changes. Due to the convolution with inode locking, it can't keep cgroup_mutex locked while iterating. It currently holds only RCU read lock around the actual iteration and then pins the found cgroup using dget(). Unfortunately, this is incorrect. Although the iteration does check cgroup_is_dead() before invoking dget(), there's nothing which prevents the dentry from going away inbetween. Note that this is different from the usual css iterations where css_tryget() is used to pin the css - css_tryget() tests whether the css can be pinned and fails if not. The problem can be solved by simply holding cgroup_mutex instead of RCU read lock around the iteration, which actually reduces LOC. Signed-off-by:
-
Tejun Heo authored
cgroup_create() was returning 0 after allocation failures. Fix it. Signed-off-by:
-
Tejun Heo authored
When cgroup_mount() fails to allocate an id for the root, it didn't set ret before jumping to unlock_drop ending up returning 0 after a failure. Fix it. Signed-off-by:
-
Tejun Heo authored
As sysfs was kernfs's only user, kernfs has been piggybacking on CONFIG_SYSFS; however, kernfs is scheduled to grow a new user very soon. Introduce a separate config option CONFIG_KERNFS which is to be selected by kernfs users. Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
Tejun Heo authored
Currently, kobject is invoking kernfs_enable_ns() directly. This is fine now as sysfs and kernfs are enabled and disabled together. If sysfs is disabled, kernfs_enable_ns() is switched to dummy implementation too and everything is fine; however, kernfs will soon have its own config option CONFIG_KERNFS and !SYSFS && KERNFS will be possible, which can make kobject call into non-dummy kernfs_enable_ns() with NULL kernfs_node pointers leading to an oops. Introduce sysfs_enable_ns() which is a wrapper around kernfs_enable_ns() so that it can be made a noop depending only on CONFIG_SYSFS regardless of the planned CONFIG_KERNFS. Signed-off-by:
Fengguang Wu <fengguang.wu@intel.com> Signed-off-by:
-
Tejun Heo authored
kernfs_node->parent and ->name are currently marked as "published" indicating that kernfs users may access them directly; however, those fields may get updated by kernfs_rename[_ns]() and unrestricted access may lead to erroneous values or oops. Protect ->parent and ->name updates with a irq-safe spinlock kernfs_rename_lock and implement the following accessors for these fields. * kernfs_name() - format the node's name into the specified buffer * kernfs_path() - format the node's path into the specified buffer * pr_cont_kernfs_name() - pr_cont a node's name (doesn't need buffer) * pr_cont_kernfs_path() - pr_cont a node's path (doesn't need buffer) * kernfs_get_parent() - pin and return a node's parent All can be called under any context. The recursive sysfs_pathname() in fs/sysfs/dir.c is replaced with kernfs_path() and sysfs_rename_dir_ns() is updated to use kernfs_get_parent() instead of dereferencing parent directly. v2: Dummy definition of kernfs_path() for !CONFIG_KERNFS was missing static inline making it cause a lot of build warnings. Add it. Signed-off-by: -
Tejun Heo authored
Implement helpers to determine node from dentry and root from super_block. Also add a kernfs_rename_ns() wrapper which assumes NULL namespace. These generally make sense and will be used by cgroup. v2: Some dummy implementations for !CONFIG_SYSFS was missing. Fixed. Reported by kbuild test robot. Signed-off-by: -
Tejun Heo authored
Add a private data field to be used by kernfs file operations. This generally makes sense and will be used by cgroup. Signed-off-by:
-
- 07 Feb, 2014 20 commits
-
-
Tejun Heo authored
A write to a kernfs_node is buffered through a kernel buffer. Writes <= PAGE_SIZE are performed atomically, while larger ones are executed in PAGE_SIZE chunks. While this is enough for sysfs, cgroup which is scheduled to be converted to use kernfs needs a bit more control over it. This patch adds kernfs_ops->atomic_write_len. If not set (zero), the behavior stays the same. If set, writes upto the size are executed atomically and larger writes are rejected with -E2BIG. A different implementation strategy would be allowing configuring chunking size while making the original write size available to the write method; however, such strategy, while being more complicated, doesn't really buy anything. If the write implementation has to handle chunking, the specific chunk size shouldn't matter all that much. Signed-off-by:
-
Tejun Heo authored
Currently, kernfs_nodes are made visible to userland on creation, which makes it difficult for kernfs users to atomically succeed or fail creation of multiple nodes. In addition, if something fails after creating some nodes, the created nodes might already be in use and their active refs need to be drained for removal, which has the potential to introduce tricky reverse locking dependency on active_ref depending on how the error path is synchronized. This patch introduces per-root flag KERNFS_ROOT_CREATE_DEACTIVATED. If set, all nodes under the root are created in the deactivated state and stay invisible to userland until explicitly enabled by the new kernfs_activate() API. Also, nodes which have never been activated are guaranteed to bypass draining on removal thus allowing error paths to not worry about lockding dependency on active_ref draining. Signed-off-by:
-
Tejun Heo authored
kernfs_iop_lookup(), kernfs_dir_pos() and kernfs_dir_next_pos() were missing kernfs_active() tests before using the found kernfs_node. As deactivated state is currently visible only while a node is being removed, this doesn't pose an actual problem. e.g. lookup succeeding on a deactivated node doesn't harm anything as the eventual file operations are gonna fail and those failures are indistinguishible from the cases in which the lookups had happened before the node was deactivated. However, we're gonna allow new nodes to be created deactivated and then activated explicitly by the kernfs user when it sees fit. This is to support atomically making multiple nodes visible to userland and thus those nodes must not be visible to userland before activated. Let's plug the lookup and readdir holes so that deactivated nodes are invisible to userland. Signed-off-by:
-
Tejun Heo authored
Add two super_block related syscall callbacks ->remount_fs() and ->show_options() to kernfs_syscall_ops. These simply forward the matching super_operations. Signed-off-by:
-
Tejun Heo authored
We're gonna need non-dir syscall callbacks, which will make dir_ops a misnomer. Let's rename kernfs_dir_ops to kernfs_syscall_ops. This is pure rename. Signed-off-by:
-
Tejun Heo authored
kernfs_dir_ops are currently being invoked without any active reference, which makes it tricky for the invoked operations to determine whether the objects associated those nodes are safe to access and will remain that way for the duration of such operations. kernfs already has active_ref mechanism to deal with this which makes the removal of a given node the synchronization point for gating the file operations. There's no reason for dir_ops to be any different. Update the dir_ops handling so that active_ref is held while the dir_ops are executing. This guarantees that while a dir_ops is executing the target nodes stay alive. As kernfs_dir_ops doesn't have any in-kernel user at this point, this doesn't affect anybody. Signed-off-by:
-
Tejun Heo authored
All device_schedule_callback_owner() users are converted to use device_remove_file_self(). Remove now unused {sysfs|device}_schedule_callback_owner(). Signed-off-by: -
Tejun Heo authored
driver-core now supports synchrnous self-deletion of attributes and the asynchrnous removal mechanism is scheduled for removal. Use it instead of device_schedule_callback(). * Conversions in arch/s390/pci/pci_sysfs.c and drivers/s390/block/dcssblk.c are straightforward. * drivers/s390/cio/ccwgroup.c is a bit more tricky because ccwgroup_notifier() was (ab)using device_schedule_callback() to purely obtain a process context to kick off ungroup operation which may block from a notifier callback. Rename ccwgroup_ungroup_callback() to ccwgroup_ungroup() and make it take ccwgroup_device * instead. The new function is now called directly from ccwgroup_ungroup_store(). ccwgroup_notifier() chain is updated to explicitly bounce through ccwgroup_device->ungroup_work. This also removes possible failure from memory pressure. Only compile-tested. Signed-off-by:
-
Tejun Heo authored
driver-core now supports synchrnous self-deletion of attributes and the asynchrnous removal mechanism is scheduled for removal. Use it instead of device_schedule_callback(). This makes "delete" behave synchronously. Signed-off-by:
-
Tejun Heo authored
driver-core now supports synchrnous self-deletion of attributes and the asynchrnous removal mechanism is scheduled for removal. Use it instead of device_schedule_callback(). This makes "remove" behave synchronously. Signed-off-by:
-
Tejun Heo authored
Sometimes it's necessary to implement a node which wants to delete nodes including itself. This isn't straightforward because of kernfs active reference. While a file operation is in progress, an active reference is held and kernfs_remove() waits for all such references to drain before completing. For a self-deleting node, this is a deadlock as kernfs_remove() ends up waiting for an active reference that itself is sitting on top of. This currently is worked around in the sysfs layer using sysfs_schedule_callback() which makes such removals asynchronous. While it works, it's rather cumbersome and inherently breaks synchronicity of the operation - the file operation which triggered the operation may complete before the removal is finished (or even started) and the removal may fail asynchronously. If a removal operation is immmediately followed by another operation which expects the specific name to be available (e.g. removal followed by rename onto the same name), there's no way to make the latter operation reliable. The thing is there's no inherent reason for this to be asynchrnous. All that's necessary to do this synchronous is a dedicated operation which drops its own active ref and deactivates self. This patch implements kernfs_remove_self() and its wrappers in sysfs and driver core. kernfs_remove_self() is to be called from one of the file operations, drops the active ref the task is holding, removes the self node, and restores active ref to the dead node so that the ref is balanced afterwards. __kernfs_remove() is updated so that it takes an early exit if the target node is already fully removed so that the active ref restored by kernfs_remove_self() after removal doesn't confuse the deactivation path. This makes implementing self-deleting nodes very easy. The normal removal path doesn't even need to be changed to use kernfs_remove_self() for the self-deleting node. The method can invoke kernfs_remove_self() on itself before proceeding the normal removal path. kernfs_remove() invoked on the node by the normal deletion path will simply be ignored. This will replace sysfs_schedule_callback(). A subtle feature of sysfs_schedule_callback() is that it collapses multiple invocations - even if multiple removals are triggered, the removal callback is run only once. An equivalent effect can be achieved by testing the return value of kernfs_remove_self() - only the one which gets %true return value should proceed with actual deletion. All other instances of kernfs_remove_self() will wait till the enclosing kernfs operation which invoked the winning instance of kernfs_remove_self() finishes and then return %false. This trivially makes all users of kernfs_remove_self() automatically show correct synchronous behavior even when there are multiple concurrent operations - all "echo 1 > delete" instances will finish only after the whole operation is completed by one of the instances. Note that manipulation of active ref is implemented in separate public functions - kernfs_[un]break_active_protection(). kernfs_remove_self() is the only user at the moment but this will be used to cater to more complex cases. v2: For !CONFIG_SYSFS, dummy version kernfs_remove_self() was missing and sysfs_remove_file_self() had incorrect return type. Fix it. Reported by kbuild test bot. v3: kernfs_[un]break_active_protection() separated out from kernfs_remove_self() and exposed as public API. Signed-off-by: -
Tejun Heo authored
KERNFS_REMOVED is used to mark half-initialized and dying nodes so that they don't show up in lookups and deny adding new nodes under or renaming it; however, its role overlaps that of deactivation. It's necessary to deny addition of new children while removal is in progress; however, this role considerably intersects with deactivation - KERNFS_REMOVED prevents new children while deactivation prevents new file operations. There's no reason to have them separate making things more complex than necessary. This patch removes KERNFS_REMOVED. * Instead of KERNFS_REMOVED, each node now starts its life deactivated. This means that we now use both atomic_add() and atomic_sub() on KN_DEACTIVATED_BIAS, which is INT_MIN. The compiler generates an overflow warnings when negating INT_MIN as the negation can't be represented as a positive number. Nothing is actually broken but let's bump BIAS by one to avoid the warnings for archs which negates the subtrahend.. * A new helper kernfs_active() which tests whether kn->active >= 0 is added for convenience and lockdep annotation. All KERNFS_REMOVED tests are replaced with negated kernfs_active() tests. * __kernfs_remove() is updated to deactivate, but not drain, all nodes in the subtree instead of setting KERNFS_REMOVED. This removes deactivation from kernfs_deactivate(), which is now renamed to kernfs_drain(). * Sanity check on KERNFS_REMOVED in kernfs_put() is replaced with checks on the active ref. * Some comment style updates in the affected area. v2: Reordered before removal path restructuring. kernfs_active() dropped and kernfs_get/put_active() used instead. RB_EMPTY_NODE() used in the lookup paths. v3: Reverted most of v2 except for creating a new node with KN_DEACTIVATED_BIAS. Signed-off-by: -
Tejun Heo authored
There currently are two mechanisms gating active ref lockdep annotations - KERNFS_LOCKDEP flag and KERNFS_ACTIVE_REF type mask. The former disables lockdep annotations in kernfs_get/put_active() while the latter disables all of kernfs_deactivate(). While KERNFS_ACTIVE_REF also behaves as an optimization to skip the deactivation step for non-file nodes, the benefit is marginal and it needlessly diverges code paths. Let's drop KERNFS_ACTIVE_REF. While at it, add a test helper kernfs_lockdep() to test KERNFS_LOCKDEP flag so that it's more convenient and the related code can be compiled out when not enabled. v2: Refreshed on top of ("kernfs: make kernfs_deactivate() honor KERNFS_LOCKDEP flag"). As the earlier patch already added KERNFS_LOCKDEP tests to kernfs_deactivate(), those additions are dropped from this patch and the existing ones are simply converted to kernfs_lockdep(). Signed-off-by: -
Tejun Heo authored
kernfs_addrm_cxt and the accompanying kernfs_addrm_start/finish() were added because there were operations which should be performed outside kernfs_mutex after adding and removing kernfs_nodes. The necessary operations were recorded in kernfs_addrm_cxt and performed by kernfs_addrm_finish(); however, after the recent changes which relocated deactivation and unmapping so that they're performed directly during removal, the only operation kernfs_addrm_finish() performs is kernfs_put(), which can be moved inside the removal path too. This patch moves the kernfs_put() of the base ref to __kernfs_remove() and remove kernfs_addrm_cxt and kernfs_addrm_start/finish(). * kernfs_add_one() is updated to grab and release kernfs_mutex itself. sysfs_addrm_start/finish() invocations around it are removed from all users. * __kernfs_remove() puts an unlinked node directly instead of chaining it to kernfs_addrm_cxt. Its callers are updated to grab and release kernfs_mutex instead of calling kernfs_addrm_start/finish() around it. v2: Rebased on top of "kernfs: associate a new kernfs_node with its parent on creation" which dropped @parent from kernfs_add_one(). Signed-off-by: -
Tejun Heo authored
kernfs_unmap_bin_file() is supposed to unmap all memory mappings of the target file before kernfs_remove() finishes; however, it currently is being called from kernfs_addrm_finish() and has the same race problem as the original implementation of deactivation when there are multiple removers - only the remover which snatches the node to its addrm_cxt->removed list is guaranteed to wait for its completion before returning. It can be easily fixed by moving kernfs_unmap_bin_file() invocation from kernfs_addrm_finish() to kernfs_deactivated(). The function may be called multiple times but that shouldn't do any harm. Signed-off-by:
-
Tejun Heo authored
The recursive nature of kernfs_remove() means that, even if kernfs_remove() is not allowed to be called multiple times on the same node, there may be race conditions between removal of parent and its descendants. While we can claim that kernfs_remove() shouldn't be called on one of the descendants while the removal of an ancestor is in progress, such rule is unnecessarily restrictive and very difficult to enforce. It's better to simply allow invoking kernfs_remove() as the caller sees fit as long as the caller ensures that the node is accessible. The current behavior in such situations is broken. Whoever enters removal path first takes the node off the hierarchy and then deactivates. Following removers either return as soon as it notices that it's not the first one or can't even find the target node as it has already been removed from the hierarchy. In both cases, the following removers may finish prematurely while the nodes which should be removed and drained are still being processed by the first one. This patch restructures so that multiple removers, whether through recursion or direction invocation, always follow the following rules. * When there are multiple concurrent removers, only one puts the base ref. * Regardless of which one puts the base ref, all removers are blocked until the target node is fully deactivated and removed. To achieve the above, removal path now first marks all descendants including self REMOVED and then deactivates and unlinks leftmost descendant one-by-one. kernfs_deactivate() is called directly from __kernfs_removal() and drops and regrabs kernfs_mutex for each descendant to drain active refs. As this means that multiple removers can enter kernfs_deactivate() for the same node, the function is updated so that it can handle multiple deactivators of the same node - only one actually deactivates but all wait till drain completion. The restructured removal path guarantees that a removed node gets unlinked only after the node is deactivated and drained. Combined with proper multiple deactivator handling, this guarantees that any invocation of kernfs_remove() returns only after the node itself and all its descendants are deactivated, drained and removed. v2: Draining separated into a separate loop (used to be in the same loop as unlink) and done from __kernfs_deactivate(). This is to allow exposing deactivation as a separate interface later. Root node removal was broken in v1 patch. Fixed. v3: Revert most of v2 except for root node removal fix and simplification of KERNFS_REMOVED setting loop. v4: Refreshed on top of ("kernfs: make kernfs_deactivate() honor KERNFS_LOCKDEP flag"). Signed-off-by: -
Tejun Heo authored
kernfs_node->u.completion is used to notify deactivation completion from kernfs_put_active() to kernfs_deactivate(). We now allow multiple racing removals of the same node and the current removal scheme is no longer correct - kernfs_remove() invocation may return before the node is properly deactivated if it races against another removal. The removal path will be restructured to address the issue. To help such restructure which requires supporting multiple waiters, this patch replaces kernfs_node->u.completion with kernfs_root->deactivate_waitq. This makes deactivation event notifications share a per-root waitqueue_head; however, the wait path is quite cold and this will also allow shaving one pointer off kernfs_node. v2: Refreshed on top of ("kernfs: make kernfs_deactivate() honor KERNFS_LOCKDEP flag"). Signed-off-by: -
Tejun Heo authored
kernfs_deactivate() forgot to check whether KERNFS_LOCKDEP is set before performing lockdep annotations and ends up feeding uninitialized lockdep_map to lockdep triggering warning like the following on USB stick hotunplug. usb 1-2: USB disconnect, device number 2 INFO: trying to register non-static key. the code is fine but needs lockdep annotation. turning off the locking correctness validator. CPU: 1 PID: 62 Comm: khubd Not tainted 3.13.0-work+ #82 Hardware name: empty empty/S3992, BIOS 080011 10/26/2007 ffff880065ca7f60 ffff88013a4ffa08 ffffffff81cfb6bd 0000000000000002 ffff88013a4ffac8 ffffffff810f8530 ffff88013a4fc710 0000000000000002 ffff880100000000 ffffffff82a3db50 0000000000000001 ffff88013a4fc710 Call Trace: [<ffffffff81cfb6bd>] dump_stack+0x4e/0x7a [<ffffffff810f8530>] __lock_acquire+0x1910/0x1e70 [<ffffffff810f931a>] lock_acquire+0x9a/0x1d0 [<ffffffff8127c75e>] kernfs_deactivate+0xee/0x130 [<ffffffff8127d4c8>] kernfs_addrm_finish+0x38/0x60 [<ffffffff8127d701>] kernfs_remove_by_name_ns+0x51/0xa0 [<ffffffff8127b4f1>] remove_files.isra.1+0x41/0x80 [<ffffffff8127b7e7>] sysfs_remove_group+0x47/0xa0 [<ffffffff8127b873>] sysfs_remove_groups+0x33/0x50 [<ffffffff8177d66d>] device_remove_attrs+0x4d/0x80 [<ffffffff8177e25e>] device_del+0x12e/0x1d0 [<ffffffff819722c2>] usb_disconnect+0x122/0x1a0 [<ffffffff819749b5>] hub_thread+0x3c5/0x1290 [<ffffffff810c6a6d>] kthread+0xed/0x110 [<ffffffff81d0a56c>] ret_from_fork+0x7c/0xb0 Fix it by making kernfs_deactivate() perform lockdep annotations only if KERNFS_LOCKDEP is set. Signed-off-by:
Fabio Estevam <festevam@gmail.com> Reported-by:
Alan Stern <stern@rowland.harvard.edu> Signed-off-by:
-
Colin Cross authored
dma_buf_map_attachment and dma_buf_vmap can return NULL or ERR_PTR on a error. This encourages a common buggy pattern in callers: sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL); if (IS_ERR_OR_NULL(sgt)) return PTR_ERR(sgt); This causes the caller to return 0 on an error. IS_ERR_OR_NULL is almost always a sign of poorly-defined error handling. This patch converts dma_buf_map_attachment to always return ERR_PTR, and fixes the callers that incorrectly handled NULL. There are a few more callers that were not checking for NULL at all, which would have dereferenced a NULL pointer later. There are also a few more callers that correctly handled NULL and ERR_PTR differently, I left those alone but they could also be modified to delete the NULL check. This patch also converts dma_buf_vmap to always return NULL. All the callers to dma_buf_vmap only check for NULL, and would have dereferenced an ERR_PTR and panic'd if one was ever returned. This is not consistent with the rest of the dma buf APIs, but matches the expectations of all of the callers. Signed-off-by:Colin Cross <ccross@android.com> Reviewed-by:
Rob Clark <robdclark@gmail.com> Signed-off-by:
-
Hugh Dickins authored
Sometimes the cleanup after memcg hierarchy testing gets stuck in mem_cgroup_reparent_charges(), unable to bring non-kmem usage down to 0. There may turn out to be several causes, but a major cause is this: the workitem to offline parent can get run before workitem to offline child; parent's mem_cgroup_reparent_charges() circles around waiting for the child's pages to be reparented to its lrus, but it's holding cgroup_mutex which prevents the child from reaching its mem_cgroup_reparent_charges(). Just use an ordered workqueue for cgroup_destroy_wq. tj: Committing as the temporary fix until the reverse dependency can be removed from memcg. Comment updated accordingly. Fixes: e5fca243 ("cgroup: use a dedicated workqueue for cgroup destruction") Suggested-by:Filipe Brandenburger <filbranden@google.com> Signed-off-by:
Hugh Dickins <hughd@google.com> Cc: stable@vger.kernel.org # 3.10+ Signed-off-by:
-
