|
|
<!DOCTYPE HTML PUBLIC "-//IETF//DTD HTML//EN"><html><head><title>GeorgeRe's ToDo List for LKRhash</title><link rel=Stylesheet type="text/css" media=all href="./docs/lkr.css"></head>
<body>
<h1>GeorgeRe's ToDo List for LKRhash</h1>
<ul> <li> <a href="#testing"> Testing</a></li> <li> <a href="#kernel-mode"> Kernel Mode</a></li> <li> <a href="#lock-implementation"> Lock Implementation</a></li> <li> <a href="#iis6"> IIS 6.0</a></li> <li> <a href="#misc"> Miscellaneous</a></li> <li> <a href="#hashfn"> Hash Functions</a></li> <li> <a href="#docs"> Documentation</a></li></ul>
<p>These file is primarily for George Reilly's benefit, so don'tworry if you don't understand all of the points.</p>
<a name="testing"><h2>Testing</h2></a>
<ul> <li> Write a C test, to directly test the C API. HashTest half does it, as it can use the public API and TypedHashTable.
<li> Build a publicly distributable test program, as a sample for customers. E.g., a dictionary component for ASP.
<li> hashtest: modify so that each thread can work on a completely separate hashtable, instead of sharing a global hashtable. This will allow us to discover the maximum possible scaling, which is probably less than the `ideal' scaling, #CPUs * 1-thread-perf.
<li> hashtest\kernel: split into driver and usermode test program. Usermode test program should take care of parsing the arguments and loading the test data into memory, then use an ioctl to pass it down to the kernel driver. The only code in the kernel driver,apart from LKRhash itself, should be the goo to crack the ioctl and DriverEntry. Usermode should print results.
<li> Check that a numeric key of 0 actually works in debug version.
<li> Better tests for ApplyIf family</ul>
<a name="kernel-mode"><h2>Kernel Mode</h2></a>
<ul> <li> kernel locks: should they block at all? consider implications if running on some usermode program's thread, or of running at DISPATCH_LEVEL. Add support for using NtDelayExecution (implements Sleep) in SwitchOrSleep.
<li> Fix global objects in driver, including the global list of tables; i.e., put in the necessary magic to ensure that constructors and destructors of globals objects get called.
<li> implement kernel-mode version of !lkrhash
<li> Write a template wrapper for operator new/delete, keyed on (Non)PagedPool.
<li> Think about running at DISPATCH_LEVEL instead of PASSIVE_LEVEL. Which functions should be pageable vs. non-pageable? Memory allocators? Lock types?
<li> Use <i>zwapi.h</i>? Which set of NT headers should we rely upon?
<li> Memory allocation pool is a parameter for LKRhash public constructors (and LKR_Initialize), but it's ignored. Need virtual base for allocators? Note: m_paDirSegs allocates an array of CDirEntry*s.
<li> DONE: Control debug spew: provide some way of setting g_fDebugOutputEnabled. (See flags to LKR_Initialize)</ul>
<a name="lock-implementation"><h2>Lock Implementation</h2></a>
<ul> <li> More volatiles for Lock_AtomicCompareAndSwap calls.
<li> Put delays into the inner loop of the spins, to see if that reduces cache line sloshing. See BLAM paper.
<li> Use SleepEx(ms, TRUE) instead of Sleep(ms) to enable APCs. Does this leave any danger of deadlocking a COM STA thread?
<li> Provide some way to expose the bit31 (precreate event) functionality of InitializeCriticalSectionAndSpinCount. Also, this function can fail.
<li> Document that the locks are not suitable for cross-process use.
<li> Enable per-class instrumentation for locks, as opposed to the current all-or-nothing system for per-lock instrumentation. Mostly present already, just needs a little factoring. Keep track of wall-clock time and sleep time too.
<li> Check asserts for Is(Read|Write)Unlocked. Disable Is(Read|Write)(L|Unl)ocked except in debug builds. They're fundamentally broken. In general, you can't tell if this thread owns the lock for reading (sometimes, not even for writing). It's even harder to tell that you DON'T own the lock, especially when some other thread does.
<li> Locks code. Move all member function implementations and enums into locks.cpp. Locks.h should be opaque declarations only.
<li> Experiment with using bitfields or a union of WORDS and BYTEs, as opposed to the current explicit masking and shifting.
<li> True waiter count in hiword? As opposed to waiters + writer, as currently.
<li> Add TryConvertSharedToExclusive member to MRSW locks.
<li> Look at the new EX_PUSH_LOCK in \nt\public\internal\base\inc\ntosp.h. Wrap in kLocks.h.
<li> Is the current kernel mode GetCurrentThreadId adequate; i.e., is the Cid unique across all processes, including the System one?
<li> Change Lock_AtomicIncrement to use `lock inc'. Ditto for Decrement. Don't need result of operation.
<li> Remove x86 `lock' prefixes and measure difference on a 1P machine.
<li> Experiment with putting the writer recursion count in m_lRW: 0xFFFF, 0xFFFE, ..., -n.
<li> Rewrite a few key functions, such as CSmallSpinLock::WriteLock or CReaderWriterLock3::ReadOrWriteLock, in x86 assembler. Keep portable implementation, of course.
<li> Add safe versions of the main entry points that do a lot more error checking: e.g., SafeWriteUnlock should check that the current thread actually owns the lock; check for over- or underflows; etc.
<li> Orphaned lock detection?
<li> Build locks as a separate statically linked library. Or dll?
<li> Add per-class initialize and cleanup functions, to be called from LKR_Initialize
<li> Make the locks throw exceptions (C++ or SEH?) in "impossible" conditions; e.g., recursively acquiring a CSmallSpinLock (=> instant deadlock), acquiring a writelock too many times, freeing a lock too many times.
<li> Experiment with queued spinlocks: see \\bustard\contrib\ReneS\mcslock
<li> Refcount Locks_Initialize and Locks_Cleanup?
<li> Move CKSpinLock, CEResource, and CFastMutex into <Locks.h>.
<li> Build another variant of the locks that waits on handles instead of using SwitchToThread. Perhaps a pool of handles?
<li> Add a timeout feature to Try(Read|Write)Lock
<li> Deadlock detection?
<li> In the Spin routines, keep track of how long it's been since spinning started. DebugBreak if we've spun for a long time (e.g., 10 minutes)
<li> For debugging purposes, keep track of all read owners of a multireader lock. Hang something out of TLS to keep track of all locks acquired by this thread.
<li> InitializeCriticalSectionAndSpinCount is broken on Win98. Declared with wrong prototype or something. Only do a GetProcAddr for it on NT.
<li> DONE: Put an InterlockedCompareExchange at the beginning of each outer loop in the spin routines, to ensure that there's an unconditional memory barrier.
<li> DONE: Increase SL_OWNER_BITS from 4 to 8 to accommodate likely scenarios with locking iterators.
<li> DONE: Sprinkle KeEnterCriticalRegion (and KeLeaveCriticalRegion) in the various locks, to prevent APCs being delivered that might suspend the thread that holds the lock.
<li> NOT NEEDED, as I realized that the return value can be stored as a member variable that's written after the lock is acquired. WriteLock: make all return a value instead of `void' that's passed into WriteUnlock. Ditto for ReadLock, etc. Needed for CKSpinLock and OldIrql. <sigh/></ul>
<a name="iis6"><h2>IIS 6.0</h2></a>
<ul> <li> Reduce the three versions of LKRhash in IIS6 to just this one. Definitely get rid of the IISUTIL version, even if not allowed to fix the IISRTL/LISRTL version.
<li> Remove all uses of the deprecated iterators from the IIS6 code base.
<li> Port !lkrhash to ulkd. Update !cache or whatever it's called.
<li> Copy the debug instrumentation from http.sys into CKSpinLock and CEResource.</ul>
<a name="misc"><h2>Miscellaneous</h2></a>
<ul> <li> Don't __declspec(dllexport) CLKRLinearHashTable. Don't __declspec(dllexport) whole classes, just public methods.
<li> Use ASSERTIONAL, PRECONDITION, etc, specialized macros.
<li> Add typedefs for RecordPointer and HashSignature, to ease porting to AWE-type environment.
<li> Run hashtest under the AppVerifier.
<li> Compile with PREFast and /W4
<li> Get rid of IsBad*Ptr wrappers in irtldbg.h. Clean up irtldbg.h and irtlmisc.h.
<li> Experiment with __assume().
<li> Get rid of all calls to Interlocked* in src\LKR-*.cpp. Replace with trivial methods. Have the methods use the lock type to decide whether to use interlocked or non-threadsafe operations. Goal is to have non-threadsafe version use no interlocked ops whatsoever.
<li> Intel's Hyper-Threading presentation says to use a "Fill line size" of 128 bytes.
<li> Provide BulkAdd and BulkDelete functions that work in terms of a collection of records. It should be cheaper to do _Expand (or _Contract) once for all affected buckets, than to have to do them after each insertion (deletion).
<li> Allow a NULL AddRefRecord(), so that reference counting can be entirely dispensed with.
<li> placement new for inline SubTable[0]
<li> Clear() should abandon immediately if m_cRecords==0.
<li> Replace use of NODES_PER_CLUMP in functions with a call to inline function _NodesPerClump(). This will allow experiments with NodeClumps whose size is set at runtime, not compile time. Also need to access m_dwKeySigs[i] and m_pvNode[i] through accessor functions.
<li> :%s/CListEntry/LIST_ENTRY/g. This allows us to get rid of the global ctors/dtors (which are a pain in the kernel) and just initialize/test empty in LKR_Initialize and LKR_Terminate.
<li> Use a custom allocator. CMediumSegment is 4KB on x86, but the heap overhead means that it's actually two pages, rather than exactly one page.
<li> 'typedef INT_PTR NodeIndex' so that loop induction works better on Win64.
<li> Experiment with disabling linear hashing's expansion and contraction. This would mean that we could avoid taking the table lock to calculate the bucket address. However, this would only work if the table lock is not exposed or explicitly disabled. It would also require the user to set the table size correctly in the constructor, if they don't use LK_{SMALL|MEDIUM|LARGE}_TABLESIZE.
<li> In Win64/Debug version, InsertRecord should check for 32-bit overflow of m_cRecords. (Though it's still improbable. A record takes 8 bytes for the pointer + 4 bytes for the hash signature plus some amortized overhead, which means that the storage for the table itself would be more than 12 * 4GB. And that excludes any storage for the > 2^32 records.)
<li> Add a "contention level" flag to the constructor. If the number of subtables is not specified explicitly (i.e., LK_DFLT_NUM_SUBTBLS is passed to ctor), we key the number of subtables off LK_TABLESIZE multiplied by a function of the number of CPUs. However, we only need a lot of subtables under two circumstances: (a) many millions of elements (esp. on Win64, where total number of elements might someday exceed 2^32), or (b) high contention. There isn't necessarily a correlation between large table size and high contention. With the multi-reader locks, high contention only arises if there are a lot of insertions/deletions. If the table is not modified much after it's built, contention shouldn't be an issue and there's no real advantage to having a lot of subtables.
<li> [Suggestion by BAlam] Provide a way to "seal" hashtable. Once the records have been inserted, call Seal(). All valid operations thereafter (except dtor) are readonly, so no table or bucket locks need be taken.
<li> Provide a way to disable all locking. Caller takes responsibility for guaranteeing that all operations are threadsafe, either because table has thread affinity or because a higher-level lock is being used.
<li> Add optional refcount parameter to template wrapper. Refcount all operations, hide dtor, and provide a Destroy method instead. Fix issues with some tables not shutting down gracefully.
<li> Benchmark CLKRHashTable with one subtable vs. CLKRLinearHashTable.
<li> Experiment with faster functions for finding subtable, esp. when number of subtables is not a power of 2. Probably just use middle 6 bits of ((hash << 13) - hash) and use that to index into a lookup table of byte-sized remainders.
<li> Use .w -> .h stuff (hsplit) to filter out MS-confidential stuff from the headers, so that LKRhash.h and Locks.h can be rendered fit for public consumption?
<li> Check that CLKRLinearHashTable can still be used as the base class for CTypedHashTable.
<li> Add a default constructor to CTypedHashTable that uses a <code>static const char* ClassName()</code> method.
<li> Should operator== and operator!= be made friend functions for the iterators, instead of members?
<li> Add version.subversion number to CLKRLinearHashTable
<li> Ensure all relevant fields are printed in !lkrhash
<li> Add -k flag to !lkrhash to enumerate all known custom extensions.
<li> The publicly exposed versions of the Lock accessors should use the Safe versions of the lock entrypoints. The internal versions (as called by the LKRhash code itself) can continue to use the fast versions.
<li> Remove the deprecated iterators from the main code before releasing it. Bundle them in a separate subdirectory for the few holdouts. Clarify that they will not be released again.
<li> Double-check <code>nCmp = pstrKey1->m_cch - pstrKey2->m_cch;</code> in WordHash.h
<li> Use C#-style XML documentation comments
<li> Write some documentation. Does there need to be a sanitized version for external consumption? Think about using XSL to achieve this.
<li> Write a tutorial. Perhaps a little phone book class (email => phone number, then firstname+lastname => phone number). Extend it with a reverse map (phone number => name); mention canonicalization
<li> See if division or multiplication is better in InsertRecord expansion test. Ditto for DeleteKey/Record. Experiment with using shifts for division. For example, NumRecords/7 ~= NumRecords * 9 / 64 = (NumRecords >> 3) + (NumRecords >> 6). Don't need perfect mathematical accuracy.
<li> Don't use modulo in CLKRHashTable::_SubTable, if m_cSubTables is not a power of 2. Either round the m_cSubTables up or down to the nearest power of 2, or compute a modulo table.
<li> Add a context pointer to all of the *Key/Apply functions that will be passed down to CompareKeys and HashKey. E.g., do a case-sensitive or -insensitive match.
<li> Make the m_pfn's be const pointers and member-initialize them.
<li> Get rid of __LKRHASH_NO_NAMESPACE__; i.e., require the namespace.
<li> Refcount LKR_Initialize and LKR_Terminate?
<li> Add some flags to LKR_Initialize: default size, output tracing, etc
<li> Should ApplyIf(LKP_DELETE) call the Action function before deleting? Or add LKP_PERFORM_DELETE[_STOP]
<li> ApplyIf should not take the bucket lock. This would permit recursive calls to routines that need to take the bucket lock. Careful! There may be race conditions, if another thread holds an incompatible bucket lock but has been suspended.
<li> Implement fMultiKeys to provide support for multiple, identical keys. Needed for EqualRange, hash_multiset, and hash_multimap. See CLKRLinearHashTable::_InsertRecord for details on what needs to be changed. Be sure to actually compare keys for equality: matching hash signatures is not enough.
<li> Provide FindMulti, DeleteMulti(OUT Record*** pppRecords OPTIONAL), and FreeMulti.
<li> Provide implementations of the STL collection classes: hash_map, hash_set, hash_multimap, and hash_multiset. Must provide full implementation of STL iterator semantics (pair<key,value>). Use the SGI version of hash_map as an inspiration, not the Dinkumware version. See Austern's book.
<li> Provide const_iterators for STL-style iterators. But see what Scott Meyers has to say about this in "Effective STL."
<li> Consider providing some kind of implicit locking with STL-style iterators. Either per-subtable or all subtables.
<li> We need to call _ExtractKey a lot. Would it be better to cache the DWORD_PTR in each node? On the other hand, _ExtractKey is usually very cheap (typically add the offset of the key to the base address of the record), so the space-time tradeoff may not be worth it.
<li> Add debug print routines for the other enumerations and for lock state to LKRhash. Place in a separate file that can be shared with lkrdbg.
<li> Provide mapping functions, LKRetcodeToHResult, LKRetcodeToWin32Error, and LKRetcodeToNtStatus, for canonical translations.
<li> Build lkrhash as a separate statically linked library
<li> Make CBucket be a union with BYTE[BUCKET_BYTE_SIZE], to improve cache-line alignment. But really ought to be using an allocator that produces page-aligned locations for maximal benefit: i.e., something like RockAll instead of the NT heap or the CRT.
<li> Better Statistics: #buckets, density, by subtable
<li> Experiment with new hash function from Paul Larson and others.
<li> Experiment with GUID hashing functions. Supposedly the first DWORD is enough. Rip all the IIDs, CLSIDs, etc out of the registry and create random GUIDs with UuidCreate() and MAC-based GUIDs with UuidCreateSequential() to test the distributions.
<li> Public API in LKR-hash.h: remove dependency on irtlmisc.h and irtldbg.h
<li> Coalesce CLKRLinearHashTable_Iterator members and CLKRHashTable_Iterator members for space; i.e., put `short m_iNode', `BYTE m_LockType', and `BYTE m_ist' together in CLKRLinearHashTable_Iterator.
<li> Add some assertions in terms of MAX_DIRSIZE_BITS and NODE_CLUMP bits.
<li> Use IsBadCodePtr to validate callback functions in constructor. Of course, this is no guarantee that they'll still be good functions later.
<li> Make LKRhash exception-safe. What happens if callback routines (AddRefRecord/ExtractKeys or ApplyIf) access violate (throw an SEH exception) or throw a C++ exception? Table and bucket locks won't be released, state variables may be corrupted, etc. LKRhash code should never throw any kind of exception
<li> Add throw() specifications to appropriate functions?
<li> Add some kind of auto object for readlocking or writelocking a table, so that the table automatically gets unlocked by auto-obj's destructor.
<li> Use auto_ptrs.
<li> Port to C# (Chris Tracy has started on this). Andre Rosenthal has started a port to Managed C++
<li> DONE: Still need NULL implementations for copy ctor and op=? Or is this an obsolete vestige of the long-gone templated inner classes?
<li> DONE: Factor out memory allocation stuff into LKR-mem.cpp; initialization stuff into LKR-init.cpp; LKR-locks.cpp, etc.
<li> DONE: Add TryWriteLock and TryReadLock methods to table. Also ConvertExclusiveToShared and vice versa.
<li> DONE: Couch tests in _IsBucketChainMultiKeySorted et al in terms of a macro, CheckAndAdd.
<li> DONE: Experiment with having no bucket locks whatsoever, just subtable locks. This should make operations a little faster, but presumably will hurt multiprocessor scalability. (Have to have table locks in multithreaded version as linear hashing's expansion/contraction means that bucket address calculation is a function of several variables that can change at any time.)
<li> DONE: Add some instrumentation: #allocs, #expands, #contracts, #ExtractKeys, etc.
<li> DONE: Reduce MIN_DIRSIZE to 4 and inline m_aDirSegs[4]. Only allocate if directory gets bigger than MIN_DIRSIZE.
<li> DONE: Add an index member to subtable
<li> DONE: Inline the array of pointers to subtables within CLKRHashTable, instead of dynamically allocating it.
<li> DONE: Make AddRefRecord return the new reference count. Should it take a `const void*' or a `void*'? (TypedHashTable takes a non-const `Record*'.)
<li> DONE: Provide a DeleteKey variant that returns a pointer to the record that is being removed from the table. In this case, instead of releasing reference, would implicitly transfer ownership of the record to the caller.
<li> DONE: Inline _FindBucket by hand into Delete(Key|Record), InsertRecord, and Find(Key|Record).
<li> DONE: Replace <code>IRTLVERIFY(x)</code> with <code>if (!x) IRTLASSERT(! "message")</code>.
<li> DONE: InsertRecord should not LKAR_INSERT_RELEASE if record addresses are same. This guarantees that calls to _AddRefRecord with LKAR_MIN_DELETE_FROM_TABLE <= lkar <= LKAR_MAX_DELETE_FROM_TABLE works.
<li> DONE: If m_cSubTables==1 don't bother to scramble the hash index in CLKRHashTable::_SubTable
<li> DONE: break apart lkrhash.cpp: iterators, applyif, stats, etc
<li> DONE: Always step forward through all subtables, to avoid possible deadlock, when acquiring subtable locks; i.e., ensure that we have a valid lock hierarchy.
<li> DONE: Test new contraction algorithm.
<li> DONE: sublocks for ApplyIf
<li> DONE: Provide a C API wrapper</ul>
<a name="hashfn"><h2>Hash Functions</h2></a>
<ul> <li> Follow up on case-insensitive hash function for Unicode. Note that (wch & 0xFFDF) is inadequate. Note: NTFS keeps a per-volume UpcaseMap for Unicode.
<li> Provide A/W string hash functions that take a count, instead of relying on zero-termination.</ul>
<a name="docs"><h2>Documentation</h2></a>
<ul> <li> Write a tutorial</ul></ul>
<hr>
<address></address>
<!-- hhmts start -->Last modified:Fri Feb 15 13:15:54 2002 <!-- hhmts end -->
</body></html>
|
