On 21/05/16 05:48, Emilio G. Cota wrote:
> On Sat, May 21, 2016 at 01:13:20 +0300, Sergey Fedorov wrote:
>> Although the API is mostly intuitive some kernel-doc-style comments
>> wouldn’t hurt, I think. ;-)
> The nit that bothered me is the "external lock needed" bit, but it's
> removed by the subsequent patch (which once it gets reviewed should be merged
> onto this patch); I think the interface is simple enough that comments
> would just add noise and maintenance burden. Plus, there are tests under
> tests/.
The interface is simple enough but e.g. the return value convention for
some of the functions may not be clear from a first glance. Regarding
maintenance burden, as soon as we have a good stable API it shouldn't be
painful.
> (snip)
>>> +/* define these to keep sizeof(qht_bucket) within QHT_BUCKET_ALIGN */
>>> +#if HOST_LONG_BITS == 32
>>> +#define QHT_BUCKET_ENTRIES 6
>>> +#else /* 64-bit */
>>> +#define QHT_BUCKET_ENTRIES 4
>>> +#endif
>>> +
>>> +struct qht_bucket {
>>> + QemuSpin lock;
>>> + QemuSeqLock sequence;
>>> + uint32_t hashes[QHT_BUCKET_ENTRIES];
>>> + void *pointers[QHT_BUCKET_ENTRIES];
>>> + struct qht_bucket *next;
>>> +} QEMU_ALIGNED(QHT_BUCKET_ALIGN);
>>> +
>>> +QEMU_BUILD_BUG_ON(sizeof(struct qht_bucket) > QHT_BUCKET_ALIGN);
>> Have you considered using separate structures for head buckets and
>> non-head buckets, e.g. "struct qht_head_bucket" and "struct
>> qht_added_bucket"? This would give us a little more entries per cache-line.
> I considered it. Note however that the gain would only apply to
> 32-bit hosts, since on 64-bit we'd only save 8 bytes but we'd
> need 12 to store hash+pointer. (lock+sequence=8, hashes=4*4=16,
> pointers=4*8=32, next=8, that is 8+16+32+8=32+32=64).
>
> On 32-bits with 6 entries we have 4 bytes of waste; we could squeeze in
> an extra entry. I'm reluctant to do this because (1) it would complicate
> code and (2) I don't think we should care too much about performance on
> 32-bit hosts.
Fair enough.
> (snip)
>>> +static inline
>>> +void *qht_do_lookup(struct qht_bucket *head, qht_lookup_func_t func,
>>> + const void *userp, uint32_t hash)
>>> +{
>>> + struct qht_bucket *b = head;
>>> + int i;
>>> +
>>> + do {
>>> + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
>>> + if (atomic_read(&b->hashes[i]) == hash) {
>>> + void *p = atomic_read(&b->pointers[i]);
>> Why do we need this atomic_read() and other (looking a bit inconsistent)
>> atomic operations on 'b->pointers' and 'b->hash'? if we always have to
>> access them protected properly by a seqlock together with a spinlock?
> [ There should be consistency: read accesses use the atomic ops to read,
> while write accesses have acquired the bucket lock so don't need them.
> Well, they need care when they write, since there may be concurrent
> readers. ]
Well, I see the consistency now =)
> I'm using atomic_read but what I really want is ACCESS_ONCE. That is:
> (1) Make sure that the accesses are done in a single instruction (even
> though gcc doesn't explicitly guarantee it even to aligned addresses
> anymore[1])
> (2) Make sure the pointer value is only read once, and never refetched.
> This is what comes right after the pointer is read:
>> + if (likely(p) && likely(func(p, userp))) {
>> + return p;
>> + }
> Refetching the pointer value might result in us passing something
> a NULL p value to the comparison function (since there may be
> concurrent updaters!), with an immediate segfault. See [2] for a
> discussion on this (essentially the compiler assumes that there's
> only a single thread).
>
> Given that even reading a garbled hash is OK (we don't really need (1),
> since the seqlock will make us retry anyway), I've changed the code to:
>
> for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
> - if (atomic_read(&b->hashes[i]) == hash) {
> + if (b->hashes[i] == hash) {
> + /* make sure the pointer is read only once */
> void *p = atomic_read(&b->pointers[i]);
>
> if (likely(p) && likely(func(p, userp))) {
>
> Performance-wise this is the impact after 10 tries for:
> $ taskset -c 0 tests/qht-bench \
> -d 5 -n 1 -u 0 -k 4096 -K 4096 -l 4096 -r 4096 -s 4096
> on my Haswell machine I get, in Mops/s:
> atomic_read() for all 40.389 +- 0.20888327415622
> atomic_read(p) only 40.759 +- 0.212835356294224
> no atomic_read(p) (unsafe) 40.559 +- 0.121422128680622
>
> Note that the unsafe version is slightly slower; I guess the CPU is trying
> to speculate too much and is gaining little from it.
>
> [1] "Linux-Kernel Memory Model" by Paul McKenney
> http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4374.html
> [2] https://lwn.net/Articles/508991/
Okay.
> (snip)
>>> +/*
>>> + * Find the last valid entry in @head, and swap it with @orig[pos], which
>>> has
>>> + * just been invalidated.
>>> + */
>>> +static inline void qht_bucket_fill_hole(struct qht_bucket *orig, int pos)
>>> +{
>>> + struct qht_bucket *b = orig;
>>> + struct qht_bucket *prev = NULL;
>>> + int i;
>>> +
>>> + if (qht_entry_is_last(orig, pos)) {
>>> + return;
>>> + }
>>> + do {
>>> + for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
>> We could iterate in the opposite direction: from the last entry in a
>> qht_bucket to the first. It would allow us to fast-forward to the next
>> qht_bucket in a chain in case of non-NULL last entry and speed-up the
>> search.
> But it would slow us down if--say--only the first entry is set. Also
> it would complicate the code a bit.
>
> Note that with the resizing threshold that we have, we're guaranteed to
> have only up to 1/8 of the head buckets full. We should therefore optimize
> for the case where the head bucket isn't full.
Okay.
Kind regards,
Sergey
