Add mechanism to support invalidateAll and deleteAll

Here's a very rough patch which implements a randomly generated key pre/suffix per bin, which we then regenerate when invalidateAll() is called.

Tried a few different ways of using state properly (e.g. injecting it in services) but that seems to lead to "Circular reference" issues.

Here's an alternative approach; using memcache itself to store the key suffixes. Seems quite a bit simpler, and AFAICS it works.

Seems like a good approach @mcdruid if no one else have time to review this I try do it the coming week.

Tagging this as a beta blocker, per #2989594: [META] Create a stable release of Memcache module for Drupal 8.

Ahem.

Overall looks almost RTBC.

1. +++ b/src/MemcacheBackend.php
   @@ -1,13 +1,8 @@
   -use Drupal\Core\Cache\Cache;
   +use Drupal\Component\Utility\Crypt;
   

   no longer needed

2. +++ b/src/MemcacheBackend.php
   @@ -316,7 +325,62 @@ class MemcacheBackend implements CacheBackendInterface {
   +  protected function timeIsGreaterThanBinDeletionTime(float $item_microtime) {
   ...
   +  protected function getBinLastDeletionTime() {
   ...
   +  protected function updateBinLastDeletionTime() {
   ...
   +  protected function ensureBinDeletionTimeIsSet() {
   ...
   +  protected function getBinLastDeletionTimeKey() {
   

   Let's mark all of that as @internal.

   I have not given up hope to use an injected timestamp service instead ;).

3. +++ b/src/MemcacheBackend.php
   @@ -316,7 +325,62 @@ class MemcacheBackend implements CacheBackendInterface {
   +    if (!$last_bin_deletion) {
   +      return FALSE;
   +    }
   

   Should create the last_bin_deletion time here - though it's really unlikely we would not have one here as we do it on __construct(), so likely not needed.

4. +++ b/src/MemcacheBackend.php
   @@ -316,7 +325,62 @@ class MemcacheBackend implements CacheBackendInterface {
   +    $microtime = microtime(TRUE);
   +
   +    $return = $this->memcache->set($this->getBinLastDeletionTimeKey(), $microtime);
   +
   +    $this->lastBinDeletionTime = $microtime;
   

   Are we sure we want to use microtime for this?

   Usually timedrift can mean cache corruption, but using normal seconds we just loose some data.

   There was a patch in the D7 queue on that to convert to microtime [which gave horrible effects], so I am tending to use normal seconds instead.

As discussed we'll continue to use microtime() rounded plus some skew like ChainedFast.

As we are dealing with volatile storage, which can evict keys without notice, we can't rely on atomic increment / decrement behavior unfortunately.

This is the hunk from ChainedFastBackend for reference:

  protected function markAsOutdated() {
    // Clocks on a single server can drift. Multiple servers may have slightly
    // differing opinions about the current time. Given that, do not assume
    // 'now' on this server is always later than our stored timestamp.
    // Also add 1 millisecond, to ensure that caches written earlier in the same
    // millisecond are invalidated. It is possible that caches will be later in
    // the same millisecond and are then incorrectly invalidated, but that only
    // costs one additional roundtrip to the persistent cache.
    $now = round(microtime(TRUE) + .001, 3);

Approach here looks right to me, agreed with tags for invalidations and state for deletions.

#12 applied fine against the current version.

With PHP 5.6 I get the following error:

PHP Catchable fatal error:  Argument 1 passed to Drupal\memcache\MemcacheBackend::timeIsGreaterThanBinDeletionTime() must be an instance of Drupal\memcache\float, double given, called in MemcacheBackend.php on line 109

if (!$this->timeIsGreaterThanBinDeletionTime($result->created)) {
...
protected function timeIsGreaterThanBinDeletionTime(float $item_microtime) {

When I remove the float it seems to work fine.

A crude performance test with "time drush cr" shows that the new version is about 10% slower than the old one (with global flush).

Thanks for the work! I am happy to see that memcache is moving forward.