| Message ID | 20240813042917.506057-14-andrii@kernel.org (mailing list archive) |
|---|---|
| State | Handled Elsewhere |
| Headers | show |
| Series | uprobes: RCU-protected hot path optimizations | expand |
On Mon, Aug 12, 2024 at 09:29:17PM -0700, Andrii Nakryiko wrote: > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > attempting uprobe look up speculatively. > > We rely on newly added mmap_lock_speculation_{start,end}() helpers to > validate that mm_struct stays intact for entire duration of this > speculation. If not, we fall back to mmap_lock-protected lookup. > > This allows to avoid contention on mmap_lock in absolutely majority of > cases, nicely improving uprobe/uretprobe scalability. > Here I have to admit to being mostly ignorant about the mm, so bear with me. :> I note the result of find_active_uprobe_speculative is immediately stale in face of modifications. The thing I'm after is that the mmap_lock_speculation business adds overhead on archs where a release fence is not a de facto nop and I don't believe the commit message justifies it. Definitely a bummer to add merely it for uprobes. If there are bigger plans concerning it that's a different story of course. With this in mind I have to ask if instead you could perhaps get away with the already present per-vma sequence counter?
On Mon, Aug 12, 2024 at 11:18 PM Mateusz Guzik <mjguzik@gmail.com> wrote: > > On Mon, Aug 12, 2024 at 09:29:17PM -0700, Andrii Nakryiko wrote: > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > attempting uprobe look up speculatively. > > > > We rely on newly added mmap_lock_speculation_{start,end}() helpers to > > validate that mm_struct stays intact for entire duration of this > > speculation. If not, we fall back to mmap_lock-protected lookup. > > > > This allows to avoid contention on mmap_lock in absolutely majority of > > cases, nicely improving uprobe/uretprobe scalability. > > > > Here I have to admit to being mostly ignorant about the mm, so bear with > me. :> > > I note the result of find_active_uprobe_speculative is immediately stale > in face of modifications. > > The thing I'm after is that the mmap_lock_speculation business adds > overhead on archs where a release fence is not a de facto nop and I > don't believe the commit message justifies it. Definitely a bummer to > add merely it for uprobes. If there are bigger plans concerning it > that's a different story of course. > > With this in mind I have to ask if instead you could perhaps get away > with the already present per-vma sequence counter? per-vma sequence counter does not implement acquire/release logic, it relies on vma->vm_lock for synchronization. So if we want to use it, we would have to add additional memory barriers here. This is likely possible but as I mentioned before we would need to ensure the pagefault path does not regress. OTOH mm->mm_lock_seq already halfway there (it implements acquire/release logic), we just had to ensure mmap_write_lock() increments mm->mm_lock_seq. So, from the release fence overhead POV I think whether we use mm->mm_lock_seq or vma->vm_lock, we would still need a proper fence here.
On Tue, Aug 13, 2024 at 08:36:03AM -0700, Suren Baghdasaryan wrote: > On Mon, Aug 12, 2024 at 11:18 PM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > On Mon, Aug 12, 2024 at 09:29:17PM -0700, Andrii Nakryiko wrote: > > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > > attempting uprobe look up speculatively. > > > > > > We rely on newly added mmap_lock_speculation_{start,end}() helpers to > > > validate that mm_struct stays intact for entire duration of this > > > speculation. If not, we fall back to mmap_lock-protected lookup. > > > > > > This allows to avoid contention on mmap_lock in absolutely majority of > > > cases, nicely improving uprobe/uretprobe scalability. > > > > > > > Here I have to admit to being mostly ignorant about the mm, so bear with > > me. :> > > > > I note the result of find_active_uprobe_speculative is immediately stale > > in face of modifications. > > > > The thing I'm after is that the mmap_lock_speculation business adds > > overhead on archs where a release fence is not a de facto nop and I > > don't believe the commit message justifies it. Definitely a bummer to > > add merely it for uprobes. If there are bigger plans concerning it > > that's a different story of course. > > > > With this in mind I have to ask if instead you could perhaps get away > > with the already present per-vma sequence counter? > > per-vma sequence counter does not implement acquire/release logic, it > relies on vma->vm_lock for synchronization. So if we want to use it, > we would have to add additional memory barriers here. This is likely > possible but as I mentioned before we would need to ensure the > pagefault path does not regress. OTOH mm->mm_lock_seq already halfway > there (it implements acquire/release logic), we just had to ensure > mmap_write_lock() increments mm->mm_lock_seq. > > So, from the release fence overhead POV I think whether we use > mm->mm_lock_seq or vma->vm_lock, we would still need a proper fence > here. > Per my previous e-mail I'm not particularly familiar with mm internals, so I'm going to handwave a little bit with my $0,03 concerning multicore in general and if you disagree with it that's your business. For the time being I have no interest in digging into any of this. Before I do, to prevent this thread from being a total waste, here are some remarks concerning the patch with the assumption that the core idea lands. From the commit message: > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > attempting uprobe look up speculatively. Just in case I'll note a nit that this paragraph will need to be removed since the patch adding the flag is getting dropped. A non-nit which may or may not end up mattering is that the flag (which *is* set on the filep slab cache) makes things more difficult to validate. Normal RCU usage guarantees that the object itself wont be freed as long you follow the rules. However, the SLAB_TYPESAFE_BY_RCU flag weakens it significantly -- the thing at hand will always be a 'struct file', but it may get reallocated to *another* file from under you. Whether this aspect plays a role here I don't know. > +static struct uprobe *find_active_uprobe_speculative(unsigned long bp_vaddr) > +{ > + const vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE; > + struct mm_struct *mm = current->mm; > + struct uprobe *uprobe; > + struct vm_area_struct *vma; > + struct file *vm_file; > + struct inode *vm_inode; > + unsigned long vm_pgoff, vm_start; > + int seq; > + loff_t offset; > + > + if (!mmap_lock_speculation_start(mm, &seq)) > + return NULL; > + > + rcu_read_lock(); > + I don't think there is a correctness problem here, but entering rcu *after* deciding to speculatively do the lookup feels backwards. > + vma = vma_lookup(mm, bp_vaddr); > + if (!vma) > + goto bail; > + > + vm_file = data_race(vma->vm_file); > + if (!vm_file || (vma->vm_flags & flags) != VM_MAYEXEC) > + goto bail; > + If vma teardown is allowed to progress and the file got fput'ed... > + vm_inode = data_race(vm_file->f_inode); ... the inode can be NULL, I don't know if that's handled. More importantly though, per my previous description of SLAB_TYPESAFE_BY_RCU, by now the file could have been reallocated and the inode you did find is completely unrelated. I understand the intent is to backpedal from everything should the mm seqc change, but the above may happen to matter. > + vm_pgoff = data_race(vma->vm_pgoff); > + vm_start = data_race(vma->vm_start); > + > + offset = (loff_t)(vm_pgoff << PAGE_SHIFT) + (bp_vaddr - vm_start); > + uprobe = find_uprobe_rcu(vm_inode, offset); > + if (!uprobe) > + goto bail; > + > + /* now double check that nothing about MM changed */ > + if (!mmap_lock_speculation_end(mm, seq)) > + goto bail; This leaks the reference obtained by find_uprobe_rcu(). > + > + rcu_read_unlock(); > + > + /* happy case, we speculated successfully */ > + return uprobe; > +bail: > + rcu_read_unlock(); > + return NULL; > +} Now to some handwaving, here it is: The core of my concern is that adding more work to down_write on the mmap semaphore comes with certain side-effects and plausibly more than a sufficient speed up can be achieved without doing it. An mm-wide mechanism is just incredibly coarse-grained and it may happen to perform poorly when faced with a program which likes to mess with its address space -- the fast path is going to keep failing and only inducing *more* overhead as the code decides to down_read the mmap semaphore. Furthermore there may be work currently synchronized with down_write which perhaps can transition to "merely" down_read, but by the time it happens this and possibly other consumers expect a change in the sequence counter, messing with it. To my understanding the kernel supports parallel faults with per-vma locking. I would find it surprising if the same machinery could not be used to sort out uprobe handling above. I presume a down_read on vma around all the work would also sort out any issues concerning stability of the file or inode objects. Of course single-threaded performance would take a hit due to atomic stemming from down/up_read and parallel uprobe lookups on the same vma would also get slower, but I don't know if that's a problem for a real workload. I would not have any comments if all speed ups were achieved without modifying non-uprobe code.
On Thu, Aug 15, 2024 at 6:44 AM Mateusz Guzik <mjguzik@gmail.com> wrote: > > On Tue, Aug 13, 2024 at 08:36:03AM -0700, Suren Baghdasaryan wrote: > > On Mon, Aug 12, 2024 at 11:18 PM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > > > On Mon, Aug 12, 2024 at 09:29:17PM -0700, Andrii Nakryiko wrote: > > > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > > > attempting uprobe look up speculatively. > > > > > > > > We rely on newly added mmap_lock_speculation_{start,end}() helpers to > > > > validate that mm_struct stays intact for entire duration of this > > > > speculation. If not, we fall back to mmap_lock-protected lookup. > > > > > > > > This allows to avoid contention on mmap_lock in absolutely majority of > > > > cases, nicely improving uprobe/uretprobe scalability. > > > > > > > > > > Here I have to admit to being mostly ignorant about the mm, so bear with > > > me. :> > > > > > > I note the result of find_active_uprobe_speculative is immediately stale > > > in face of modifications. > > > > > > The thing I'm after is that the mmap_lock_speculation business adds > > > overhead on archs where a release fence is not a de facto nop and I > > > don't believe the commit message justifies it. Definitely a bummer to > > > add merely it for uprobes. If there are bigger plans concerning it > > > that's a different story of course. > > > > > > With this in mind I have to ask if instead you could perhaps get away > > > with the already present per-vma sequence counter? > > > > per-vma sequence counter does not implement acquire/release logic, it > > relies on vma->vm_lock for synchronization. So if we want to use it, > > we would have to add additional memory barriers here. This is likely > > possible but as I mentioned before we would need to ensure the > > pagefault path does not regress. OTOH mm->mm_lock_seq already halfway > > there (it implements acquire/release logic), we just had to ensure > > mmap_write_lock() increments mm->mm_lock_seq. > > > > So, from the release fence overhead POV I think whether we use > > mm->mm_lock_seq or vma->vm_lock, we would still need a proper fence > > here. > > > > Per my previous e-mail I'm not particularly familiar with mm internals, > so I'm going to handwave a little bit with my $0,03 concerning multicore > in general and if you disagree with it that's your business. For the > time being I have no interest in digging into any of this. > > Before I do, to prevent this thread from being a total waste, here are > some remarks concerning the patch with the assumption that the core idea > lands. > > From the commit message: > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > attempting uprobe look up speculatively. > > Just in case I'll note a nit that this paragraph will need to be removed > since the patch adding the flag is getting dropped. Yep, of course, I'll update all that for the next revision (I'll wait for non-RFC patches to land first before reposting). > > A non-nit which may or may not end up mattering is that the flag (which > *is* set on the filep slab cache) makes things more difficult to > validate. Normal RCU usage guarantees that the object itself wont be > freed as long you follow the rules. However, the SLAB_TYPESAFE_BY_RCU > flag weakens it significantly -- the thing at hand will always be a > 'struct file', but it may get reallocated to *another* file from under > you. Whether this aspect plays a role here I don't know. Yes, that's ok and is accounted for. We care about that memory not going even from under us (I'm not even sure if it matters that it is still a struct file, tbh; I think that shouldn't matter as we are prepared to deal with completely garbage values read from struct file). > > > +static struct uprobe *find_active_uprobe_speculative(unsigned long bp_vaddr) > > +{ > > + const vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE; > > + struct mm_struct *mm = current->mm; > > + struct uprobe *uprobe; > > + struct vm_area_struct *vma; > > + struct file *vm_file; > > + struct inode *vm_inode; > > + unsigned long vm_pgoff, vm_start; > > + int seq; > > + loff_t offset; > > + > > + if (!mmap_lock_speculation_start(mm, &seq)) > > + return NULL; > > + > > + rcu_read_lock(); > > + > > I don't think there is a correctness problem here, but entering rcu > *after* deciding to speculatively do the lookup feels backwards. RCU should protect VMA and file, mm itself won't go anywhere, so this seems ok. > > > + vma = vma_lookup(mm, bp_vaddr); > > + if (!vma) > > + goto bail; > > + > > + vm_file = data_race(vma->vm_file); > > + if (!vm_file || (vma->vm_flags & flags) != VM_MAYEXEC) > > + goto bail; > > + > > If vma teardown is allowed to progress and the file got fput'ed... > > > + vm_inode = data_race(vm_file->f_inode); > > ... the inode can be NULL, I don't know if that's handled. > Yep, inode pointer value is part of RB-tree key, so if it's NULL, we just won't find a matching uprobe. Same for any other "garbage" f_inode value. Importantly, we never should dereference such inode pointers, at least until we find a valid uprobe (in which case we keep inode reference to it). > More importantly though, per my previous description of > SLAB_TYPESAFE_BY_RCU, by now the file could have been reallocated and > the inode you did find is completely unrelated. > > I understand the intent is to backpedal from everything should the mm > seqc change, but the above may happen to matter. Yes, I think we took that into account. All that we care about is memory "type safety", i.e., even if struct file's memory is reused, it's ok, we'll eventually detect the change and will discard wrong uprobe that we might by accident lookup (though probably in most cases we just won't find a uprobe at all). > > > + vm_pgoff = data_race(vma->vm_pgoff); > > + vm_start = data_race(vma->vm_start); > > + > > + offset = (loff_t)(vm_pgoff << PAGE_SHIFT) + (bp_vaddr - vm_start); > > + uprobe = find_uprobe_rcu(vm_inode, offset); > > + if (!uprobe) > > + goto bail; > > + > > + /* now double check that nothing about MM changed */ > > + if (!mmap_lock_speculation_end(mm, seq)) > > + goto bail; > > This leaks the reference obtained by find_uprobe_rcu(). find_uprobe_rcu() doesn't obtain a reference, uprobe is RCU-protected, and if caller need a refcount bump it will have to use try_get_uprobe() (which might fail). > > > + > > + rcu_read_unlock(); > > + > > + /* happy case, we speculated successfully */ > > + return uprobe; > > +bail: > > + rcu_read_unlock(); > > + return NULL; > > +} > > Now to some handwaving, here it is: > > The core of my concern is that adding more work to down_write on the > mmap semaphore comes with certain side-effects and plausibly more than a > sufficient speed up can be achieved without doing it. > > An mm-wide mechanism is just incredibly coarse-grained and it may happen > to perform poorly when faced with a program which likes to mess with its > address space -- the fast path is going to keep failing and only > inducing *more* overhead as the code decides to down_read the mmap > semaphore. > > Furthermore there may be work currently synchronized with down_write > which perhaps can transition to "merely" down_read, but by the time it > happens this and possibly other consumers expect a change in the > sequence counter, messing with it. > > To my understanding the kernel supports parallel faults with per-vma > locking. I would find it surprising if the same machinery could not be > used to sort out uprobe handling above. per-vma locking is still *locking*. Which means memory sharing between multiple CPUs, which means limited scalability. Lots of work in this series went to avoid even refcounting (as I pointed out for find_uprobe_rcu()) due to the same reason, and so relying on per-VMA locking is just shifting the bottleneck from mmap_lock to vma->vm_lock. Worst (and not uncommon) case is the same uprobe in the same process (and thus vma) being hit on multiple CPUs at the same time. Whether that's protected by mmap_lock or vma->vm_lock is immaterial at that point (from scalability standpoint). > > I presume a down_read on vma around all the work would also sort out any > issues concerning stability of the file or inode objects. > > Of course single-threaded performance would take a hit due to atomic > stemming from down/up_read and parallel uprobe lookups on the same vma > would also get slower, but I don't know if that's a problem for a real > workload. > > I would not have any comments if all speed ups were achieved without > modifying non-uprobe code. I'm also not a mm-focused person, so I'll let Suren and others address mm-specific concerns, but I (hopefully) addressed all the uprobe-related questions and concerns you had.
On Thu, Aug 15, 2024 at 9:47 AM Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote: > > On Thu, Aug 15, 2024 at 6:44 AM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > On Tue, Aug 13, 2024 at 08:36:03AM -0700, Suren Baghdasaryan wrote: > > > On Mon, Aug 12, 2024 at 11:18 PM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > > > > > On Mon, Aug 12, 2024 at 09:29:17PM -0700, Andrii Nakryiko wrote: > > > > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > > > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > > > > attempting uprobe look up speculatively. > > > > > > > > > > We rely on newly added mmap_lock_speculation_{start,end}() helpers to > > > > > validate that mm_struct stays intact for entire duration of this > > > > > speculation. If not, we fall back to mmap_lock-protected lookup. > > > > > > > > > > This allows to avoid contention on mmap_lock in absolutely majority of > > > > > cases, nicely improving uprobe/uretprobe scalability. > > > > > > > > > > > > > Here I have to admit to being mostly ignorant about the mm, so bear with > > > > me. :> > > > > > > > > I note the result of find_active_uprobe_speculative is immediately stale > > > > in face of modifications. > > > > > > > > The thing I'm after is that the mmap_lock_speculation business adds > > > > overhead on archs where a release fence is not a de facto nop and I > > > > don't believe the commit message justifies it. Definitely a bummer to > > > > add merely it for uprobes. If there are bigger plans concerning it > > > > that's a different story of course. > > > > > > > > With this in mind I have to ask if instead you could perhaps get away > > > > with the already present per-vma sequence counter? > > > > > > per-vma sequence counter does not implement acquire/release logic, it > > > relies on vma->vm_lock for synchronization. So if we want to use it, > > > we would have to add additional memory barriers here. This is likely > > > possible but as I mentioned before we would need to ensure the > > > pagefault path does not regress. OTOH mm->mm_lock_seq already halfway > > > there (it implements acquire/release logic), we just had to ensure > > > mmap_write_lock() increments mm->mm_lock_seq. > > > > > > So, from the release fence overhead POV I think whether we use > > > mm->mm_lock_seq or vma->vm_lock, we would still need a proper fence > > > here. > > > > > > > Per my previous e-mail I'm not particularly familiar with mm internals, > > so I'm going to handwave a little bit with my $0,03 concerning multicore > > in general and if you disagree with it that's your business. For the > > time being I have no interest in digging into any of this. > > > > Before I do, to prevent this thread from being a total waste, here are > > some remarks concerning the patch with the assumption that the core idea > > lands. > > > > From the commit message: > > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > > attempting uprobe look up speculatively. > > > > Just in case I'll note a nit that this paragraph will need to be removed > > since the patch adding the flag is getting dropped. > > Yep, of course, I'll update all that for the next revision (I'll wait > for non-RFC patches to land first before reposting). > > > > > A non-nit which may or may not end up mattering is that the flag (which > > *is* set on the filep slab cache) makes things more difficult to > > validate. Normal RCU usage guarantees that the object itself wont be > > freed as long you follow the rules. However, the SLAB_TYPESAFE_BY_RCU > > flag weakens it significantly -- the thing at hand will always be a > > 'struct file', but it may get reallocated to *another* file from under > > you. Whether this aspect plays a role here I don't know. > > Yes, that's ok and is accounted for. We care about that memory not > going even from under us (I'm not even sure if it matters that it is > still a struct file, tbh; I think that shouldn't matter as we are > prepared to deal with completely garbage values read from struct > file). Correct, with SLAB_TYPESAFE_BY_RCU we do need an additional check that vma->vm_file has not been freed and reused. That's where mmap_lock_speculation_{start|end} helps us. For vma->vm_file to change from under us one would have to take mmap_lock for write. If that happens mmap_lock_speculation_{start|end} should detect that and terminate our speculation. > > > > > > +static struct uprobe *find_active_uprobe_speculative(unsigned long bp_vaddr) > > > +{ > > > + const vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE; > > > + struct mm_struct *mm = current->mm; > > > + struct uprobe *uprobe; > > > + struct vm_area_struct *vma; > > > + struct file *vm_file; > > > + struct inode *vm_inode; > > > + unsigned long vm_pgoff, vm_start; > > > + int seq; > > > + loff_t offset; > > > + > > > + if (!mmap_lock_speculation_start(mm, &seq)) > > > + return NULL; > > > + > > > + rcu_read_lock(); > > > + > > > > I don't think there is a correctness problem here, but entering rcu > > *after* deciding to speculatively do the lookup feels backwards. > > RCU should protect VMA and file, mm itself won't go anywhere, so this seems ok. > > > > > > + vma = vma_lookup(mm, bp_vaddr); > > > + if (!vma) > > > + goto bail; > > > + > > > + vm_file = data_race(vma->vm_file); > > > + if (!vm_file || (vma->vm_flags & flags) != VM_MAYEXEC) > > > + goto bail; > > > + > > > > If vma teardown is allowed to progress and the file got fput'ed... > > > > > + vm_inode = data_race(vm_file->f_inode); > > > > ... the inode can be NULL, I don't know if that's handled. > > > > Yep, inode pointer value is part of RB-tree key, so if it's NULL, we > just won't find a matching uprobe. Same for any other "garbage" > f_inode value. Importantly, we never should dereference such inode > pointers, at least until we find a valid uprobe (in which case we keep > inode reference to it). > > > More importantly though, per my previous description of > > SLAB_TYPESAFE_BY_RCU, by now the file could have been reallocated and > > the inode you did find is completely unrelated. > > > > I understand the intent is to backpedal from everything should the mm > > seqc change, but the above may happen to matter. > > Yes, I think we took that into account. All that we care about is > memory "type safety", i.e., even if struct file's memory is reused, > it's ok, we'll eventually detect the change and will discard wrong > uprobe that we might by accident lookup (though probably in most cases > we just won't find a uprobe at all). > > > > > > + vm_pgoff = data_race(vma->vm_pgoff); > > > + vm_start = data_race(vma->vm_start); > > > + > > > + offset = (loff_t)(vm_pgoff << PAGE_SHIFT) + (bp_vaddr - vm_start); > > > + uprobe = find_uprobe_rcu(vm_inode, offset); > > > + if (!uprobe) > > > + goto bail; > > > + > > > + /* now double check that nothing about MM changed */ > > > + if (!mmap_lock_speculation_end(mm, seq)) > > > + goto bail; > > > > This leaks the reference obtained by find_uprobe_rcu(). > > find_uprobe_rcu() doesn't obtain a reference, uprobe is RCU-protected, > and if caller need a refcount bump it will have to use > try_get_uprobe() (which might fail). > > > > > > + > > > + rcu_read_unlock(); > > > + > > > + /* happy case, we speculated successfully */ > > > + return uprobe; > > > +bail: > > > + rcu_read_unlock(); > > > + return NULL; > > > +} > > > > Now to some handwaving, here it is: > > > > The core of my concern is that adding more work to down_write on the > > mmap semaphore comes with certain side-effects and plausibly more than a > > sufficient speed up can be achieved without doing it. AFAIK writers of mmap_lock are not considered a fast path. In a sense yes, we made any writer a bit heavier but OTOH we also made mm->mm_lock_seq a proper sequence count which allows us to locklessly check if mmap_lock is write-locked. I think you asked whether there will be other uses for mmap_lock_speculation_{start|end} and yes. For example, I am planning to use them for printing /proc/{pid}/maps without taking mmap_lock (when it's uncontended). If we have VMA seq counter-based detection it would be better (see below). > > > > An mm-wide mechanism is just incredibly coarse-grained and it may happen > > to perform poorly when faced with a program which likes to mess with its > > address space -- the fast path is going to keep failing and only > > inducing *more* overhead as the code decides to down_read the mmap > > semaphore. > > > > Furthermore there may be work currently synchronized with down_write > > which perhaps can transition to "merely" down_read, but by the time it > > happens this and possibly other consumers expect a change in the > > sequence counter, messing with it. > > > > To my understanding the kernel supports parallel faults with per-vma > > locking. I would find it surprising if the same machinery could not be > > used to sort out uprobe handling above. From all the above, my understanding of your objection is that checking mmap_lock during our speculation is too coarse-grained and you would prefer to use the VMA seq counter to check that the VMA we are working on is unchanged. I agree, that would be ideal. I had a quick chat with Jann about this and the conclusion we came to is that we would need to add an additional smp_wmb() barrier inside vma_start_write() and a smp_rmb() in the speculation code: static inline void vma_start_write(struct vm_area_struct *vma) { int mm_lock_seq; if (__is_vma_write_locked(vma, &mm_lock_seq)) return; down_write(&vma->vm_lock->lock); /* * We should use WRITE_ONCE() here because we can have concurrent reads * from the early lockless pessimistic check in vma_start_read(). * We don't really care about the correctness of that early check, but * we should use WRITE_ONCE() for cleanliness and to keep KCSAN happy. */ WRITE_ONCE(vma->vm_lock_seq, mm_lock_seq); + smp_wmb(); up_write(&vma->vm_lock->lock); } Note: up_write(&vma->vm_lock->lock) in the vma_start_write() is not enough because it's one-way permeable (it's a "RELEASE operation") and later vma->vm_file store (or any other VMA modification) can move before our vma->vm_lock_seq store. This makes vma_start_write() heavier but again, it's write-locking, so should not be considered a fast path. With this change we can use the code suggested by Andrii in https://lore.kernel.org/all/CAEf4BzZeLg0WsYw2M7KFy0+APrPaPVBY7FbawB9vjcA2+6k69Q@mail.gmail.com/ with an additional smp_rmb(): rcu_read_lock() vma = find_vma(...) if (!vma) /* bail */ vm_lock_seq = smp_load_acquire(&vma->vm_lock_seq); mm_lock_seq = smp_load_acquire(&vma->mm->mm_lock_seq); /* I think vm_lock has to be acquired first to avoid the race */ if (mm_lock_seq == vm_lock_seq) /* bail, vma is write-locked */ ... perform uprobe lookup logic based on vma->vm_file->f_inode ... smp_rmb(); if (vma->vm_lock_seq != vm_lock_seq) /* bail, VMA might have changed */ The smp_rmb() is needed so that vma->vm_lock_seq load does not get reordered and moved up before speculation. I'm CC'ing Jann since he understands memory barriers way better than me and will keep me honest. > > per-vma locking is still *locking*. Which means memory sharing between > multiple CPUs, which means limited scalability. Lots of work in this > series went to avoid even refcounting (as I pointed out for > find_uprobe_rcu()) due to the same reason, and so relying on per-VMA > locking is just shifting the bottleneck from mmap_lock to > vma->vm_lock. Worst (and not uncommon) case is the same uprobe in the > same process (and thus vma) being hit on multiple CPUs at the same > time. Whether that's protected by mmap_lock or vma->vm_lock is > immaterial at that point (from scalability standpoint). > > > > > I presume a down_read on vma around all the work would also sort out any > > issues concerning stability of the file or inode objects. > > > > Of course single-threaded performance would take a hit due to atomic > > stemming from down/up_read and parallel uprobe lookups on the same vma > > would also get slower, but I don't know if that's a problem for a real > > workload. > > > > I would not have any comments if all speed ups were achieved without > > modifying non-uprobe code. > > I'm also not a mm-focused person, so I'll let Suren and others address > mm-specific concerns, but I (hopefully) addressed all the > uprobe-related questions and concerns you had.
On Thu, Aug 15, 2024 at 10:45:45AM -0700, Suren Baghdasaryan wrote: > >From all the above, my understanding of your objection is that > checking mmap_lock during our speculation is too coarse-grained and > you would prefer to use the VMA seq counter to check that the VMA we > are working on is unchanged. I agree, that would be ideal. I had a > quick chat with Jann about this and the conclusion we came to is that > we would need to add an additional smp_wmb() barrier inside > vma_start_write() and a smp_rmb() in the speculation code: > > static inline void vma_start_write(struct vm_area_struct *vma) > { > int mm_lock_seq; > > if (__is_vma_write_locked(vma, &mm_lock_seq)) > return; > > down_write(&vma->vm_lock->lock); > /* > * We should use WRITE_ONCE() here because we can have concurrent reads > * from the early lockless pessimistic check in vma_start_read(). > * We don't really care about the correctness of that early check, but > * we should use WRITE_ONCE() for cleanliness and to keep KCSAN happy. > */ > WRITE_ONCE(vma->vm_lock_seq, mm_lock_seq); > + smp_wmb(); > up_write(&vma->vm_lock->lock); > } > > Note: up_write(&vma->vm_lock->lock) in the vma_start_write() is not > enough because it's one-way permeable (it's a "RELEASE operation") and > later vma->vm_file store (or any other VMA modification) can move > before our vma->vm_lock_seq store. > > This makes vma_start_write() heavier but again, it's write-locking, so > should not be considered a fast path. > With this change we can use the code suggested by Andrii in > https://lore.kernel.org/all/CAEf4BzZeLg0WsYw2M7KFy0+APrPaPVBY7FbawB9vjcA2+6k69Q@mail.gmail.com/ > with an additional smp_rmb(): > > rcu_read_lock() > vma = find_vma(...) > if (!vma) /* bail */ > > vm_lock_seq = smp_load_acquire(&vma->vm_lock_seq); > mm_lock_seq = smp_load_acquire(&vma->mm->mm_lock_seq); > /* I think vm_lock has to be acquired first to avoid the race */ > if (mm_lock_seq == vm_lock_seq) > /* bail, vma is write-locked */ > ... perform uprobe lookup logic based on vma->vm_file->f_inode ... > smp_rmb(); > if (vma->vm_lock_seq != vm_lock_seq) > /* bail, VMA might have changed */ > > The smp_rmb() is needed so that vma->vm_lock_seq load does not get > reordered and moved up before speculation. > > I'm CC'ing Jann since he understands memory barriers way better than > me and will keep me honest. > So I briefly noted that maybe down_read on the vma would do it, but per Andrii parallel lookups on the same vma on multiple CPUs are expected, which whacks that out. When I initially mentioned per-vma sequence counters I blindly assumed they worked the usual way. I don't believe any fancy rework here is warranted especially given that the per-mm counter thing is expected to have other uses. However, chances are decent this can still be worked out with per-vma granualarity all while avoiding any stores on lookup and without invasive (or complicated) changes. The lockless uprobe code claims to guarantee only false negatives and the miss always falls back to the mmap semaphore lookup. There may be something here, I'm going to chew on it. That said, thank you both for writeup so far.
+brauner for "struct file" lifetime On Thu, Aug 15, 2024 at 7:45 PM Suren Baghdasaryan <surenb@google.com> wrote: > On Thu, Aug 15, 2024 at 9:47 AM Andrii Nakryiko > <andrii.nakryiko@gmail.com> wrote: > > > > On Thu, Aug 15, 2024 at 6:44 AM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > > > On Tue, Aug 13, 2024 at 08:36:03AM -0700, Suren Baghdasaryan wrote: > > > > On Mon, Aug 12, 2024 at 11:18 PM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > > > > > > > On Mon, Aug 12, 2024 at 09:29:17PM -0700, Andrii Nakryiko wrote: > > > > > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > > > > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > > > > > attempting uprobe look up speculatively. Stupid question: Is this uprobe stuff actually such a hot codepath that it makes sense to optimize it to be faster than the page fault path? (Sidenote: I find it kinda interesting that this is sort of going back in the direction of the old Speculative Page Faults design.) > > > > > > We rely on newly added mmap_lock_speculation_{start,end}() helpers to > > > > > > validate that mm_struct stays intact for entire duration of this > > > > > > speculation. If not, we fall back to mmap_lock-protected lookup. > > > > > > > > > > > > This allows to avoid contention on mmap_lock in absolutely majority of > > > > > > cases, nicely improving uprobe/uretprobe scalability. > > > > > > > > > > > > > > > > Here I have to admit to being mostly ignorant about the mm, so bear with > > > > > me. :> > > > > > > > > > > I note the result of find_active_uprobe_speculative is immediately stale > > > > > in face of modifications. > > > > > > > > > > The thing I'm after is that the mmap_lock_speculation business adds > > > > > overhead on archs where a release fence is not a de facto nop and I > > > > > don't believe the commit message justifies it. Definitely a bummer to > > > > > add merely it for uprobes. If there are bigger plans concerning it > > > > > that's a different story of course. > > > > > > > > > > With this in mind I have to ask if instead you could perhaps get away > > > > > with the already present per-vma sequence counter? > > > > > > > > per-vma sequence counter does not implement acquire/release logic, it > > > > relies on vma->vm_lock for synchronization. So if we want to use it, > > > > we would have to add additional memory barriers here. This is likely > > > > possible but as I mentioned before we would need to ensure the > > > > pagefault path does not regress. OTOH mm->mm_lock_seq already halfway > > > > there (it implements acquire/release logic), we just had to ensure > > > > mmap_write_lock() increments mm->mm_lock_seq. > > > > > > > > So, from the release fence overhead POV I think whether we use > > > > mm->mm_lock_seq or vma->vm_lock, we would still need a proper fence > > > > here. > > > > > > > > > > Per my previous e-mail I'm not particularly familiar with mm internals, > > > so I'm going to handwave a little bit with my $0,03 concerning multicore > > > in general and if you disagree with it that's your business. For the > > > time being I have no interest in digging into any of this. > > > > > > Before I do, to prevent this thread from being a total waste, here are > > > some remarks concerning the patch with the assumption that the core idea > > > lands. > > > > > > From the commit message: > > > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > > > attempting uprobe look up speculatively. > > > > > > Just in case I'll note a nit that this paragraph will need to be removed > > > since the patch adding the flag is getting dropped. > > > > Yep, of course, I'll update all that for the next revision (I'll wait > > for non-RFC patches to land first before reposting). > > > > > > > > A non-nit which may or may not end up mattering is that the flag (which > > > *is* set on the filep slab cache) makes things more difficult to > > > validate. Normal RCU usage guarantees that the object itself wont be > > > freed as long you follow the rules. However, the SLAB_TYPESAFE_BY_RCU > > > flag weakens it significantly -- the thing at hand will always be a > > > 'struct file', but it may get reallocated to *another* file from under > > > you. Whether this aspect plays a role here I don't know. > > > > Yes, that's ok and is accounted for. We care about that memory not > > going even from under us (I'm not even sure if it matters that it is > > still a struct file, tbh; I think that shouldn't matter as we are > > prepared to deal with completely garbage values read from struct > > file). > > Correct, with SLAB_TYPESAFE_BY_RCU we do need an additional check that > vma->vm_file has not been freed and reused. That's where > mmap_lock_speculation_{start|end} helps us. For vma->vm_file to change > from under us one would have to take mmap_lock for write. If that > happens mmap_lock_speculation_{start|end} should detect that and > terminate our speculation. > > > > > > > > > > +static struct uprobe *find_active_uprobe_speculative(unsigned long bp_vaddr) > > > > +{ > > > > + const vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE; > > > > + struct mm_struct *mm = current->mm; > > > > + struct uprobe *uprobe; > > > > + struct vm_area_struct *vma; > > > > + struct file *vm_file; > > > > + struct inode *vm_inode; > > > > + unsigned long vm_pgoff, vm_start; > > > > + int seq; > > > > + loff_t offset; > > > > + > > > > + if (!mmap_lock_speculation_start(mm, &seq)) > > > > + return NULL; > > > > + > > > > + rcu_read_lock(); > > > > + > > > > > > I don't think there is a correctness problem here, but entering rcu > > > *after* deciding to speculatively do the lookup feels backwards. > > > > RCU should protect VMA and file, mm itself won't go anywhere, so this seems ok. > > > > > > > > > + vma = vma_lookup(mm, bp_vaddr); > > > > + if (!vma) > > > > + goto bail; > > > > + > > > > + vm_file = data_race(vma->vm_file); > > > > + if (!vm_file || (vma->vm_flags & flags) != VM_MAYEXEC) > > > > + goto bail; > > > > + > > > > > > If vma teardown is allowed to progress and the file got fput'ed... > > > > > > > + vm_inode = data_race(vm_file->f_inode); > > > > > > ... the inode can be NULL, I don't know if that's handled. > > > > > > > Yep, inode pointer value is part of RB-tree key, so if it's NULL, we > > just won't find a matching uprobe. Same for any other "garbage" > > f_inode value. Importantly, we never should dereference such inode > > pointers, at least until we find a valid uprobe (in which case we keep > > inode reference to it). > > > > > More importantly though, per my previous description of > > > SLAB_TYPESAFE_BY_RCU, by now the file could have been reallocated and > > > the inode you did find is completely unrelated. > > > > > > I understand the intent is to backpedal from everything should the mm > > > seqc change, but the above may happen to matter. > > > > Yes, I think we took that into account. All that we care about is > > memory "type safety", i.e., even if struct file's memory is reused, > > it's ok, we'll eventually detect the change and will discard wrong > > uprobe that we might by accident lookup (though probably in most cases > > we just won't find a uprobe at all). > > > > > > > > > + vm_pgoff = data_race(vma->vm_pgoff); > > > > + vm_start = data_race(vma->vm_start); > > > > + > > > > + offset = (loff_t)(vm_pgoff << PAGE_SHIFT) + (bp_vaddr - vm_start); > > > > + uprobe = find_uprobe_rcu(vm_inode, offset); > > > > + if (!uprobe) > > > > + goto bail; > > > > + > > > > + /* now double check that nothing about MM changed */ > > > > + if (!mmap_lock_speculation_end(mm, seq)) > > > > + goto bail; > > > > > > This leaks the reference obtained by find_uprobe_rcu(). > > > > find_uprobe_rcu() doesn't obtain a reference, uprobe is RCU-protected, > > and if caller need a refcount bump it will have to use > > try_get_uprobe() (which might fail). > > > > > > > > > + > > > > + rcu_read_unlock(); > > > > + > > > > + /* happy case, we speculated successfully */ > > > > + return uprobe; > > > > +bail: > > > > + rcu_read_unlock(); > > > > + return NULL; > > > > +} > > > > > > Now to some handwaving, here it is: > > > > > > The core of my concern is that adding more work to down_write on the > > > mmap semaphore comes with certain side-effects and plausibly more than a > > > sufficient speed up can be achieved without doing it. > > AFAIK writers of mmap_lock are not considered a fast path. In a sense > yes, we made any writer a bit heavier but OTOH we also made > mm->mm_lock_seq a proper sequence count which allows us to locklessly > check if mmap_lock is write-locked. I think you asked whether there > will be other uses for mmap_lock_speculation_{start|end} and yes. For > example, I am planning to use them for printing /proc/{pid}/maps > without taking mmap_lock (when it's uncontended). What would be the goal of this - to avoid cacheline bouncing of the mmap lock between readers? Or to allow mmap_write_lock() to preempt /proc/{pid}/maps readers who started out uncontended? Is the idea that you'd change show_map_vma() to first do something like get_file_active() to increment the file refcount (because otherwise the dentry can be freed under you and you need the dentry for path printing), then recheck your sequence count on the mm or vma (to avoid accessing the dentry of an unrelated file that hasn't become userspace-visible yet and may not have a proper dentry pointer yet), then print the file path, drop the file reference again, and in the end recheck the sequence count again before actually returning the printed data to userspace? > If we have VMA seq > counter-based detection it would be better (see below). > > > > > > > An mm-wide mechanism is just incredibly coarse-grained and it may happen > > > to perform poorly when faced with a program which likes to mess with its > > > address space -- the fast path is going to keep failing and only > > > inducing *more* overhead as the code decides to down_read the mmap > > > semaphore. > > > > > > Furthermore there may be work currently synchronized with down_write > > > which perhaps can transition to "merely" down_read, but by the time it > > > happens this and possibly other consumers expect a change in the > > > sequence counter, messing with it. > > > > > > To my understanding the kernel supports parallel faults with per-vma > > > locking. I would find it surprising if the same machinery could not be > > > used to sort out uprobe handling above. > > From all the above, my understanding of your objection is that > checking mmap_lock during our speculation is too coarse-grained and > you would prefer to use the VMA seq counter to check that the VMA we > are working on is unchanged. I agree, that would be ideal. I had a > quick chat with Jann about this and the conclusion we came to is that > we would need to add an additional smp_wmb() barrier inside > vma_start_write() and a smp_rmb() in the speculation code: > > static inline void vma_start_write(struct vm_area_struct *vma) > { > int mm_lock_seq; > > if (__is_vma_write_locked(vma, &mm_lock_seq)) > return; > > down_write(&vma->vm_lock->lock); > /* > * We should use WRITE_ONCE() here because we can have concurrent reads > * from the early lockless pessimistic check in vma_start_read(). > * We don't really care about the correctness of that early check, but > * we should use WRITE_ONCE() for cleanliness and to keep KCSAN happy. > */ > WRITE_ONCE(vma->vm_lock_seq, mm_lock_seq); > + smp_wmb(); > up_write(&vma->vm_lock->lock); > } > > Note: up_write(&vma->vm_lock->lock) in the vma_start_write() is not > enough because it's one-way permeable (it's a "RELEASE operation") and > later vma->vm_file store (or any other VMA modification) can move > before our vma->vm_lock_seq store. > > This makes vma_start_write() heavier but again, it's write-locking, so > should not be considered a fast path. > With this change we can use the code suggested by Andrii in > https://lore.kernel.org/all/CAEf4BzZeLg0WsYw2M7KFy0+APrPaPVBY7FbawB9vjcA2+6k69Q@mail.gmail.com/ > with an additional smp_rmb(): > > rcu_read_lock() > vma = find_vma(...) > if (!vma) /* bail */ And maybe add some comments like: /* * Load the current VMA lock sequence - we will detect if anyone concurrently * locks the VMA after this point. * Pairs with smp_wmb() in vma_start_write(). */ > vm_lock_seq = smp_load_acquire(&vma->vm_lock_seq); /* * Now we just have to detect if the VMA is already locked with its current * sequence count. * * The following load is ordered against the vm_lock_seq load above (using * smp_load_acquire() for the load above), and pairs with implicit memory * ordering between the mm_lock_seq write in mmap_write_unlock() and the * vm_lock_seq write in the next vma_start_write() after that (which can only * occur after an mmap_write_lock()). */ > mm_lock_seq = smp_load_acquire(&vma->mm->mm_lock_seq); > /* I think vm_lock has to be acquired first to avoid the race */ > if (mm_lock_seq == vm_lock_seq) > /* bail, vma is write-locked */ > ... perform uprobe lookup logic based on vma->vm_file->f_inode ... /* * Order the speculative accesses above against the following vm_lock_seq * recheck. */ > smp_rmb(); > if (vma->vm_lock_seq != vm_lock_seq) (As I said on the other thread: Since this now relies on vma->vm_lock_seq not wrapping back to the same value for correctness, I'd like to see vma->vm_lock_seq being at least an "unsigned long", or even better, an atomic64_t... though I realize we don't currently do that for seqlocks either.) > /* bail, VMA might have changed */ > > The smp_rmb() is needed so that vma->vm_lock_seq load does not get > reordered and moved up before speculation. > > I'm CC'ing Jann since he understands memory barriers way better than > me and will keep me honest.
On Thu, Aug 15, 2024 at 8:58 PM Jann Horn <jannh@google.com> wrote: > Stupid question: Is this uprobe stuff actually such a hot codepath > that it makes sense to optimize it to be faster than the page fault > path? > That's what I implicitly asked, hoping a down_read on vma would do it, but Andrii claims multiple parallel lookups on the same vma are a problem. Even so, I suspect something *simple* is doable here which avoids any writes to vmas and does not need the mm-wide sequence counter. It may be requirements are lax enough that merely observing some state is the same before and after uprobe lookup will be sufficient, or maybe some other hackery is viable without messing with fences in vma_start_write.
On Thu, Aug 15, 2024 at 11:58 AM Jann Horn <jannh@google.com> wrote: > > +brauner for "struct file" lifetime > > On Thu, Aug 15, 2024 at 7:45 PM Suren Baghdasaryan <surenb@google.com> wrote: > > On Thu, Aug 15, 2024 at 9:47 AM Andrii Nakryiko > > <andrii.nakryiko@gmail.com> wrote: > > > > > > On Thu, Aug 15, 2024 at 6:44 AM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > > > > > On Tue, Aug 13, 2024 at 08:36:03AM -0700, Suren Baghdasaryan wrote: > > > > > On Mon, Aug 12, 2024 at 11:18 PM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > > > > > > > > > On Mon, Aug 12, 2024 at 09:29:17PM -0700, Andrii Nakryiko wrote: > > > > > > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > > > > > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > > > > > > attempting uprobe look up speculatively. > > Stupid question: Is this uprobe stuff actually such a hot codepath > that it makes sense to optimize it to be faster than the page fault > path? > > (Sidenote: I find it kinda interesting that this is sort of going back > in the direction of the old Speculative Page Faults design.) > > > > > > > > We rely on newly added mmap_lock_speculation_{start,end}() helpers to > > > > > > > validate that mm_struct stays intact for entire duration of this > > > > > > > speculation. If not, we fall back to mmap_lock-protected lookup. > > > > > > > > > > > > > > This allows to avoid contention on mmap_lock in absolutely majority of > > > > > > > cases, nicely improving uprobe/uretprobe scalability. > > > > > > > > > > > > > > > > > > > Here I have to admit to being mostly ignorant about the mm, so bear with > > > > > > me. :> > > > > > > > > > > > > I note the result of find_active_uprobe_speculative is immediately stale > > > > > > in face of modifications. > > > > > > > > > > > > The thing I'm after is that the mmap_lock_speculation business adds > > > > > > overhead on archs where a release fence is not a de facto nop and I > > > > > > don't believe the commit message justifies it. Definitely a bummer to > > > > > > add merely it for uprobes. If there are bigger plans concerning it > > > > > > that's a different story of course. > > > > > > > > > > > > With this in mind I have to ask if instead you could perhaps get away > > > > > > with the already present per-vma sequence counter? > > > > > > > > > > per-vma sequence counter does not implement acquire/release logic, it > > > > > relies on vma->vm_lock for synchronization. So if we want to use it, > > > > > we would have to add additional memory barriers here. This is likely > > > > > possible but as I mentioned before we would need to ensure the > > > > > pagefault path does not regress. OTOH mm->mm_lock_seq already halfway > > > > > there (it implements acquire/release logic), we just had to ensure > > > > > mmap_write_lock() increments mm->mm_lock_seq. > > > > > > > > > > So, from the release fence overhead POV I think whether we use > > > > > mm->mm_lock_seq or vma->vm_lock, we would still need a proper fence > > > > > here. > > > > > > > > > > > > > Per my previous e-mail I'm not particularly familiar with mm internals, > > > > so I'm going to handwave a little bit with my $0,03 concerning multicore > > > > in general and if you disagree with it that's your business. For the > > > > time being I have no interest in digging into any of this. > > > > > > > > Before I do, to prevent this thread from being a total waste, here are > > > > some remarks concerning the patch with the assumption that the core idea > > > > lands. > > > > > > > > From the commit message: > > > > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > > > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > > > > attempting uprobe look up speculatively. > > > > > > > > Just in case I'll note a nit that this paragraph will need to be removed > > > > since the patch adding the flag is getting dropped. > > > > > > Yep, of course, I'll update all that for the next revision (I'll wait > > > for non-RFC patches to land first before reposting). > > > > > > > > > > > A non-nit which may or may not end up mattering is that the flag (which > > > > *is* set on the filep slab cache) makes things more difficult to > > > > validate. Normal RCU usage guarantees that the object itself wont be > > > > freed as long you follow the rules. However, the SLAB_TYPESAFE_BY_RCU > > > > flag weakens it significantly -- the thing at hand will always be a > > > > 'struct file', but it may get reallocated to *another* file from under > > > > you. Whether this aspect plays a role here I don't know. > > > > > > Yes, that's ok and is accounted for. We care about that memory not > > > going even from under us (I'm not even sure if it matters that it is > > > still a struct file, tbh; I think that shouldn't matter as we are > > > prepared to deal with completely garbage values read from struct > > > file). > > > > Correct, with SLAB_TYPESAFE_BY_RCU we do need an additional check that > > vma->vm_file has not been freed and reused. That's where > > mmap_lock_speculation_{start|end} helps us. For vma->vm_file to change > > from under us one would have to take mmap_lock for write. If that > > happens mmap_lock_speculation_{start|end} should detect that and > > terminate our speculation. > > > > > > > > > > > > > > +static struct uprobe *find_active_uprobe_speculative(unsigned long bp_vaddr) > > > > > +{ > > > > > + const vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE; > > > > > + struct mm_struct *mm = current->mm; > > > > > + struct uprobe *uprobe; > > > > > + struct vm_area_struct *vma; > > > > > + struct file *vm_file; > > > > > + struct inode *vm_inode; > > > > > + unsigned long vm_pgoff, vm_start; > > > > > + int seq; > > > > > + loff_t offset; > > > > > + > > > > > + if (!mmap_lock_speculation_start(mm, &seq)) > > > > > + return NULL; > > > > > + > > > > > + rcu_read_lock(); > > > > > + > > > > > > > > I don't think there is a correctness problem here, but entering rcu > > > > *after* deciding to speculatively do the lookup feels backwards. > > > > > > RCU should protect VMA and file, mm itself won't go anywhere, so this seems ok. > > > > > > > > > > > > + vma = vma_lookup(mm, bp_vaddr); > > > > > + if (!vma) > > > > > + goto bail; > > > > > + > > > > > + vm_file = data_race(vma->vm_file); > > > > > + if (!vm_file || (vma->vm_flags & flags) != VM_MAYEXEC) > > > > > + goto bail; > > > > > + > > > > > > > > If vma teardown is allowed to progress and the file got fput'ed... > > > > > > > > > + vm_inode = data_race(vm_file->f_inode); > > > > > > > > ... the inode can be NULL, I don't know if that's handled. > > > > > > > > > > Yep, inode pointer value is part of RB-tree key, so if it's NULL, we > > > just won't find a matching uprobe. Same for any other "garbage" > > > f_inode value. Importantly, we never should dereference such inode > > > pointers, at least until we find a valid uprobe (in which case we keep > > > inode reference to it). > > > > > > > More importantly though, per my previous description of > > > > SLAB_TYPESAFE_BY_RCU, by now the file could have been reallocated and > > > > the inode you did find is completely unrelated. > > > > > > > > I understand the intent is to backpedal from everything should the mm > > > > seqc change, but the above may happen to matter. > > > > > > Yes, I think we took that into account. All that we care about is > > > memory "type safety", i.e., even if struct file's memory is reused, > > > it's ok, we'll eventually detect the change and will discard wrong > > > uprobe that we might by accident lookup (though probably in most cases > > > we just won't find a uprobe at all). > > > > > > > > > > > > + vm_pgoff = data_race(vma->vm_pgoff); > > > > > + vm_start = data_race(vma->vm_start); > > > > > + > > > > > + offset = (loff_t)(vm_pgoff << PAGE_SHIFT) + (bp_vaddr - vm_start); > > > > > + uprobe = find_uprobe_rcu(vm_inode, offset); > > > > > + if (!uprobe) > > > > > + goto bail; > > > > > + > > > > > + /* now double check that nothing about MM changed */ > > > > > + if (!mmap_lock_speculation_end(mm, seq)) > > > > > + goto bail; > > > > > > > > This leaks the reference obtained by find_uprobe_rcu(). > > > > > > find_uprobe_rcu() doesn't obtain a reference, uprobe is RCU-protected, > > > and if caller need a refcount bump it will have to use > > > try_get_uprobe() (which might fail). > > > > > > > > > > > > + > > > > > + rcu_read_unlock(); > > > > > + > > > > > + /* happy case, we speculated successfully */ > > > > > + return uprobe; > > > > > +bail: > > > > > + rcu_read_unlock(); > > > > > + return NULL; > > > > > +} > > > > > > > > Now to some handwaving, here it is: > > > > > > > > The core of my concern is that adding more work to down_write on the > > > > mmap semaphore comes with certain side-effects and plausibly more than a > > > > sufficient speed up can be achieved without doing it. > > > > AFAIK writers of mmap_lock are not considered a fast path. In a sense > > yes, we made any writer a bit heavier but OTOH we also made > > mm->mm_lock_seq a proper sequence count which allows us to locklessly > > check if mmap_lock is write-locked. I think you asked whether there > > will be other uses for mmap_lock_speculation_{start|end} and yes. For > > example, I am planning to use them for printing /proc/{pid}/maps > > without taking mmap_lock (when it's uncontended). > > What would be the goal of this - to avoid cacheline bouncing of the > mmap lock between readers? Or to allow mmap_write_lock() to preempt > /proc/{pid}/maps readers who started out uncontended? The latter, from my early patchset which I need to refine (https://lore.kernel.org/all/20240123231014.3801041-3-surenb@google.com/): 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. > > Is the idea that you'd change show_map_vma() to first do something > like get_file_active() to increment the file refcount (because > otherwise the dentry can be freed under you and you need the dentry > for path printing), then recheck your sequence count on the mm or vma > (to avoid accessing the dentry of an unrelated file that hasn't become > userspace-visible yet and may not have a proper dentry pointer yet), > then print the file path, drop the file reference again, and in the > end recheck the sequence count again before actually returning the > printed data to userspace? Yeah, you can see the details in that link I posted above. See get_vma_snapshot() function. > > > If we have VMA seq > > counter-based detection it would be better (see below). > > > > > > > > > > An mm-wide mechanism is just incredibly coarse-grained and it may happen > > > > to perform poorly when faced with a program which likes to mess with its > > > > address space -- the fast path is going to keep failing and only > > > > inducing *more* overhead as the code decides to down_read the mmap > > > > semaphore. > > > > > > > > Furthermore there may be work currently synchronized with down_write > > > > which perhaps can transition to "merely" down_read, but by the time it > > > > happens this and possibly other consumers expect a change in the > > > > sequence counter, messing with it. > > > > > > > > To my understanding the kernel supports parallel faults with per-vma > > > > locking. I would find it surprising if the same machinery could not be > > > > used to sort out uprobe handling above. > > > > From all the above, my understanding of your objection is that > > checking mmap_lock during our speculation is too coarse-grained and > > you would prefer to use the VMA seq counter to check that the VMA we > > are working on is unchanged. I agree, that would be ideal. I had a > > quick chat with Jann about this and the conclusion we came to is that > > we would need to add an additional smp_wmb() barrier inside > > vma_start_write() and a smp_rmb() in the speculation code: > > > > static inline void vma_start_write(struct vm_area_struct *vma) > > { > > int mm_lock_seq; > > > > if (__is_vma_write_locked(vma, &mm_lock_seq)) > > return; > > > > down_write(&vma->vm_lock->lock); > > /* > > * We should use WRITE_ONCE() here because we can have concurrent reads > > * from the early lockless pessimistic check in vma_start_read(). > > * We don't really care about the correctness of that early check, but > > * we should use WRITE_ONCE() for cleanliness and to keep KCSAN happy. > > */ > > WRITE_ONCE(vma->vm_lock_seq, mm_lock_seq); > > + smp_wmb(); > > up_write(&vma->vm_lock->lock); > > } > > > > Note: up_write(&vma->vm_lock->lock) in the vma_start_write() is not > > enough because it's one-way permeable (it's a "RELEASE operation") and > > later vma->vm_file store (or any other VMA modification) can move > > before our vma->vm_lock_seq store. > > > > This makes vma_start_write() heavier but again, it's write-locking, so > > should not be considered a fast path. > > With this change we can use the code suggested by Andrii in > > https://lore.kernel.org/all/CAEf4BzZeLg0WsYw2M7KFy0+APrPaPVBY7FbawB9vjcA2+6k69Q@mail.gmail.com/ > > with an additional smp_rmb(): > > > > rcu_read_lock() > > vma = find_vma(...) > > if (!vma) /* bail */ > > And maybe add some comments like: > > /* > * Load the current VMA lock sequence - we will detect if anyone concurrently > * locks the VMA after this point. > * Pairs with smp_wmb() in vma_start_write(). > */ > > vm_lock_seq = smp_load_acquire(&vma->vm_lock_seq); > /* > * Now we just have to detect if the VMA is already locked with its current > * sequence count. > * > * The following load is ordered against the vm_lock_seq load above (using > * smp_load_acquire() for the load above), and pairs with implicit memory > * ordering between the mm_lock_seq write in mmap_write_unlock() and the > * vm_lock_seq write in the next vma_start_write() after that (which can only > * occur after an mmap_write_lock()). > */ > > mm_lock_seq = smp_load_acquire(&vma->mm->mm_lock_seq); > > /* I think vm_lock has to be acquired first to avoid the race */ > > if (mm_lock_seq == vm_lock_seq) > > /* bail, vma is write-locked */ > > ... perform uprobe lookup logic based on vma->vm_file->f_inode ... > /* > * Order the speculative accesses above against the following vm_lock_seq > * recheck. > */ > > smp_rmb(); > > if (vma->vm_lock_seq != vm_lock_seq) > > (As I said on the other thread: Since this now relies on > vma->vm_lock_seq not wrapping back to the same value for correctness, > I'd like to see vma->vm_lock_seq being at least an "unsigned long", or > even better, an atomic64_t... though I realize we don't currently do > that for seqlocks either.) > > > /* bail, VMA might have changed */ > > > > The smp_rmb() is needed so that vma->vm_lock_seq load does not get > > reordered and moved up before speculation. > > > > I'm CC'ing Jann since he understands memory barriers way better than > > me and will keep me honest.
On Thu, Aug 15, 2024 at 11:58 AM Jann Horn <jannh@google.com> wrote: > > +brauner for "struct file" lifetime > > On Thu, Aug 15, 2024 at 7:45 PM Suren Baghdasaryan <surenb@google.com> wrote: > > On Thu, Aug 15, 2024 at 9:47 AM Andrii Nakryiko > > <andrii.nakryiko@gmail.com> wrote: > > > > > > On Thu, Aug 15, 2024 at 6:44 AM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > > > > > On Tue, Aug 13, 2024 at 08:36:03AM -0700, Suren Baghdasaryan wrote: > > > > > On Mon, Aug 12, 2024 at 11:18 PM Mateusz Guzik <mjguzik@gmail.com> wrote: > > > > > > > > > > > > On Mon, Aug 12, 2024 at 09:29:17PM -0700, Andrii Nakryiko wrote: > > > > > > > Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access > > > > > > > vma->vm_file->f_inode lockless only under rcu_read_lock() protection, > > > > > > > attempting uprobe look up speculatively. > > Stupid question: Is this uprobe stuff actually such a hot codepath > that it makes sense to optimize it to be faster than the page fault > path? Not a stupid question, but yes, generally speaking uprobe performance is critical for a bunch of tracing use cases. And having independent threads implicitly contending with each other just because of uprobe's internal implementation detail (while conceptually there should be no dependencies for triggering uprobe from multiple parallel threads) is a big surprise to users and affects production use cases beyond just uprobe-handling BPF logic overhead ("useful overhead") they assume. > > (Sidenote: I find it kinda interesting that this is sort of going back > in the direction of the old Speculative Page Faults design.) > > > > > > > > We rely on newly added mmap_lock_speculation_{start,end}() helpers to > > > > > > > validate that mm_struct stays intact for entire duration of this > > > > > > > speculation. If not, we fall back to mmap_lock-protected lookup. > > > > > > > > > > > > > > This allows to avoid contention on mmap_lock in absolutely majority of > > > > > > > cases, nicely improving uprobe/uretprobe scalability. > > > > > > > > > > > > > [...] > > Note: up_write(&vma->vm_lock->lock) in the vma_start_write() is not > > enough because it's one-way permeable (it's a "RELEASE operation") and > > later vma->vm_file store (or any other VMA modification) can move > > before our vma->vm_lock_seq store. > > > > This makes vma_start_write() heavier but again, it's write-locking, so > > should not be considered a fast path. > > With this change we can use the code suggested by Andrii in > > https://lore.kernel.org/all/CAEf4BzZeLg0WsYw2M7KFy0+APrPaPVBY7FbawB9vjcA2+6k69Q@mail.gmail.com/ > > with an additional smp_rmb(): > > > > rcu_read_lock() > > vma = find_vma(...) > > if (!vma) /* bail */ > > And maybe add some comments like: > > /* > * Load the current VMA lock sequence - we will detect if anyone concurrently > * locks the VMA after this point. > * Pairs with smp_wmb() in vma_start_write(). > */ > > vm_lock_seq = smp_load_acquire(&vma->vm_lock_seq); > /* > * Now we just have to detect if the VMA is already locked with its current > * sequence count. > * > * The following load is ordered against the vm_lock_seq load above (using > * smp_load_acquire() for the load above), and pairs with implicit memory > * ordering between the mm_lock_seq write in mmap_write_unlock() and the > * vm_lock_seq write in the next vma_start_write() after that (which can only > * occur after an mmap_write_lock()). > */ > > mm_lock_seq = smp_load_acquire(&vma->mm->mm_lock_seq); > > /* I think vm_lock has to be acquired first to avoid the race */ > > if (mm_lock_seq == vm_lock_seq) > > /* bail, vma is write-locked */ > > ... perform uprobe lookup logic based on vma->vm_file->f_inode ... > /* > * Order the speculative accesses above against the following vm_lock_seq > * recheck. > */ > > smp_rmb(); > > if (vma->vm_lock_seq != vm_lock_seq) > thanks, will incorporate these comments into the next revision > (As I said on the other thread: Since this now relies on > vma->vm_lock_seq not wrapping back to the same value for correctness, > I'd like to see vma->vm_lock_seq being at least an "unsigned long", or > even better, an atomic64_t... though I realize we don't currently do > that for seqlocks either.) > > > /* bail, VMA might have changed */ > > > > The smp_rmb() is needed so that vma->vm_lock_seq load does not get > > reordered and moved up before speculation. > > > > I'm CC'ing Jann since he understands memory barriers way better than > > me and will keep me honest.
diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c index 713824c8ca77..12f3edf2ffb1 100644 --- a/kernel/events/uprobes.c +++ b/kernel/events/uprobes.c @@ -2286,6 +2286,53 @@ static int is_trap_at_addr(struct mm_struct *mm, unsigned long vaddr) return is_trap_insn(&opcode); } +static struct uprobe *find_active_uprobe_speculative(unsigned long bp_vaddr) +{ + const vm_flags_t flags = VM_HUGETLB | VM_MAYEXEC | VM_MAYSHARE; + struct mm_struct *mm = current->mm; + struct uprobe *uprobe; + struct vm_area_struct *vma; + struct file *vm_file; + struct inode *vm_inode; + unsigned long vm_pgoff, vm_start; + int seq; + loff_t offset; + + if (!mmap_lock_speculation_start(mm, &seq)) + return NULL; + + rcu_read_lock(); + + vma = vma_lookup(mm, bp_vaddr); + if (!vma) + goto bail; + + vm_file = data_race(vma->vm_file); + if (!vm_file || (vma->vm_flags & flags) != VM_MAYEXEC) + goto bail; + + vm_inode = data_race(vm_file->f_inode); + vm_pgoff = data_race(vma->vm_pgoff); + vm_start = data_race(vma->vm_start); + + offset = (loff_t)(vm_pgoff << PAGE_SHIFT) + (bp_vaddr - vm_start); + uprobe = find_uprobe_rcu(vm_inode, offset); + if (!uprobe) + goto bail; + + /* now double check that nothing about MM changed */ + if (!mmap_lock_speculation_end(mm, seq)) + goto bail; + + rcu_read_unlock(); + + /* happy case, we speculated successfully */ + return uprobe; +bail: + rcu_read_unlock(); + return NULL; +} + /* assumes being inside RCU protected region */ static struct uprobe *find_active_uprobe_rcu(unsigned long bp_vaddr, int *is_swbp) { @@ -2293,6 +2340,10 @@ static struct uprobe *find_active_uprobe_rcu(unsigned long bp_vaddr, int *is_swb struct uprobe *uprobe = NULL; struct vm_area_struct *vma; + uprobe = find_active_uprobe_speculative(bp_vaddr); + if (uprobe) + return uprobe; + mmap_read_lock(mm); vma = vma_lookup(mm, bp_vaddr); if (vma) {
Now that files_cachep is SLAB_TYPESAFE_BY_RCU, we can safely access vma->vm_file->f_inode lockless only under rcu_read_lock() protection, attempting uprobe look up speculatively. We rely on newly added mmap_lock_speculation_{start,end}() helpers to validate that mm_struct stays intact for entire duration of this speculation. If not, we fall back to mmap_lock-protected lookup. This allows to avoid contention on mmap_lock in absolutely majority of cases, nicely improving uprobe/uretprobe scalability. BEFORE ====== uprobe-nop ( 1 cpus): 3.417 ± 0.013M/s ( 3.417M/s/cpu) uprobe-nop ( 2 cpus): 5.724 ± 0.006M/s ( 2.862M/s/cpu) uprobe-nop ( 3 cpus): 8.543 ± 0.012M/s ( 2.848M/s/cpu) uprobe-nop ( 4 cpus): 11.094 ± 0.004M/s ( 2.774M/s/cpu) uprobe-nop ( 5 cpus): 13.703 ± 0.006M/s ( 2.741M/s/cpu) uprobe-nop ( 6 cpus): 16.350 ± 0.010M/s ( 2.725M/s/cpu) uprobe-nop ( 7 cpus): 19.100 ± 0.031M/s ( 2.729M/s/cpu) uprobe-nop ( 8 cpus): 20.138 ± 0.029M/s ( 2.517M/s/cpu) uprobe-nop (10 cpus): 20.161 ± 0.020M/s ( 2.016M/s/cpu) uprobe-nop (12 cpus): 15.129 ± 0.011M/s ( 1.261M/s/cpu) uprobe-nop (14 cpus): 15.013 ± 0.013M/s ( 1.072M/s/cpu) uprobe-nop (16 cpus): 13.352 ± 0.007M/s ( 0.834M/s/cpu) uprobe-nop (24 cpus): 12.470 ± 0.005M/s ( 0.520M/s/cpu) uprobe-nop (32 cpus): 11.252 ± 0.042M/s ( 0.352M/s/cpu) uprobe-nop (40 cpus): 10.308 ± 0.001M/s ( 0.258M/s/cpu) uprobe-nop (48 cpus): 11.037 ± 0.007M/s ( 0.230M/s/cpu) uprobe-nop (56 cpus): 12.055 ± 0.002M/s ( 0.215M/s/cpu) uprobe-nop (64 cpus): 12.895 ± 0.004M/s ( 0.201M/s/cpu) uprobe-nop (72 cpus): 13.995 ± 0.005M/s ( 0.194M/s/cpu) uprobe-nop (80 cpus): 15.224 ± 0.030M/s ( 0.190M/s/cpu) AFTER ===== uprobe-nop ( 1 cpus): 3.562 ± 0.006M/s ( 3.562M/s/cpu) uprobe-nop ( 2 cpus): 6.751 ± 0.007M/s ( 3.376M/s/cpu) uprobe-nop ( 3 cpus): 10.121 ± 0.007M/s ( 3.374M/s/cpu) uprobe-nop ( 4 cpus): 13.100 ± 0.007M/s ( 3.275M/s/cpu) uprobe-nop ( 5 cpus): 16.321 ± 0.008M/s ( 3.264M/s/cpu) uprobe-nop ( 6 cpus): 19.612 ± 0.004M/s ( 3.269M/s/cpu) uprobe-nop ( 7 cpus): 22.910 ± 0.037M/s ( 3.273M/s/cpu) uprobe-nop ( 8 cpus): 24.705 ± 0.011M/s ( 3.088M/s/cpu) uprobe-nop (10 cpus): 30.772 ± 0.020M/s ( 3.077M/s/cpu) uprobe-nop (12 cpus): 33.614 ± 0.009M/s ( 2.801M/s/cpu) uprobe-nop (14 cpus): 39.166 ± 0.004M/s ( 2.798M/s/cpu) uprobe-nop (16 cpus): 41.692 ± 0.014M/s ( 2.606M/s/cpu) uprobe-nop (24 cpus): 64.802 ± 0.048M/s ( 2.700M/s/cpu) uprobe-nop (32 cpus): 84.226 ± 0.223M/s ( 2.632M/s/cpu) uprobe-nop (40 cpus): 102.071 ± 0.067M/s ( 2.552M/s/cpu) uprobe-nop (48 cpus): 106.603 ± 1.198M/s ( 2.221M/s/cpu) uprobe-nop (56 cpus): 117.695 ± 0.059M/s ( 2.102M/s/cpu) uprobe-nop (64 cpus): 124.291 ± 0.485M/s ( 1.942M/s/cpu) uprobe-nop (72 cpus): 135.527 ± 0.134M/s ( 1.882M/s/cpu) uprobe-nop (80 cpus): 146.195 ± 0.230M/s ( 1.827M/s/cpu) Previously total throughput was maxing out at 20mln/s with 8-10 cores, declining afterwards. With this change, it now keeps growing with each added CPU, reaching 146mln/s at 80 CPUs (this was measured on a 80-core Intel(R) Xeon(R) Gold 6138 CPU @ 2.00GHz). Suggested-by: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrii Nakryiko <andrii@kernel.org> --- kernel/events/uprobes.c | 51 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 51 insertions(+)