Settings

Theme

Cmapv2: A high performance, concurrent map

github.com

41 points by sirgallo 6 months ago · 27 comments

Reader

kiitos 6 months ago

This repo is completely unsound, code like [1] is pervasive and demonstrates a total misunderstanding of what guarantees are provided -- or, really, not provided -- by "atomic" reads of unsafe.Pointer values. Data races everywhere!

Not safe, do not pass go, do not collect $200, absolutely do not use.

[1] https://github.com/sirgallo/cmapv2/blob/280e3017ae4ba212f6f8...

  • sapiogram 6 months ago

    Thank you for debunking it so I didn't have to. I don't think I've ever seen someone post a low-level/concurrent data structure in Go that wasn't wildly unsound, so I assumed this was too.

    • ecshafer 6 months ago

      Go is made by Google. Do they not have someone writing an equivalent of Java.util.concurrent.ConcurrentHashMap?

      • jbn 6 months ago

        or a port of the lock-free hash map from https://preshing.com/20130605/the-worlds-simplest-lock-free-...

        But really, the premise of using a shared map with concurrent readers and writers seems like a good generator of hard-to-reproduce bugs. IMHO shared-nothing (when feasible) is much easier to reason about, and possibly do periodic merging of thread-local updates, but I would avoid concurrent updates entirely (in particular if 2 threads race to update the same key... that goes to deeper design issues in the application).

      • sapiogram 6 months ago

        They do, there's a concurrent hashmap in the standard library. It doesn't get posted here or on Reddit, though.

    • diagraphic 6 months ago

      There are many good implementations.

  • kiitos 6 months ago

    Be more specific? OK

    ---

    CopyNode broken

    `CopyNode` duplicates only the parent; every child pointer is still shared

        nodeCopy.setChildren(make([]*node, len(n.children)))
        copy(nodeCopy.children, n.children) // pointers reused
    
    https://github.com/sirgallo/cmapv2/blob/main/node.go#L11-L17

    Any later mutation (for example `setValue`) writes through those shared pointers, so readers and historical snapshots are modified concurrently -- invalid, data race, memory model violation

    ---

    Bitmap corruption

    `SetBit` uses XOR rather than “set”:

        func SetBit(bitmap uint32, position int) uint32 { return bitmap ^ (1 << position) }
    
    https://github.com/sirgallo/cmapv2/blob/main/utils.go#L41-L4...

    Calling it twice on the same index flips the bit back to 0. During branch-creation on insert and during delete, this function is invoked multiple times on the same index, clearing a bit that should remain set and leaving orphaned children.

    ---

    Invalid assumptions re: interior pointers

    Only the root pointer is read with `atomic.LoadPointer`. All deeper fields like `children[pos]`, `bitmap`, and the byte-slice keys/values, are accessed directly after a successful CAS. Readers therefore race with writers that mutate these fields in place -- race condition, memory model violation, etc.

        pos := cMap.getPosition(node.Bitmap(), hash, level)
        if node.Child(pos).IsLeaf() && bytes.Equal(key, node.Child(pos).Key()) {
            return node.Child(pos).Value()
        }
    
    
    https://github.com/sirgallo/cmapv2/blob/main/operation.go#L5...

    ---

    All xxxRecursive functions rely on those invalid interior pointer assumptions

    Sequence in `putRecursive` / `deleteRecursive` is

        1. `curr := atomic.LoadPointer(ptr)`
        2. Build `nodeCopy`
        3. Recurse; grandchildren are mutated in place
        4. `atomic.CompareAndSwap(ptr, curr, nodeCopy)`
    
    https://github.com/sirgallo/cmapv2/blob/main/operation.go#L1...

    If another goroutine has already swapped in a different copy of `curr` (and mutated it) the CAS still succeeds because the pointer value is unchanged, merging incompatible sub-tries and corrupting the data

    ---

    Use-after-free in sync.Pool

    On CAS failure the freshly built `nodeCopy` is immediately returned to a `sync.Pool` -- undefined behavior

        cMap.pool.PutNode(nodeCopy) // may race with outstanding readers
    
    https://github.com/sirgallo/cmapv2/blob/main/operation.go#L1...

    Other goroutines still holding references to that node can now access a reclaimed object, oops.

    ---

    K/V Aliasing

    Keys and values (both []byte slices, which are not safe for concurrent r/w access) are stored by reference, a mistake:

        n.setKey(key)
        n.setValue(value)
    
    If the caller mutates those slices later (or concurrently in another goroutine), data races ahoy

    ---

    Reader panics, etc.

        - `getRecursive` accesses `children[pos]` without bounds or nil checks, concurrent shrink can make `pos` invalid
        - `GetIndex` allows a negative `shiftSize` once `level >= 7` with `chunkSize = 5`, producing nonsense indices and potential slice-out-of-bounds
    • sirgalloOP 6 months ago

      Hi, thank you for the in depth response. I really needed to hear these things. I have gone ahead and addressed almost all of the issues that you pointed out. I updated the root to be the only point for compare and swap and on mutations I do a clone instead of mutating the shared pointer. I updated set bit from xor to set and have an explicit clearbit fn for deletions. I also updated k/v aliasing to copy slices on write so that old copies do not share ref to same slice. The node pool has been removed completely for the time being as well. I added in additional tests to test for these cases and added in more concurrent tests as well. this was huge feedback and I really appreciate it.

      • kiitos 6 months ago

        The point of enumerating all of those specific issues, wasn't to say "here are some bugs" and if you fix them you're good. It was to say "here are some examples of the much more fundamental problem", which seemed to be a fundamental misunderstanding of the language memory model and the guarantees offered by assignments, atomic.CompareAndSwap, etc., and those operations' interactions with package unsafe.

        For example this code

        https://github.com/sirgallo/cmapv2/blob/6bcaa0253b1b0b261e8a...

        and in particular its use of this code

        https://github.com/sirgallo/cmapv2/blob/6bcaa0253b1b0b261e8a...

        is still completely unsound.

        Looking at *only this code path* -- and there are *many more* --

        ---

        Put

        - Snapshots the current root pointer with atomic.LoadPointer

        - Makes an updated root pointer via putRecursive, given the snapshotted root pointer

        - Spins on a CAS of the root ptr and the updated ptr with runtime.Gosched() between attempts

        ---

        atomic.LoadPointer isn't a real snapshot

        - It's atomic only over the root ptr, not any interior field

        - Those interior fields are mutated in-place via e.g. setBitmap, setChild, etc.

        - Any goroutine can see partial data, violating the memory model, etc.

        ---

        putRecursive is unsound

        - copyNode performs a shallow copy, child pointers are shared, subsequent setChild, extendTable, etc. mutate nodes other goroutines can still hold -- this is a fundamental bug that seems to remain un-addressed from the previous review

        - Those mutations use plain writes (no atomics/locks/etc.) -- data race, memory model violation, etc.

        - Get later returns the internal []byte slice directly -- data race, memory model violation, etc.

        - Newly created nodes are cast to unsafe.Pointer without an atomic store, bypassing the write barrier required by the GC

        ---

        That compareAndSwap is unsound

        - It compares only the root pointer, a shallow copy

        - After a successful CAS other writers can still mutate any shared children (see above), so readers following any shared path see data races, memory model violations, etc.

        - The retry loop can livelock, details elided

        ---

        The implementation still seems to confuse "atomic pointer swap" with "atomic update of a complex, shared value", misunderstands the requirements of the Go memory model, and consistently mis-uses unsafe.Pointer.

        tl;dr here is probably to just stop using package unsafe altogether, until you have some time to properly understand its semantics, requirements, and limitations...

        • sirgalloOP 6 months ago

          Hey, this is going to sound crazy, but I have been looking for someone to critique my code with as much care as you have and give real genuine feedback. I am going to take your input as learning experience.

        • sirgalloOP 6 months ago

          Understood, and again, thank you for picking apart my code. I will take some time to fully understand Go mem model and unsafe package before trying to tackle this problem again. In the meantime, do you have any resources I could take a look at to better my understanding?

    • johnisgood 6 months ago

      > both []byte slices, which are not safe for concurrent r/w access

      You must clone the slice on both write and read, right?

      I get that cloning incurs a memory allocation and a copy operation, but this is the price for safety when concurrent access is possible or your data may be bodified outside your structure.

      You could probably intern immutable keys, or avoid storing if keys already exist and are immutable, or use an object pool (like sync.Pool) to reduce allocations if this happens at scale. Anything else I am missing?

      • sapiogram 6 months ago

        > You must clone the slice on both write and read, right?

        I haven't looked at the code, but that doesn't make sense to me. If you can't read the slice safely, you also can't clone it safely.

        • johnisgood 6 months ago

          So, what are the solutions if they are indeed not safe to do read / write concurrently?

          Like okay, I read "both []byte slices, which are not safe for concurrent r/w access", but then, what is the solution? If the claim is indeed true.

          • sapiogram 6 months ago

            Some options:

            - Make sure no one mutates the slice ever, making it safe to read

            - Guarding the slice behind a mutex, requiring anyone who reads or writes it to lock the mutex first

            - Using some kind of thread-safe slice implementation

            • sirgalloOP 6 months ago

              I went with option 1.

              On a mutation, I do a complete node copy where I also copy the key/value slices. When I set a child node for the first time or update a child, I create a branch new leaf node with a copy of the key/value. This way previous nodes maintain the original copy.

derekperkins 6 months ago

Here's a better implementation if anyone is interested (not mine) https://github.com/puzpuzpuz/xsync#map

sirgalloOP 6 months ago

Performance comparisons are made against go sync.Map, with cmapv2 on par or sometimes exceeding is on different workloads. It is both lock-free and thread safe, using atomic operations. It also supports sharding out of the box. cmapv2 with 10m k/v pairs where keys and values are 64bytes was able to achieve 1.5m w/s and 3m r/s.

gttalbot 6 months ago

Is it anywhere close to swisstables?

  • vlovich123 6 months ago

    No because swisstables generally don't do concurrency (i.e. concurrency ==> atomics which are inherently more expensive due to HW reasons).

Keyboard Shortcuts

j
Next item
k
Previous item
o / Enter
Open selected item
?
Show this help
Esc
Close modal / clear selection