mm: make vm_area_struct anon_name field RCU-safe
Subsystems (8)
Suren Baghdasaryan
maintainerPATCH v2 1/3 mm: make vm_area_struct anon_name field RCU-safe
For lockless /proc/pid/maps reading we have to ensure all the fields used when generating the output are RCU-safe. The only pointer fields in vm_area_struct which are used to generate that file's output are vm_file and anon_name. vm_file is RCU-safe but anon_name is not. Make anon_name RCU-safe as well.
--- include/linux/mm_inline.h | 10 +++++++++- include/linux/mm_types.h | 3 ++- mm/madvise.c | 30 ++++++++++++++++++++++++++---- 3 files changed, 37 insertions(+), 6 deletions(-)
include/linux/mm_inline.h +9 -1
@@ -389,7 +389,7 @@ static inline void dup_anon_vma_name(struct vm_area_struct *orig_vma,struct anon_vma_name *anon_name =anon_vma_name (orig_vma);if (anon_name) - new_vma->anon_name =anon_vma_name_reuse (anon_name); +rcu_assign_pointer (new_vma->anon_name,anon_vma_name_reuse (anon_name)); }static inline void free_anon_vma_name (struct vm_area_struct *vma)
@@ -411,6 +411,8 @@ static inline bool anon_vma_name_eq(struct anon_vma_name *anon_name1, !strcmp(anon_name1->name, anon_name2->name); } +struct anon_vma_name *anon_vma_name_get_rcu (struct vm_area_struct *vma); +#else /* CONFIG_ANON_VMA_NAME */ static inline void anon_vma_name_get (struct anon_vma_name *anon_name) {}static inline void anon_vma_name_put (struct anon_vma_name *anon_name) {}
@@ -424,6 +426,12 @@ static inline bool anon_vma_name_eq(struct anon_vma_name *anon_name1, return true; } +static inline +struct anon_vma_name *anon_vma_name_get_rcu (struct vm_area_struct *vma) +{ +return NULL ; +} +#endif /* CONFIG_ANON_VMA_NAME */ static inline void init_tlb_flush_pending (struct mm_struct *mm)
include/linux/mm_types.h +2 -1
@@ -545,6 +545,7 @@ struct vm_userfaultfd_ctx {};struct anon_vma_name {struct kref kref; +struct rcu_head rcu;/* The name needs to be at the end because it is dynamically sized. */ char name[]; };
@@ -699,7 +700,7 @@ struct vm_area_struct {* terminated string containing the name given to the vma, or NULL if * unnamed. Serialized by mmap_lock. Use anon_vma_name to access. */ -struct anon_vma_name *anon_name; +struct anon_vma_name __rcu *anon_name;#endif #ifdef CONFIG_SWAP atomic_long_t swap_readahead_info;
mm/madvise.c +26 -4
@@ -88,14 +88,15 @@ void anon_vma_name_free(struct kref *kref) {struct anon_vma_name *anon_name = container_of(kref, struct anon_vma_name, kref); -kfree (anon_name); +kfree_rcu (anon_name, rcu); }struct anon_vma_name *anon_vma_name (struct vm_area_struct *vma) {mmap_assert_locked (vma->vm_mm); -return vma->anon_name; +return rcu_dereference_protected (vma->anon_name, +rwsem_is_locked (&vma->vm_mm->mmap_lock)); }/* mmap_lock should be write-locked */
@@ -105,7 +106,7 @@ static int replace_anon_vma_name(struct vm_area_struct *vma,struct anon_vma_name *orig_name =anon_vma_name (vma);if (!anon_name) { - vma->anon_name =NULL ; +rcu_assign_pointer (vma->anon_name,NULL );anon_vma_name_put (orig_name);return 0 ; }
@@ -113,11 +114,32 @@ static int replace_anon_vma_name(struct vm_area_struct *vma,if (anon_vma_name_eq (orig_name, anon_name))return 0 ; - vma->anon_name =anon_vma_name_reuse (anon_name); +rcu_assign_pointer (vma->anon_name,anon_vma_name_reuse (anon_name));anon_vma_name_put (orig_name); return 0; } + +/* +* Returned anon_vma_name is stable due to elevated refcount but not guaranteed +* to be assigned to the original VMA after the call. +*/ +struct anon_vma_name *anon_vma_name_get_rcu (struct vm_area_struct *vma) +{ +struct anon_vma_name __rcu *anon_name; + +WARN_ON_ONCE (!rcu_read_lock_held ()); + + anon_name =rcu_dereference (vma->anon_name); +if (!anon_name) +return NULL ; + +if (unlikely (!kref_get_unless_zero (&anon_name->kref))) +return NULL ; + +return anon_name; +} +#else /* CONFIG_ANON_VMA_NAME */ static int replace_anon_vma_name (struct vm_area_struct *vma,struct anon_vma_name *anon_name)
Signature · 1 line
2.43.0.429.g432eaa2c6b-goog
Suren Baghdasaryan
maintainerPATCH v2 2/3 mm: add mm_struct sequence number to detect write locks
Provide a way for lockless mm_struct users to detect whether mm might have been changed since some specific point in time. The API provided allows the user to record a counter when it starts using the mm and later use that counter to check if anyone write-locked mmap_lock since the counter was recorded. Recording the counter value should be done while holding mmap_lock at least for reading to prevent the counter from concurrent changes. Every time mmap_lock is write-locked mm_struct updates its mm_wr_seq counter so that checks against counters recorded before that would fail, indicating a possibility of mm being modified.
--- include/linux/mm_types.h | 2 ++ include/linux/mmap_lock.h | 22 ++++++++++++++++++++++ 2 files changed, 24 insertions(+)
include/linux/mm_types.h +2 -0
@@ -846,6 +846,8 @@ struct mm_struct {*/ int mm_lock_seq;#endif +/* Counter incremented each time mm gets write-locked */ +unsigned long mm_wr_seq;unsigned long hiwater_rss;/* High-watermark of RSS usage */
include/linux/mmap_lock.h +22 -0
@@ -106,6 +106,8 @@ static inline void mmap_write_lock(struct mm_struct *mm) {__mmap_lock_trace_start_locking (mm, true);down_write (&mm->mmap_lock); +/* Pairs with ACQUIRE semantics in mmap_write_seq_read */ +smp_store_release (&mm->mm_wr_seq, mm->mm_wr_seq +1 );__mmap_lock_trace_acquire_returned (mm, true, true); }
@@ -113,6 +115,8 @@ static inline void mmap_write_lock_nested(struct mm_struct *mm, int subclass) {__mmap_lock_trace_start_locking (mm, true);down_write_nested (&mm->mmap_lock, subclass); +/* Pairs with ACQUIRE semantics in mmap_write_seq_read */ +smp_store_release (&mm->mm_wr_seq, mm->mm_wr_seq +1 );__mmap_lock_trace_acquire_returned (mm, true, true); }
@@ -122,6 +126,10 @@ static inline int mmap_write_lock_killable(struct mm_struct *mm)__mmap_lock_trace_start_locking (mm, true); ret =down_write_killable (&mm->mmap_lock); +if (!ret) { +/* Pairs with ACQUIRE semantics in mmap_write_seq_read */ +smp_store_release (&mm->mm_wr_seq, mm->mm_wr_seq +1 ); + }__mmap_lock_trace_acquire_returned (mm, true, ret ==0 );return ret; }
@@ -140,6 +148,20 @@ static inline void mmap_write_downgrade(struct mm_struct *mm)downgrade_write (&mm->mmap_lock); } +static inline unsigned long mmap_write_seq_read (struct mm_struct *mm) +{ +/* Pairs with RELEASE semantics in mmap_write_lock */ +return smp_load_acquire (&mm->mm_wr_seq); +} + +static inline void mmap_write_seq_record (struct mm_struct *mm, +unsigned long *mm_wr_seq) +{ +mmap_assert_locked (mm); +/* Nobody can concurrently modify since we hold the mmap_lock */ + *mm_wr_seq = mm->mm_wr_seq; +} +static inline void mmap_read_lock (struct mm_struct *mm) {__mmap_lock_trace_start_locking (mm, false);
Signature · 1 line
2.43.0.429.g432eaa2c6b-goog
Suren Baghdasaryan
maintainerPATCH v2 3/3 mm/maps: read proc/pid/maps under RCU
In reply to Suren BaghdasaryanWith maple_tree supporting vma tree traversal under RCU and per-vma locks making vma access RCU-safe, /proc/pid/maps can be read under RCU and without the need to read-lock mmap_lock. However vma content can change from under us, therefore we make a copy of the vma and we pin pointer fields used when generating the output (currently only vm_file and anon_name). Afterwards we check for concurrent address space modifications, wait for them to end and retry. That last check is needed to avoid possibility of missing a vma during concurrent maple_tree node replacement, which might report a NULL when a vma is replaced with another one. While we take the mmap_lock for reading during such contention, we do that momentarily only to record new mm_wr_seq counter. This change is designed to reduce mmap_lock contention and prevent a process reading /proc/pid/maps files (often a low priority task, such as monitoring/data collection services) from blocking address space updates.
Note that this change has a userspace visible disadvantage: it allows for sub-page data tearing as opposed to the previous mechanism where data tearing could happen only between pages of generated output data. Since current userspace considers data tearing between pages to be acceptable, we assume is will be able to handle sub-page data tearing as well.
---
Changes since v1 [1]:
- Fixed CONFIG_ANON_VMA_NAME=n build by introducing anon_vma_name_{get|put}_if_valid, per SeongJae Park
- Fixed misspelling of get_vma_snapshot()
[1] https://lore.kernel.org/all/[email protected]/
fs/proc/internal.h | 2 + fs/proc/task_mmu.c | 113 +++++++++++++++++++++++++++++++++++--- include/linux/mm_inline.h | 18 ++++++ 3 files changed, 126 insertions(+), 7 deletions(-)
fs/proc/internal.h +2 -0
@@ -290,6 +290,8 @@ struct proc_maps_private {struct task_struct *task;struct mm_struct *mm;struct vma_iterator iter; +unsigned long mm_wr_seq; +struct vm_area_struct vma_copy;#ifdef CONFIG_NUMA struct mempolicy *task_mempolicy;#endif
fs/proc/task_mmu.c +106 -7
157 lines not on this page · raw message
include/linux/mm_inline.h +18 -0
@@ -413,6 +413,21 @@ static inline bool anon_vma_name_eq(struct anon_vma_name *anon_name1,struct anon_vma_name *anon_vma_name_get_rcu (struct vm_area_struct *vma); +/* +* Takes a reference if anon_vma is valid and stable (has references). +* Fails only if anon_vma is valid but we failed to get a reference. +*/ +static inline bool anon_vma_name_get_if_valid (struct vm_area_struct *vma) +{ +return !vma->anon_name ||anon_vma_name_get_rcu (vma); +} + +static inline void anon_vma_name_put_if_valid (struct vm_area_struct *vma) +{ +if (vma->anon_name) +anon_vma_name_put (vma->anon_name); +} +#else /* CONFIG_ANON_VMA_NAME */ static inline void anon_vma_name_get (struct anon_vma_name *anon_name) {}static inline void anon_vma_name_put (struct anon_vma_name *anon_name) {}
@@ -432,6 +447,9 @@ struct anon_vma_name *anon_vma_name_get_rcu(struct vm_area_struct *vma) return NULL; } +static inline bool anon_vma_name_get_if_valid (struct vm_area_struct *vma) {return true; } +static inline void anon_vma_name_put_if_valid (struct vm_area_struct *vma) {} +#endif /* CONFIG_ANON_VMA_NAME */ static inline void init_tlb_flush_pending (struct mm_struct *mm)
Signature · 1 line
2.43.0.429.g432eaa2c6b-goog
Suren Baghdasaryan
maintainerRe: [PATCH v2 3/3] mm/maps: read proc/pid/maps under RCU
Quoted 65 lines · Suren Baghdasaryan
With
maple_treesupporting vma tree traversal under RCU and per-vma locks making vma access RCU-safe,/proc/pid/mapscan be read under RCU and without the need to read-lockmmap_lock. However vma content can change from under us, therefore we make a copy of the vma and we pin pointer fields used when generating the output (currently onlyvm_fileandanon_name). Afterwards we check for concurrent address space modifications, wait for them to end and retry. That last check is needed to avoid possibility of missing a vma during concurrentmaple_treenode replacement, which might report a NULL when a vma is replaced with another one. While we take themmap_lockfor reading during such contention, we do that momentarily only to record newmm_wr_seqcounter. This change is designed to reducemmap_lockcontention and prevent a process reading/proc/pid/mapsfiles (often a low priority task, such as monitoring/data collection services) from blocking address space updates.Note that this change has a userspace visible disadvantage: it allows for sub-page data tearing as opposed to the previous mechanism where data tearing could happen only between pages of generated output data. Since current userspace considers data tearing between pages to be acceptable, we assume is will be able to handle sub-page data tearing as well.
Signed-off-by Suren Baghdasaryan <[email protected]>---
Changes since v1 [1]:
- FixedCONFIG_ANON_VMA_NAME=nbuild by introducing anon_vma_name_{get|put}_if_valid, per SeongJae Park
- Fixed misspelling ofget_vma_snapshot()[1] https://lore.kernel.org/all/[email protected]/
fs/proc/internal.h | 2 + fs/proc/task_mmu.c | 113 +++++++++++++++++++++++++++++++++++--- include/linux/mm_inline.h | 18 ++++++ 3 files changed, 126 insertions(+), 7 deletions(-)fs/proc/internal.h +2
-0@@ -290,6 +290,8 @@ struct proc_maps_private {struct task_struct *task;struct mm_struct *mm;struct vma_iterator iter; +unsigned long mm_wr_seq; +struct vm_area_struct vma_copy;#ifdef CONFIG_NUMA struct mempolicy *task_mempolicy;#endif fs/proc/task_mmu.c +5
-2@@ -126,11 +126,95 @@ static void release_task_mempolicy(struct proc_maps_private *priv) }#endif -static struct vm_area_struct *proc_get_vma (struct proc_maps_private *priv, -loff_t *ppos) +#ifdef CONFIG_PER_VMA_LOCK + +static const struct seq_operations proc_pid_maps_op; + +/*
fs/proc/task_mmu.c static void release_task_mempolicy(struct proc_maps_private *priv) · line 143
Quoted 10 lines
fs/proc/task_mmu.c +10
-0+* Take VMA snapshot and pin vm_file and anon_name as they are used by +* show_map_vma. +*/ +static int get_vma_snapshot (struct proc_maps_private *priv,struct vm_area_struct *vma) +{ +struct vm_area_struct *copy = &priv->vma_copy; +int ret = -EAGAIN ; + +memcpy (copy, vma,sizeof (*vma)); +if (copy->vm_file && !get_file_rcu (©->vm_file))
There is a problem in this patchset. I assumed that get_file_rcu() can be used against vma->vm_file but that's not true. vma->vm_file is freed via a call to fput() which schedules its freeing using schedule_delayed_work(..., 1) but I don't think that constitutes RCU grace period, so it can be freed from under us.
Andrew, could you please remove this patchset from your tree until I sort this out?
Thanks,
Suren.
Quoted 181 lines
fs/proc/task_mmu.c +91
-5+ goto out; + +if (!anon_vma_name_get_if_valid (copy)) + goto put_file; + +if (priv->mm_wr_seq ==mmap_write_seq_read (priv->mm)) +return 0 ; + +/* Address space got modified, vma might be stale. Wait and retry. */ +rcu_read_unlock (); + ret =mmap_read_lock_killable (priv->mm); +mmap_write_seq_record (priv->mm, &priv->mm_wr_seq); +mmap_read_unlock (priv->mm); +rcu_read_lock (); + +if (!ret) + ret = -EAGAIN ;/* no other errors, ok to retry */ + +anon_vma_name_put_if_valid (copy); +put_file: +if (copy->vm_file) +fput (copy->vm_file); +out: +return ret; +} + +static void put_vma_snapshot (struct proc_maps_private *priv) +{ +struct vm_area_struct *vma = &priv->vma_copy; + +anon_vma_name_put_if_valid (vma); +if (vma->vm_file) +fput (vma->vm_file); +} + +static inline bool needs_mmap_lock (struct seq_file *m) +{ +/* +* smaps and numa_maps perform page table walk, therefore require +* mmap_lock but maps can be read under RCU. +*/ +return m->op != &proc_pid_maps_op; +} + +#else /* CONFIG_PER_VMA_LOCK */ + +/* Without per-vma locks VMA access is not RCU-safe */ +static inline bool needs_mmap_lock (struct seq_file *m) {return true; } + +#endif /* CONFIG_PER_VMA_LOCK */ + +static struct vm_area_struct *proc_get_vma (struct seq_file *m,loff_t *ppos) { +struct proc_maps_private *priv = m->private;struct vm_area_struct *vma =vma_next (&priv->iter); +#ifdef CONFIG_PER_VMA_LOCK +if (vma && !needs_mmap_lock (m)) { +int ret; + +put_vma_snapshot (priv); +while ((ret =get_vma_snapshot (priv, vma)) == -EAGAIN ) { +/* lookup the vma at the last position again */ +vma_iter_init (&priv->iter, priv->mm, *ppos); + vma =vma_next (&priv->iter); + } + +if (ret) { +put_vma_snapshot (priv); +return NULL ; + } + vma = &priv->vma_copy; + } +#endif if (vma) { *ppos = vma->vm_start; }else {@@ -169,12 +253,20 @@ static void *m_start(struct seq_file *m, loff_t *ppos)return ERR_PTR (-EINTR ); } +/* Drop mmap_lock if possible */ +if (!needs_mmap_lock (m)) { +mmap_write_seq_record (priv->mm, &priv->mm_wr_seq); +mmap_read_unlock (priv->mm); +rcu_read_lock (); +memset (&priv->vma_copy,0 ,sizeof (priv->vma_copy)); + } +vma_iter_init (&priv->iter, mm, last_addr);hold_task_mempolicy (priv);if (last_addr ==-2UL )return get_gate_vma (mm); -return proc_get_vma (priv, ppos); +return proc_get_vma (m, ppos); }static void *m_next (struct seq_file *m,void *v,loff_t *ppos)@@ -183,7 +275,7 @@ static void *m_next(struct seq_file *m, void *v, loff_t *ppos) *ppos =-1UL ;return NULL ; } -return proc_get_vma (m->private, ppos); +return proc_get_vma (m, ppos); }static void m_stop (struct seq_file *m,void *v)@@ -195,7 +287,10 @@ static void m_stop(struct seq_file *m, void *v)return ;release_task_mempolicy (priv); -mmap_read_unlock (mm); +if (needs_mmap_lock (m)) +mmap_read_unlock (mm); +else +rcu_read_unlock ();mmput (mm);put_task_struct (priv->task); priv->task =NULL ;@@ -283,8 +378,10 @@ show_map_vma(struct seq_file *m, struct vm_area_struct *vma) start = vma->vm_start; end = vma->vm_end;show_vma_header_prefix (m, start, end, flags, pgoff, dev, ino); -if (mm) - anon_name =anon_vma_name (vma); +if (mm) { + anon_name =needs_mmap_lock (m) ?anon_vma_name (vma) : +anon_vma_name_get_rcu (vma); + }/* * Print the dentry name for named mappings, and a @@ -338,6 +435,8 @@ show_map_vma(struct seq_file *m, struct vm_area_struct *vma)seq_puts (m, name); }seq_putc (m,'\n' ); +if (anon_name && !needs_mmap_lock (m)) +anon_vma_name_put (anon_name); }static int show_map (struct seq_file *m,void *v)include/linux/mm_inline.h +18
-1@@ -413,6 +413,21 @@ static inline bool anon_vma_name_eq(struct anon_vma_name *anon_name1,struct anon_vma_name *anon_vma_name_get_rcu (struct vm_area_struct *vma); +/* +* Takes a reference if anon_vma is valid and stable (has references). +* Fails only if anon_vma is valid but we failed to get a reference. +*/ +static inline bool anon_vma_name_get_if_valid (struct vm_area_struct *vma) +{ +return !vma->anon_name ||anon_vma_name_get_rcu (vma); +} + +static inline void anon_vma_name_put_if_valid (struct vm_area_struct *vma) +{ +if (vma->anon_name) +anon_vma_name_put (vma->anon_name); +} +#else /* CONFIG_ANON_VMA_NAME */ static inline void anon_vma_name_get (struct anon_vma_name *anon_name) {}static inline void anon_vma_name_put (struct anon_vma_name *anon_name) {}@@ -432,6 +447,9 @@ struct anon_vma_name *anon_vma_name_get_rcu(struct vm_area_struct *vma) return NULL; } +static inline bool anon_vma_name_get_if_valid (struct vm_area_struct *vma) {return true; } +static inline void anon_vma_name_put_if_valid (struct vm_area_struct *vma) {} +#endif /* CONFIG_ANON_VMA_NAME */ static inline void init_tlb_flush_pending (struct mm_struct *mm) --2.43.0.429.g432eaa2c6b-goog