- Loading...
Page History
...
RFE: 8153224 Monitor deflation prolong safepoints
https://bugs.openjdk.java.net/browse/JDK-8153224
Full Webrev: http://cr.openjdk.java.net/~dcubed/8153224-webrev/10-for-jdk1417-for-jdk15+24.v2.0715.full/
Inc Webrev: http://cr.openjdk.java.net/~dcubed/8153224-webrev/1017-for-jdk14jdk15+24.v2.0715.inc/
Background
This patch for Async Monitor Deflation is based on Carsten Varming's
...
The current idle monitor deflation mechanism executes at a safepoint during cleanup operations. Due to this execution environment, the current mechanism does not have to worry about interference from concurrently executing JavaThreads. Async Monitor Deflation uses the ServiceThread to deflate idle monitors so the new mechanism has to detect interference and adapt as appropriate. In other words, data races are natural part of Async Monitor Deflation and the algorithms have to detect the races and react without data loss or corruption.
Key Parts of the Algorithm
1) Deflation With Interference Detection
ObjectSynchronizer::deflate_monitor_using_JT() is the new counterpart to ObjectSynchronizer::deflate_monitor() and does the heavy lifting of asynchronously deflating a monitor using a three part prototcol:
Async Monitor Deflation is performed in two stages: stage one performs the two part protocol described in "Deflation With Interference Detection" below and moves the async deflated ObjectMonitors from an in-use list to a global wait list; the ServiceThread performs a handshake (or a safepoint) with all other JavaThreads after stage one is complete and that forces any racing threads to make forward progress; stage two moves the ObjectMonitors from the global wait list to the global free list. The special values that mark an ObjectMonitor as async deflated remain in their fields until the ObjectMonitor is moved from the global free list to a per-thread free list which is sometime after stage two has completed.
Key Parts of the Algorithm
1) Deflation With Interference Detection
ObjectSynchronizer::deflate_monitor_using_JT() is the new counterpart to ObjectSynchronizer::deflate_monitor() and does the heavy lifting of asynchronously deflating a monitor using a two part prototcol:
- Setting a NULL owner field to DEFLATER_MARKER with cmpxchg() forces any contending thread through the slow path. A racing thread would Setting a NULL owner field to DEFLATER_MARKER with cmpxchg() forces any contending thread through the slow path. A racing thread would be trying to set the owner field.
- Making a zero ref_count contentions field a large negative value with cmpxchg() forces racing threads to retry. A racing thread would would be trying to increment the ref_count field.If the owner field is still equal to DEFLATER_MARKER, then we have won all the races and can deflate the monitorcontentions field.
If we lose any of the races, the monitor cannot be deflated at this time.
Once we know it is safe to deflate the monitor (which is mostly field resetting and monitor list management), we have to restore the object's header. That's another racy operation that is described below in "Restoring the Header With Interference Detection".
The setting of the special values that mark an ObjectMonitor as async deflated and the restoration of the object's header comprise the first stage of Async Monitor Deflation.
2) Restoring the Header With Interference Detection
ObjectMonitor::install_displaced_markword_in_object() is the new piece of code that handles all the racy situations with restoring an object's header asynchronously. The function is called from two three places (deflation and saving an ObjectMonitor* in an ObjectMonitorHandle). The restoration protocol for the object's header uses the mark bit along with the hash() value staying at zero to indicate that the object's header is being restored, ObjectMonitor::enter(), and FastHashCode). Only one of the possible racing scenarios can win and the losing scenarios all adapt to the winning scenario's object header value.
3) Using "owner" or "
...
contentions" With Interference Detection
Various code paths have been updated to recognize an owner field equal to DEFLATER_MARKER or a negative ref_count contentions field and those code paths will retry their operation. This is the shortest "Key Part" description, but don't be fooled. See "Gory Details" below.
An Example of ObjectMonitor Interference Detection
ObjectMonitor::save_om_ptrenter() is used to safely save an ObjectMonitor* in an ObjectMonitorHandlecan change an idle monitor into a busy monitor. ObjectSynchronizer::deflate_monitor_using_JT() is used to asynchronously deflate an idle monitor. save_om_ptrenter() and deflate_monitor_using_JT() can interfere with each other. The thread calling save_om_ptrenter() (T-saveenter) is potentially racing with another JavaThread (T-deflate) so both threads have to check the results of the races.
Start of the Race
T-save enter ObjectMonitor T-deflate
------------------------ +-----------------------+ ----------------------------------------
save_om_ptrenter() { | owner=NULL | deflate_monitor_using_JT() {
1> atomic inc ref_countcontentions | ref_countcontentions=0 | 1> cmpxchg(DEFLATER_MARKER, &owner, NULL)
+-----------------------+
- The data fields are at their starting values.
- The "1>" markers are showing where each thread is at for the ObjectMonitor box:
- T-deflate is about to execute cmpxchg().
- T-save enter is about to increment the ref_countcontentions.
Racing Threads
T-save enter ObjectMonitor T-deflate
------------------------ +-----------------------+ --------------------------------------------
save_om_ptrenter() { | owner=DEFLATER_MARKER | deflate_monitor_using_JT() {
1> atomic inc ref_countadd_to_contentions(1) | ref_countcontentions=0 | cmpxchg(try_set_owner_from(NULL, DEFLATER_MARKER, &owner, NULL)
+-----------------------+ :
1> prev = cmpxchg(&contentions, 0, -max_jint, &ref_count, 0)
- T-deflate has executed cmpxchg() and set owner to DEFLATEDEFLATER_MARKER.
- T-save enter still hasn't done anything yet
- The "1>" markers are showing where each thread is at for the ObjectMonitor box:
- T-save enter and T-deflate are racing to update the ref_count contentions field.
T-deflate Wins
T-save enter ObjectMonitor ObjectMonitor T-deflate
---------------------------------- +-------------------------+ --------------------------------------------
save_om_ptrenter() { | owner=DEFLATER_MARKER | deflate_monitor_using_JT() {
atomic inc ref_countadd_to_contentions(1) | ref_countcontentions=-max_jint+1 | cmpxchg(try_set_owner_from(NULL, DEFLATER_MARKER, &owner, NULL)
1> if (owner == DEFLATER_MARKER &&is_being_async_deflated()) { +-------------------------+ :
restore obj header ref_count <= 0) { || prev = cmpxchg(&contentions, 0, -max_jint, &ref_count, 0)
restore obj headeradd_to_contentions(-1) \/ 1> if (prev == 0) &&
atomic dec ref_count {
2> return false to force retry +-------------------------+ owner == DEFLATER_MARKER) {
2> return false to force retry restore obj header
} | owner=DEFLATER_MARKER | 2> finish restorethe obj headerdeflation
} | ref_countcontentions=-max_jint | 2> finish the deflation}
+-------------------------+ }
- This diagram starts after "Racing Threads".
- The "1>" markers are showing where each thread is at for that ObjectMonitor box:
- T-save enter and T-deflate both observe owner == DEFLATER_MARKER and a negative ref_count contentions field.
- T-save enter has lost the race: it restores the obj header (not shown) and decrements the ref_countcontentions.
- T-deflate restores the obj header (not shown).
- The "2>" markers are showing where each thread is at for that ObjectMonitor box.
- T-save enter returns false to cause the caller to retry.
- T-deflate finishes the deflation.
- Note: As of CR5/v2.05/8-for-jdk13, the owner == DEFLATER_MARKER value is allowed to linger until a deflated ObjectMonitor is reused for an enter operation. This narrows the C2 ObjectMonitor enter optimization race window with async deflation.
T-save Wins
T-enter Wins
T-save enter ObjectMonitor T-deflate
---------------------------------- +-------------------------+ ---------------------------------------------
save_om_ptrenter() { | owner=DEFLATER_MARKER. | deflate_monitor_using_JT() {
atomic inc ref_count add_to_contentions(1) | ref_countcontentions=1 | cmpxchg(try_set_owner_from(NULL, DEFLATER_MARKER, &owner, NULL)
1> if (owner == DEFLATER_MARKER &&is_being_async_deflated()) { +-------------------------+ :
} ref_count <= 0) { || prev = cmpxchg(-max_jint, &ref_count, 0)
} else { || prev = cmpxchg(&contentions, 0, -max_jint)
2> <continue contended enter> \/ 1> if (prev == 0) &&
{
save om_ptr in the +-------------------------+ owner == DEFLATER_MARKER)} else {
ObjectMonitorHandle | owner=NULL | } else {
try_set_owner_from(DEFLATER_MARKER, NULL)
2> return true | ref_countcontentions=1 | cmpxchg(NULL, &owner, DEFLATER_MARKER)2> return
+-------------------------+ 2> return
- This diagram starts after "Racing Threads".
- The "1>" markers are showing where each thread is at for the ObjectMonitor box:
- T-save enter and T-deflate both observe a ref_count contentions field > 0.
- T-save enter has won the race and it saves the ObjectMonitor* in the ObjectMonitorHandle (not shown)continues with the contended enter protocol.
- T-deflate detects that it has lost the race (prev != 0) and bails out on deflating the ObjectMonitor:
- Before bailing out T-deflate tries to restore the owner field to NULL if it is still DEFLATER_MARKER.
- The "2>" markers are showing where each thread is at for that ObjectMonitor box.
T-save Complication with C2
Sorry in advance for the sudden deep dive into really gory C2 details, but this is related to a majority of save_om_ptr() so this is the right place to talk about the complication.
As of CR7/v2.07/10-for-jdk14, we have added C2 inc_om_ref_count() on X64 to implement the ref_count management parts of save_om_ptr():
- inc_om_ref_count() does not implement the "restore obj header" part nor the "save om_ptr in the ObjectMonitorHandle" part mentioned in the previous two subsections.
- inc_om_ref_count() is used by C2 fast_lock(), C2 fast_unlock() and C2 rtm_inflated_locking() on the LP64 X64 platform.
- The v2.05 version of C2 fast_lock() has code to detect a deflated and recycled ObjectMonitor after acquiring ownership of the ObjectMonitor. The solution to the race was to drop ownership and take the slow enter path. We have spent a lot of time and energy analyzing this race and the solution to this race and have convinced ourselves that the solution introduces theoretical problems with succession. The proper solution is to switch to using inc_om_ref_count() to protect the ObjectMonitor* for the duration of C2 fast_lock().
- Robbin wrote a new test called MoCrazy that is targeted at the C2 optimizations. This test revealed a race in the baseline C2 fast_unlock() where ownership was reacquired in order to ensure proper succession. So baseline C2 fast_unlock() had a similar version of the race that we thought we fixed in C2 fast_lock(). The proper solution is to switch to using inc_om_ref_count() to protect the ObjectMonitor* for the duration of C2 fast_unlock().
- C2 rtm_inflated_locking() is similarly exposed to races with async deflation so inc_om_ref_count() is used to protect the ObjectMonitor* for the duration of C2 rtm_inflated_locking().
T-enter Wins By A-B-A
T-enter ObjectMonitor T-deflate
-------------------------------------------- +-------------------------+ ------------------------------------------
ObjectMonitor::enter() { | owner=DEFLATER_MARKER | deflate_monitor_using_JT() {
<owner is contended> | ref_count=1 | cmpxchg(DEFLATER_MARKER, &owner, NULL)
1> EnterI() { +-------------------------+ 1> :
if (owner == DEFLATER_MARKER && || 2> : <thread_stalls>
cmpxchg(Self, &owner, \/ :
DEFLATER_MARKER) +-------------------------+ :
== DEFLATER_MARKER) { | owner=Self/T-enter | :
// EnterI is done | ref_count=0 | : <thread_resumes>
return +-------------------------+ prev = cmpxchg(-max_jint, &ref_count, 0)
} || if (prev == 0 &&
} // enter() is done \/ 3> owner == DEFLATER_MARKER) {
~OMH: atomic dec ref_count +-------------------------+ } else {
2> : <does app work> | owner=Self/T-enter|NULL | cmpxchg(NULL, &owner, DEFLATER_MARKER)
3> : | ref_count=-max_jint | atomic add max_jint to ref_count
exit() monitor +-------------------------+ 4> bailout on deflation
4> owner = NULL || }
\/
+-------------------------+
| owner=Self/T-enter|NULL |
| ref_count=0 |
+-------------------------+
- T-deflate has executed cmpxchg() and set owner to DEFLATE_MARKER.
- T-enter has called ObjectMonitor::enter() with "ref_count == 1", noticed that the owner is contended and is about to call ObjectMonitor::EnterI().
- The first ObjectMonitor box is showing the fields at this point and the "1>" markers are showing where each thread is at for that ObjectMonitor box.
- T-deflate stalls after setting the owner field to DEFLATER_MARKER.
- T-enter calls EnterI() to do the contended enter work:
- EnterI() observes owner == DEFLATER_MARKER and uses cmpxchg() to set the owner field to Self/T-enter.
- T-enter owns the monitor, returns from EnterI(), and returns from enter().
- The ObjectMonitorHandle destructor decrements the ref_count.
- T-enter is now ready to do work that requires the monitor to be owned.
- The second ObjectMonitor box is showing the fields at this point and the "2>" markers are showing where each thread is at for that ObjectMonitor box.
- T-enter is doing app work (but it also could have finished and exited the monitor).
- T-deflate resumes, calls cmpxchg() to set the ref_count field to -max_jint, and passes the first part of the bailout expression because "prev == 0".
- The third ObjectMonitor box is showing the fields at this point and the "3>" markers are showing where each thread is at for that ObjectMonitor box.
- T-deflate performs the A-B-A check which observes that "owner != DEFLATE_MARKER" and bails out on deflation:
- Depending on when T-deflate resumes after the stall, it will see "owner == T-enter" or "owner == NULL".
- Both of those values will cause deflation to bailout so we have to conditionally undo work:
- restore the owner field to NULL if it is still DEFLATER_MARKER (it's not DEFLATER_MARKER)
- undo setting ref_count to -max_jint by atomically adding max_jint to ref_count which will restore ref_count to its proper value.
- If the T-enter thread has managed to enter but not exit the monitor during the T-deflate stall, then our owner field A-B-A transition is:
NULL → DEFLATE_MARKER → Self/T-enter
so we really have A1-B-A2, but the A-B-A principal still holds.
If the T-enter thread has managed to enter and exit the monitor during the T-deflate stall, then our owner field A-B-A transition is:
NULL → DEFLATE_MARKER → Self/T-enter → NULL
so we really have A-B1-B2-A, but the A-B-A principal still holds.
T-enter finished doing app work and is about to exit the monitor (or it has already exited the monitor).
The fourth ObjectMonitor box is showing the fields at this point and the "4>" markers are showing where each thread is at for that ObjectMonitor box.
An Example of Object Header Interference
After T-deflate has won the race for deflating an ObjectMonitor it has to restore the header in the associated object. Of course another thread can be trying to do something to the object's header at the same time. Isn't asynchronous work exciting?!?!
ObjectMonitor::install_displaced_markword_in_object() is called from two places so we can have a race between a T-save thread and a T-deflate thread:
Start of the Race
T-save object T-deflate
------------------------------------------- +-------------+ --------------------------------------------
install_displaced_markword_in_object() { | mark=om_ptr | install_displaced_markword_in_object() {
dmw = header() +-------------+ dmw = header()
if (!dmw->is_marked() && if (!dmw->is_marked() &&
dmw->hash() == 0) { dmw->hash() == 0) {
create marked_dmw create marked_dmw
dmw = cmpxchg(marked_dmw, &header, dmw) dmw = cmpxchg(marked_dmw, &header, dmw)
} }
- The data field (mark) is at its starting value.
- 'dmw' and 'marked_dmw' are local copies in each thread.
- T-save and T-deflate are both calling install_displaced_markword_in_object() at the same time.
- Both threads are poised to call cmpxchg() at the same time.
T-deflate Wins First Race
T-save object T-deflate
------------------------------------------- +-------------+ -------------------------------------------
install_displaced_markword_in_object() { | mark=om_ptr | install_displaced_markword_in_object() {
dmw = header() +-------------+ dmw = header()
if (!dmw->is_marked() && if (!dmw->is_marked() &&
dmw->hash() == 0) { dmw->hash() == 0) {
create marked_dmw create marked_dmw
dmw = cmpxchg(marked_dmw, &header, dmw) dmw = cmpxchg(marked_dmw, &header, dmw)
} }
// dmw == marked_dmw here // dmw == original dmw here
if (dmw->is_marked()) if (dmw->is_marked())
unmark dmw unmark dmw
obj = object() obj = object()
obj->cas_set_mark(dmw, this) obj->cas_set_mark(dmw, this)
- The return value from cmpxchg() in each thread will be different.
- Since T-deflate won the race, its 'dmw' variable contains the header/dmw from the ObjectMonitor.
- Since T-save lost the race, its 'dmw' variable contains the 'marked_dmw' set by T-deflate.
- T-save will unmark its 'dmw' variable.
- Both threads are poised to call cas_set_mark() at the same time.
T-save Wins First Race
- Note: The owner == DEFLATER_MARKER and contentions < 0 values that are set by T-deflate (stage one of async deflation) remain in place until after T-deflate does a handshake (or safepoint) operation with all JavaThreads. This handshake forces T-enter to make forward progress and see that the ObjectMonitor is being async deflated before T-enter checks in for the handshake.
T-enter Wins By Cancellation Via DEFLATER_MARKER Swap
T-enter ObjectMonitor T-save object T-deflate
------------------------------------------- - +-------------+ ---------------------+ ----------------------
install_displaced_markword_in_object() { | mark=om_ptr | install_displaced_markword_in_object() {
dmw = header() +-------------+ dmw = header()
if (!dmw->is_marked() && if (!dmw->is_marked() &&
dmw->hash() == 0) { dmw->hash() == 0) {
create marked_dmw create marked_dmw
dmw = cmpxchg(marked_dmw, &header, dmw) dmw = cmpxchg(marked_dmw, &header, dmw)
} }
// dmw == original dmw here // dmw == marked_dmw here
if (dmw->is_marked()) if (dmw->is_marked())
unmark dmw unmark dmw
obj = object() obj = object()
obj->cas_set_mark(dmw, this) obj->cas_set_mark(dmw, this)
- This diagram is the same as "T-deflate Wins First Race" except we've swapped the post cmpxchg() comments.
- Since T-save won the race, its 'dmw' variable contains the header/dmw from the ObjectMonitor.
- Since T-deflate lost the race, its 'dmw' variable contains the 'marked_dmw' set by T-save.
- T-deflate will unmark its 'dmw' variable.
- Both threads are poised to call cas_set_mark() at the same time.
Either Wins the Second Race
T-save object T-deflate
----------
ObjectMonitor::enter() { | owner=DEFLATER_MARKER | deflate_monitor_using_JT() {
add_to_contentions(1) | contentions=1 | try_set_owner_from(NULL, DEFLATER_MARKER)
1> EnterI() { +------------------------------------------- +-------------+ -------------------------------------------
install_displaced_markword_in_object() { | mark=dmw | install_displaced_markword_in_object() {
dmw = header() + 1> :
if (try_set_owner_from(DEFLATER_MARKER, || 2> : <thread_stalls>
Self) == DEFLATER_MARKER) { \/ :
// Add marker for cancellation +-------------------------+ dmw:
= header()
if (!dmw->is_marked() && add_to_contentions(1) | owner=Self/T-enter | :
if (!dmw->is_marked() &&
dmw->hash() == 0) { // EnterI dmw->hash() == 0) {
create marked_dmw create marked_dmw
dmw = cmpxchg(marked_dmw, &header, dmw) dmw = cmpxchg(marked_dmw, &header, dmw)
} }
// dmw == ... is done | contentions=2 | : <thread_resumes>
return // dmw == ...
if (dmw->is_marked()) if (dmw->is_marked())
unmark dmw unmark dmw
obj = object() obj = object()
obj->cas_set_mark(dmw, this) obj->cas_set_mark(dmw, this)
- It does not matter whether T-save or T-deflate won the cmpxchg() call so the comment does not say who won.
- It does not matter whether T-save or T-deflate won the cas_set_mark() call; in this scenario both were trying to restore the same value.
- The object's mark field has changed from 'om_ptr' → 'dmw'.
Please notice that install_displaced_markword_in_object() does not do any retries on any code path:
- Instead the code adapts to being the loser in a cmpxchg() by unmarking its copy of the dmw.
- In the second race, if a thread loses the cas_set_mark() race, there is also no need to retry because the object's header has been restored by the other thread.
Hashcodes and Object Header Interference
If we have a race between a T-deflate thread and a thread trying to get/set a hashcode (T-hash), then the race is between the ObjectMonitorHandle.save_om_ptr(obj, mark) call in T-hash and deflation protocol in T-deflate.
Start of the Race
T-hash ObjectMonitor T-deflate
---------------------- +-----------------------+ ----------------------------------------
save_om_ptr() { | owner=NULL | deflate_monitor_using_JT() {
: | ref_count=0 | 1> cmpxchg(DEFLATER_MARKER, &owner, NULL)
1> atomic inc ref_count +-----------------------+
- The data fields are at their starting values.
- T-deflate is about to execute cmpxchg().
- T-hash is about to increment ref_count.
- The "1>" markers are showing where each thread is at for the ObjectMonitor box.
Racing Threads
T-hash ObjectMonitor T-deflate
---------------------- +-------------------------+ prev = cmpxchg(&contentions, 0, -max_jint)
} || if (prev == 0) {
2> add_to_contentions(-1) \/ 3> } else {
} // enter() is done +-----------------------+ ------------------------------------------
save_om_ptr() { | owner=DEFLATER_MARKER | deflate_monitor_using_JT() {
: --+ if (try_set_owner_from(DEFLATER_MARKER,
: <does app work> | ref_count=0 | cmpxchg(DEFLATER_MARKER, &owner, NULL)
1> atomic inc ref_count +-----------------------+| owner=Self/T-enter|NULL | if (contentions != 0 || waitersNULL) != 0DEFLATER_MARKER) {
3> : }
| contentions=1 | add_to_contentions(-1)
exit() monitor 1> prev = cmpxchg(-max_jint, &ref_count, 0)
- T-deflate has set the owner field to DEFLATER_MARKER.
- The "1>" markers are showing where each thread is at for the ObjectMonitor box:
- T-deflate is about to execute cmpxchg().
- T-save is about to increment the ref_count.
T-deflate Wins
If T-deflate wins the race, then T-hash will have to retry at most once.
T-hash ObjectMonitor T-deflate
------------------------- +-----------------------+ +------------------------------------------
save_om_ptr() { | owner=DEFLATER_MARKER | deflate_monitor_using_JT() {
1> atomic inc ref_count | ref_count=-max_jint | cmpxchg(DEFLATER_MARKER, &owner, NULL)
if (owner == +-----------------------+ if (contentions != 0 || waiters != 0) {
DEFLATER_MARKER && + }
4> owner = NULL || 4> bailout on deflation
\/ }
|| }
ref_count <= 0) { \/ prev = cmpxchg(-max_jint, &ref_count, 0)
restore obj header +-------------------------+ 1> if (prev == 0 &&
atomic dec ref_count | owner=DEFLATER_MARKER | Self/T-enter|NULL |
owner == DEFLATER_MARKER) {
2> return false to | ref_count=-max_jint | restore obj header
cause a retry +-----------------------+ 2> finish| thecontentions=0 deflation
} |
}+-------------------------+
- T-deflate made it past the cmpxchg() of ref_count before T-hash incremented ithas set owner to DEFLATER_MARKER.
- T-deflate set the ref_count field to -max_jint and is about to make the last of the protocol checksenter has called ObjectMonitor::enter(), noticed that the owner is contended, increments contentions, and is about to call ObjectMonitor::EnterI().
- The first ObjectMonitor box is showing the fields at this point and the "1>" markers are showing where each thread is at for that ObjectMonitor box.
- T-deflate sees "prev == 0 && owner == stalls after setting the owner field to DEFLATER_MARKER" so it knows that it has won the race.
- T-deflate restores obj header (not shown).
- T-hash increments the ref_count.
- enter calls EnterI() to do the contended enter work:
- EnterI() sets the owner field from DEFLATER_MARKER to Self/T-enter.
- EnterI() increments contentions one extra time since it cancelled async deflation via a DEFLATER_MARKER swap.
- Note: The extra increment also makes the return value from is_being_async_deflated() stable; the previous A-B-A algorithm would allow the contentions field to flicker from 0 → -max_jint and back to zero. With the current algorithm, a negative contentions field value is a linearization point so once it is negative, we are committed to performing async deflation.
- T-enter owns the monitor and returns from EnterI() (contentions still has both increments)
- .
- The second ObjectMonitor box is showing the fields at this point and the "2>" markers are showing where each thread is at for that ObjectMonitor box.
- T-deflate finishes the deflation workenter decrements contentions and returns from enter() (contentions still has the extra increment).
- T-hash returns false to cause a retry and when T-hash retries:
- it observes the restored object header (done by T-hash or T-deflate):
- if the object's header does not have a hash, then generate a hash and merge it with the object's header.
- Otherwise, extract the hash from the object's header and return it.
- it observes the restored object header (done by T-hash or T-deflate):
T-hash Wins
If T-hash wins the race, then the ref_count will cause T-deflate to bail out on deflating the monitor.
Note: header is not mentioned in any of the previous sections for simplicity.
- enter is now ready to do work that requires the monitor to be owned.
- T-enter is doing app work (but it also could have finished and exited the monitor and it still has the extra increment).
- T-deflate resumes, tries to set the contentions field to -max_jint and fails because contentions == 1 (the extra increment comes into play!).
- The third ObjectMonitor box is showing the fields at this point and the "3>" markers are showing where each thread is at for that ObjectMonitor box.
- T-deflate tries to restore the owner field from DEFLATER_MARKER to NULL:
- If it does not succeed, then the EnterI() call managed to cancel async deflation via a DEFLATER_MARKER swap so T-deflate decrements contentions to get rid of the extra increment that EnterI() did as a marker for this type of cancellation.
- If it does succeed, then EnterI() did not cancel async deflation via a DEFLATER_MARKER swap and we don't have an extra increment to get rid of.
- Note: For the previous bullet, async deflation is still cancelled because the ObjectMonitor is now busy with a contended enter.
T-enter finished doing app work and is about to exit the monitor (or it has already exited the monitor).
The fourth ObjectMonitor box is showing the fields at this point and the "4>" markers are showing where each thread is at for that ObjectMonitor box.
An Example of Object Header Interference
After T-deflate has won the race for deflating an ObjectMonitor it has to restore the header in the associated object. Of course another thread can be trying to do something to the object's header at the same time. Isn't asynchronous work exciting?!?!
ObjectMonitor::install_displaced_markword_in_object() is called from two places so we can have a race between a T-enter thread and a T-deflate thread:
Start of the Race
T-enter object T-deflate
----------------------------------------------- +-------------+ ------------------------ T-hash ObjectMonitor T-deflate
------------------------- +-----------------------+ ------------------------------------------
save_om_ptr() { | header=dmw_no_hash | deflate_monitor_using_JT() {
atomic inc ref_count | owner=DEFLATER_MARKER | cmpxchg(DEFLATER_MARKER, &owner, NULL)
1> if (owner == | ref_count=1 | if (contentions != 0 || waiters != 0) {
DEFLATER_MARKER && +-----------------------+ }
ref_count <= 0) { || 1> prev = cmpxchg(-max_jint, &ref_count, 0)
} else { \/ if (prev == 0 &&
2> save om_ptr in the +-----------------------+
install_displaced_markword_in_object(oop obj) { owner| == DEFLATER_MARKERmark=om_ptr | install_displaced_markword_in_object(oop obj) {
dmw ObjectMonitorHandle= header() | header=dmw_no_hash | } else {
return true +-------------+ dmw = header()
obj->cas_set_mark(dmw, this) | owner=NULL obj->cas_set_mark(dmw, this) | cmpxchg(NULL, &owner, DEFLATER_MARKER)
} | ref_count=1 | 2> bailout on deflation
} }
- The data field (mark) is at its starting value.
- 'dmw' is a local copy in each thread.
- T-enter and T-deflate are both calling install_displaced_markword_in_object() at the same time.
- Both threads are poised to call cas_set_mark() at the same time.
Either Thread Wins the Race
T-enter object +T-deflate
----------------------------------------------- +-------------+ --------------+ }
if save_om_ptr() { ||
if no hash \/
gen hash & merge +---------------------------------+
hash = hash(header) | header=dmw_hash |
} install_displaced_markword_in_object(oop obj) { | mark=dmw | install_displaced_markword_in_object(oop obj) {
dmw = header() +-------------+ | owner=NULL |
3> atomic dec ref_count | ref_count=1 |
return hash dmw = header()
obj->cas_set_mark(dmw, this) +-----------------------+
- T-deflate has set the owner field to DEFLATER_MARKER.
- T-hash has incremented ref_count before T-deflate made it to cmpxchg().
- The first ObjectMonitor box is showing the fields at this point and the "1>" markers are showing where each thread is at for that ObjectMonitor box.
- T-deflate bails out on deflation, but first it tries to restore the owner field:
- The return value of cmpxchg() is not checked here.
- If T-deflate cannot restore the owner field to NULL, then another thread has managed to enter the monitor (or enter and exit the monitor) and we don't want to overwrite that information.
- T-hash observes:
- "owner == DEFLATER_MARKER && ref_count > 0" or
- "owner == NULL && ref_count > 0" so it gets ready to save the ObjectMonitor*.
- The second ObjectMonitor box is showing the fields at this point and the "2>" markers are showing where each thread is at for that ObjectMonitor box.
- T-hash saves the ObjectMonitor* in the ObjectMonitorHandle (not shown) and returns to the caller.
- save_om_ptr() returns true since the ObjectMonitor is safe:
- if ObjectMonitor's 'header/dmw' field does not have a hash, then generate a hash and merge it with the 'header/dmw' field.
- Otherwise, extract the hash from the ObjectMonitor's 'header/dmw' field.
- The third ObjectMonitor box is showing the fields at this point and the "3>" marker is showing where T-hash is at for that ObjectMonitor box.
- T-hash decrements the ref_count field.
- T-hash returns the hash value.
Please note that in Carsten's original prototype, there was another race in ObjectSynchronizer::FastHashCode() when the object's monitor had to be inflated. The setting of the hashcode in the ObjectMonitor's header/dmw could race with T-deflate. That race is resolved in this version by the use of an ObjectMonitorHandle in the call to ObjectSynchronizer::inflate(). The ObjectMonitor* returned by ObjectMonitorHandle.om_ptr() has a non-zero ref_count so no additional races with T-deflate are possible.
...
obj->cas_set_mark(dmw, this)- It does not matter whether T-enter or T-deflate won the cas_set_mark() call; in this scenario both were trying to restore the same value.
- The object's mark field has changed from 'om_ptr' → 'dmw'.
Please notice that install_displaced_markword_in_object() does not do any retries on any code path:
- If a thread loses the cas_set_mark() race, there is no need to retry because the object's header has been restored by the other thread.
Hashcodes and Object Header Interference
There are a few races that can occur between a T-deflate thread and a thread trying to get/set a hashcode (T-hash) in an ObjectMonitor:
- If the object has an ObjectMonitor (i.e., is inflated) and if the ObjectMonitor has a hashcode, then the hashcode value can be carefully fetched from the ObjectMonitor and returned to the caller (T-hash). If there is a race with async deflation, then we have to retry.
- There are several reasons why we might have to inflate the ObjectMonitor in order to set the hashcode:
- The object is neutral, does not contain a hashcode and we (T-hash) lost the race to try an install a hashcode in the mark word.
- The object is stack locked and does not contain a hashcode in the mark word.
- The object has an ObjectMonitor and the ObjectMonitor does not have a hashcode.
Note: In this case, the inflate() call on the common fall thru code path is almost always a no-op since the existing ObjectMonitor is not likely to be async deflated before inflate() sees that the object already has an ObjectMonitor and bails out.
The common fall thru code path (executed by T-hash) that inflates the ObjectMonitor in order to set the hashcode can race with an async deflation (T-deflate). After the hashcode has been stored in the ObjectMonitor, we (T-hash) check if the ObjectMonitor has been async deflated (by T-deflate). If it has, then we (T-hash) retry because we don't know if the hashcode was stored in the ObjectMonitor before the object's header was restored (by T-deflate). Retrying (by T-hash) will result in the hashcode being stored in either object's header or in the re-inflated ObjectMonitor's header as appropriate.
Spin-Lock Monitor List Management In Theory
Use of specialized measurement code with the CR5/v2.05/8-for-jdk13 bits revealed that the gListLock contention is responsible for much of the performance degradation observed with SPECjbb2015. Consequently the primary focus of the next round of changes is/was on switching from course grained Thread::muxAcquire(&gListLock) and Thread::muxRelease(&gListLock) pairs to spin-lock -free monitor list management. Of course, since the Java Monitor subsystem is full of special cases, the spin-lock -free list management code has to have a number of special cases which are described hereto have a number of special cases which are described here.
The Spin-Lock Monitor List management code was pushed to JDK15 using the following bug id:
JDK-8235795 replace monitor list mux{Acquire,Release}(&gListLock) with spin locks
The Async Monitor Deflation project makes a few additional changes on top of what was pushed via JDK-8235795.
The Simple Case
There is one simple case of spin-lock -free list management with the Java Monitor subsystem so we'll start with that code as a way to introduce the spin-lock -free concepts:
L1: while (true) {
L2: PaddedObjectMonitor* cur = Atomic::load(&g_block_list);
L3: Atomic::store(&new_blk[0]._next_om, = cur);
L4: if (Atomic::cmpxchg(new_blk, &g_block_list, cur, new_blk) == cur) {
L5: Atomic::add(&om_list_globals.population, _BLOCKSIZE - 1, &g_om_population);
L6: break;
L7: }
L8: }
What the above block of code does is:
- prepends a 'new_blk' to the front of 'g_block_list'
- increments the 'gom_omlist_globals.population' counter to include the number of new elements
...
- L2 loads the current 'g_block_list' value into 'cur'; g_block_list is only modified by cmpxchg so a load (instead of a load-acquire) is sufficient.
- L3 stores 'cur' into the 0th element's next field for 'new_blk'; g_block_list is only modified by cmpxchg so a store (instead of a release-store) is sufficientnext field for 'new_blk'.
- L4 is the critical decision point for this lock-free list update. cmpxchg will change 'g_block_list' to 'new_blk' iff 'g_block_list' == 'cur' (publish it).
- if the cmpxchg return value is 'cur', then we succeeded with the lock-free list update and we atomically update 'gom_omlist_globals.population' to match.
- Otherwise we loop around and do everything again from L2. This is the "spin" part of spin-lock.
At the point that cmpxchg has published the new 'g_block_list' value, 'new_blk' is now first block in the list and the 0th element's next field is used to find the previous first block; all of the monitor list blocks are chained together via the next field in the block's 0th element. It is the use of cmpxchg to update 'g_block_list' and the checking of the return value from cmpxchg that insures that we don't lose track of any blocks.
...
to achieve the safe update of the 'g_block_list' value; the atomic increment of the 'gom_omlist_globals.population' counter is considered to be just accounting (pun intended).
...
Note: This subsection is talking about "Simple Take" and "Simple Prepend" in abstract terms. The purpose of this code and A-B-A example is to introduce the race concepts. The code shown here is not an exact match for the project code and the specific A-B-A example is not (currently) found in the project code.
The left hand column shows "Thread1T1" taking a node "A" from the front of the list and it shows the simple code that does that operation. The right hand column shows "Thread2T2" prepending a node "B" to the front of the list and it shows the simple code that does that operation. We have a third thread, "Thread3T3", that does a take followed by a prepend, but we don't show a column for "Thread3T3". Instead we have a column in the middle that shows the results of the interleaved operations of all three threads:
...
When we allow simultaneous take and prepend operations on the same list, the simple algorithms are exposed to A-B-A races. An A-B-A race is a situation where the head of the list can change from node "A" to node "B" and back to node "A" again without the simple algorithm being aware that critical state has changed. In the middle column of the above diagram, we show what happens when Thread3 T3 causes the head of the list to change from node "A" to node "B" (a take operation) and back to node "A" (a prepend operation). That A-B-A race causes Thread1 T1 to lose node "B" when it updates the list head to node "X" instead of node "B" because Thread1 T1 was unaware that its local 'next' value was stale.
Here's the diagram again with the code in Thread1 T1 and Thread2 T2 lined up with the effects of the A-B-A race executed by Thread3T3:
T1: Simple Take: | | T2: Simple Prepend:
---------------- | T1 and T3 see this initial list: | -------------------
while (true) { | +---+ +---+ +---+ | :
cur = head; | head -> | A | -> | X | -> | Y | | :
next = cur->next; | +---+ +---+ +---+ | :
: | T3 takes "A", T2 sees this list: | :
: | +---+ +---+ | :
: | head -> | X | -> | Y | | :
: | +---+ +---+ | while (true) {
: | T2 prepends "B": | cur = head;
: | +---+ +---+ +---+ | new->next = cur;
: | head -> | B | -> | X | -> | Y | | if (cmpxchg(new, &head, cur) == cur) {
: | +---+ +---+ +---+ | break;
: | T3 prepends "A": | }
: | +---+ +---+ +---+ +---+ | }
: | head -> | A | -> | B | -> | X | -> | Y | |
: | +---+ +---+ +---+ +---+ |
: | T1 takes "A", loses "B": |
: | +---+ |
: | | B | ----+ |
: | +---+ | |
: | V |
: | +---+ +---+ |
if (cmpxchg(next, &head, cur) == cur) { | head -> | X | -> | Y | |
} | +---+ +---+ |
} | +---+ |
return cur; | cur -> | A | |
| +---+ |
So the simple algorithms are not sufficient when we allow simultaneous take and prepend operations.
...
Spin-Locking to Solve the A-B-A Race
Note: This subsection is talking about "MarkingSpin-Locking" as a solution to the A-B-A race in abstract terms. The purpose of this marking spin-locking code and A-B-A example is to introduce the solution concepts. The code shown here is not an exact match for the project code.
One solution to the A-B-A race is to mark spin-lock the next field in a node to indicate that the node is busy. Only one thread can successfully mark spin-lock the next field in a node at a time and other threads must loop around and retry their marking spin-locking operation until they succeed. Each thread that marks spin-locks the next field in a node must unmark unlock the next field when it is done with the node so that other threads can proceed.
Here's the take algorithm modified with marking spin-locking (still ignores the empty list for clarity):
// "take" a node with markinglocking:
while (true) {
cur = head;
if (!marktry_om_nextlock(cur, &next)) {
// could not marklock cur so try again
continue;
}
if (head != cur) {
// head changed while markinglocking cur so try again
unmarkom_nextunlock(cur);
continue;
}
next = unmarked_next(cur);
// list head is now markedlocked so switch it to next which also makes list head unmarkedunlocked
OrderAccessAtomic::release_store(&head, next);
unmarkom_nextunlock(cur); // unmarkunlock cur and return it
return cur;
}
The modified take algorithm does not change the list head pointer until it has successfully marked locked the list head node. Notice that after we mark lock the list head node we have to verify that the list head pointer hasn't changed in the mean time. Only after we have verified that the node we marked locked is still the list head is it safe to modify the list head pointer. The marking locking of the list head prevents the take algorithm from executing in parallel with a prepend algorithm and losing a node.
Also notice that we update the list head pointer with release-store instead of with cmpxchg. Since we have the list head markedlocked, we are not racing with other threads to change the list head pointer so we can use the smaller release-store hammer a simple store instead of the heavier heavy cmpxchg hammer.
Here's the prepend algorithm modified with marking locking (ignores the empty list for clarity):
// "prepend" a node with markinglocking:
while (true) {
cur = head;
if (!marktry_om_nextlock(cur, &next)) {
// could not marklock cur so try again
continue;
}
if (head != cur) {
// head changed while markinglocking cur so try again
unmarkom_nextunlock(cur);
continue;
}
next = unmarked_next(cur);
// list head is now markedlocked so switch it to 'new' which also makes list head unmarkedunlocked
Atomic::release_store(&head, new);
unmarkom_nextunlock(cur); // unmarkunlock the previous list head
}
The modified prepend algorithm does not change the list head pointer until it has successfully marked locked the list head node. Notice that after we mark lock the list head node we have to verify that the list head pointer hasn't changed in the mean time. Only after we have verified that the node we marked locked is still the list head is it safe to modify the list head pointer. The marking locking of the list head prevents the prepend algorithm from executing in parallel with the take algorithm and losing a node.
Also notice that we update the list head pointer with release-store instead of with cmpxchg for the same reasons as the previous algorithm.
...
The purpose of this subsection is to provide background information about how ObjectMonitors move between the various lists. This project changes the way these movements are implemented, but does not change the movements themselves. For example, newly allocated blocks of ObjectMonitors are always prepending to the global free list; this is true in the baseline and is true in this project. One exception is the optional addition of the global wait list (see below).
...
ObjectMonitors are deflated at a safepoint by:
ObjectSynchronizer::deflate_monitor_list() calling ObjectSynchronizer::deflate_monitor()
And when Async Monitor Deflation is enabled, they are deflated by:
ObjectSynchronizer::deflate_monitor_list_using_JT() calling ObjectSynchronizer::deflate_monitor_using_JT()Idle ObjectMonitors are deflated by the ServiceThread when Async Monitor Deflation is enabled. They can also be deflated at a safepoint by the VMThread or by a task worker thread. Safepoint deflation is used when Async Monitor Deflation is disabled or when there is a special deflation request, e.g., System.gc().
An idle ObjectMonitor is deflated and extracted from its in-use list and prepended to the global free wait list. The in-use list can be either the global in-use list or a per-thread in-use list. Deflated ObjectMonitors are always prepended to the global free wait list.
- In CR7/v2.07/10-for-jdk14, the HandshakeAfterDeflateIdleMonitors diagnostic option is added to enable a new g_wait_list that tracks deflated ObjectMonitors until after a handshake/safepoint with all JavaThreads.
- The g_The om_list_globals.wait_list allows ObjectMonitors to be safely deflated on platforms that do not have C2 inc_om_ref_count() implemented. See the "T-save Complication with C2" subsection above for the gory C2 details.deflated without reuse races.
- After a So when the option is enabled, idle ObjectMonitors are deflated and extracted from an in-use list and prepended to g_wait_list; after the handshake/safepoint with all JavaThreads, the ObjectMonitors on the gom_list_globals.wait_list are prepended to the global free list.
...
- global free list:
- prepended to by JavaThreads that allocated a new block of ObjectMonitors (malloc time)
- prepended to by JavaThreads that are exiting (and have a non-empty per-thread free list)
- taken from the head by JavaThreads that need to allocate ObjectMonitor(s) for their per-thread free list (reprovision)
- prepended to by deflation done by:
- either the VMThread or a worker thread for safepoint based
- or the ServiceThread for async monitor deflation
- global in-use list:
- prepended to by JavaThreads that are exiting (and have a non-empty per-thread free list)
- extracted from by deflation done by:
- either the VMThread or a worker thread for safepoint based
- or the ServiceThread for async monitor deflation
- global wait list:only used when HandshakeAfterDeflateIdleMonitors == true
- prepended by the ServiceThread during async deflation
- entire list detached and prepended to the global free list by the ServiceThread during async deflation
- Note: The global wait list serves the same function as Carsten's gFreeListNextSafepoint list in his prototype.
- per-thread free list:
- prepended to by a JavaThread when it needs to allocate new ObjectMonitor(s) (reprovision)
- taken from the head by a JavaThread when it needs to allocate a new ObjectMonitor (inflation)
- prepended to by a JavaThread when it isn't able to link the object to the ObjectMonitor (failed inflation)
- entire list detached and prepended to the global free list when the JavaThread is exiting
- per-thread in-use list:
- prepended to by a JavaThread when it allocates a new ObjectMonitor (inflation, optimistically in-use)
- extracted from by deflation done by:
- either the VMThread or a worker thread for safepoint based
- or the ServiceThread for async monitor deflation
- entire list detached and prepended to the global in-use list when the JavaThread is exiting
- global free list:
Spin-Lock
...
Monitor List Management In Reality
Prepending To A List That Also Allows Deletes
...
The next case to consider for spin-lock -free list management with the Java Monitor subsystem is prepending to a list that also allows deletes. As you might imagine, the possibility of a prepend racing with a delete makes things more complicated. The solution is to "mark" lock the next field in the ObjectMonitor at the head of the list we're trying to prepend to. A successful mark lock tells other prependers or deleters that the marked locked ObjectMonitor is busy and they will need to retry their own mark lock operation.
L01: while (true) {
L02: (void)mark_next_loop om_lock(m); // markLock m so we can safely update its next field.
L03: ObjectMonitor* cur = NULL;
L04: ObjectMonitor* next = NULL;
L05: // MarkLock the list head to guard against A-B-A race:
L06L05: if (mark(cur = get_list_head_locked(list_p, &cur, &next))) != NULL) {
L07L06: // List head is now markedlocked so we can safely switch it.
L08L07: setm->set_next_om(m, cur); // m now points to cur (and unmarksunlocks m)
L09L08: OrderAccessAtomic::release_store(list_p, m); // Switch list head to unmarkedunlocked m.
L10L09: setom_nextunlock(cur, next); // Unmark the previous list head.
L11L10: break;
L12L11: }
L13L12: // The list is empty so try to set the list head.
L14L13: assert(cur == NULL, "cur must be NULL: cur=" INTPTR_FORMAT, p2i(cur));
L15L14: setm->set_next_om(m, cur); // m now points to NULL (and unmarksunlocks m)
L16L15: if (Atomic::cmpxchg(m, list_p, cur, m) == cur) {
L17L16: // List head is now unmarkedunlocked m.
L18L17: break;
L19L18: }
L20L19: // Implied else: try it all again
L21L20: }
L22L21: Atomic::inc(count_p);
What the above block of code does is:
- prepends an ObjectMonitor 'm' to the front of the list referred to by list_p
- mark lock 'm's next fieldmark
- lock the list head's next field
- update 'm' to refer to the list head
- update 'list_p' to refer to 'm'
- unmark the next field in unlock the previous list head
- increments the counter referred to by 'count_p' by one
- prepends an ObjectMonitor 'm' to the front of the list referred to by list_p
The above block of code can be called by multiple prependers in parallel or with deleters running in parallel and must not lose track of any ObjectMonitor. Of course, the "must not lose track of any ObjectMonitor" part is where all the details come in:
- L02 loops to mark locks 'm's next field; internally we have to loop because another thread (T2) might have 'm' marked locked and we try again until we have marked locked it.
You might be asking yourself: why does T2 have 'm' markedlocked?- Before T1 was trying to prepend 'm' to an in-use list, T1 and T2 were racing to take an ObjectMonitor off the free list.
- T1 won the race, marked locked 'm', removed 'm' from the free list and unmarked unlocked 'm'; T2 stalled before trying to mark lock 'm'.
- T2 resumed and marked locked 'm', realized that 'm' was no longer the head of the free list, unmarked unlocked 'm' and tried it all again.
- If our thread (T1) does not mark lock 'm' before it tries to prepend it to an in-use list, then T2's umarking unlocking of 'm' could erase the next value that T1 wants to put in 'm'.
- L06 L05 tries to mark lock the list head 'list_p'; if markget_list_head_locked() returns truenon-NULL, we have the list head marked locked and can safely update it:
- L08L07: Update 'm's next field to point to the current list head (which unmarks unlocks 'm').
- L09L08: release-store 'm' into 'list_p' which switches the list head to an unmarked unlocked 'm'.
- L10L09: We unmark unlock the previous list head.
- If markget_list_head_locked() returned falseNULL, we have an empty list:
- L15L14: Update 'm's next field to NULL (which unmarks unlocks 'm').
- L16L15: Try to cmpxchg 'list_p' to 'm':
- if cmpxchg works, then we're done.
- Otherwise, another prepender won the race to update the list head so we have to try again.
- The counter referred to by 'count_p' is incremented by one.
- L02 loops to mark locks 'm's next field; internally we have to loop because another thread (T2) might have 'm' marked locked and we try again until we have marked locked it.
ObjectMonitor 'm' is safely on the list at the point that we have updated 'list_p' to refer to 'm'. In this subsection's block of code, we also called three new functions, mark_next_loop: om_lock(), markget_list_head_locked() and set_next_om(), that are explained in the next few subsections about helper functions.
Note: The above code snippet comes from prepend_to_common(); see that function for more context and a few more comments.
...
try_om_
...
lock(), mark_om_ptr(), and set_next_om() Helper Functions
Managing marks spin-locks on ObjectMonitors has been abstracted into a few helper functions. mark_next. try_om_lock() is the first interesting one:
L01L1: static bool marktry_om_nextlock(ObjectMonitor* om, ObjectMonitor** next_p) {
L02L2: // Get current next field without any markingOM_LOCK_BIT value.
L03L3: ObjectMonitor* next = unmarked_next(ObjectMonitor*om);
L04: ((intptr_t)OrderAccess::load_acquire(&om->_next_om) & ~0x1);
L05: if (Atomic::cmpxchg(L4: if (om->try_set_next_om(next, mark_om_ptr(next), &om->_next_om, next) != next) {
L06L5: return false; // CouldCannot notlock mark the next field or it was already markedObjectMonitor.
L07L6: }
L08: *next_p = next;
L09 L7: return true;
L10L8: }
The above function tries to mark lock the next field in an ObjectMonitor:
- If marking locking is successful, then the unmarked next field is returned via parameter and true is returned.
- Otherwise, false is returned.
The function can be called by multiple threads at the same time and only one thread will succeed in the marking locking operation (return == true) and all other threads will get return == false. Of course, the "only one thread will succeed" part is where all the details come in:
- L0[34] load-acquires L3 loads the ObjectMonitor's next field and strips the marking locking bit:
- The unmarked unlocked value is saved in 'next'.
- We need the unmarked unlocked next value in order to properly detect if the next field was already markedlocked.
- L05 L4 tries to cmpxchg a marked locked 'next' value into the ObjectMonitor's next field:
- if cmpxchg does not work, then we return false:
- The cmpxchg will not work if the next field changes after we load-acquired loaded the value on L04L3.
- The cmpxchg will not work if the next field is already markedlocked.
- Otherwise, we return the unmarked 'next' value via 'next_p' and return true.
- if cmpxchg does not work, then we return false:
- L0[34] load-acquires L3 loads the ObjectMonitor's next field and strips the marking locking bit:
The marktry_om_nextlock() function calls another helper function, mark_om_ptr(), that needs a quick explanation:
L1: static ObjectMonitor* mark_om_ptr(ObjectMonitor* om) {
L2: return (ObjectMonitor*)((intptr_t)om | 0x1OM_LOCK_BIT);
L3: }This function encapsulates the setting of the marking locking bit in an ObjectMonitor* for the purpose of hiding the details and making the calling code easier to read:
- L2 casts the ObjectMonitor* into a type that will allow the '|' operator to be used.
- We use the 0x1 (OM_LOCK_BIT) bit as our marking locking value because ObjectMonitors are aligned on a cache line so the low order bit is not used by the normal addressing of an ObjectMonitor*.
set_next_om() is the next interesting function and it also only needs a quick explanation:
L1: staticinline void ObjectMonitor::set_next(ObjectMonitor* om, _om(ObjectMonitor* value) {
L2: OrderAccessAtomic::release_store(&om->_next_om, value);
L3: }
...
- This function is simply a wrapper around a release-store of an ObjectMonitor* into the next field in an ObjectMonitor.
- The typical "setcur->set_next_om(cur, next)" call sequence is easier to read than "OrderAccess::release_store(&cur→_next_om, next)".
...
om_
...
lock() Helper Function
markom_next_looplock() is the next interesting helper function:
L1: static ObjectMonitor*void markom_next_looplock(ObjectMonitor* om) {
L2: ObjectMonitor* next;
L3: while (true) {
L4: if (mark_next(om, &next))(true) {
L5L3: // Marked om's next field so return the unmarked value.
L6if (try_om_lock(om)) {
L4: return next;
L7L5: }
L8L6: }
L9L7: }
The above function loops until it marks locks the next field of the target ObjectMonitor. The unmarked value of the next field is returned by the function. There is nothing particularly special about this function so we don't need any line specific annotations.
Debugging Tip: If there's a bug where an ObjectMonitor's next field is not properly unmarkedunlocked, then this function will loop forever and the caller will be stuck.
...
get_list_head_locked() Helper Function
markget_list_head_locked() is the next interesting helper function:
L01: static boolObjectMonitor* markget_list_head_locked(ObjectMonitor* volatile * list_p,
L02: ObjectMonitor** mid_p, ObjectMonitor** next_p) {
L03L02: while (true) {
L04L03: ObjectMonitor* mid = OrderAccessAtomic::load_acquire(list_p);
L05L04: if (mid == NULL) {
L06L05: return falseNULL; // The list is empty so nothing to mark.
L07L06: }
L08L07: if (marktry_om_nextlock(mid, next_p)) {
L09L08: if (OrderAccessAtomic::load_acquire(list_p) != mid) {
L10L09: // The list head changed so we have to retry.
L11L10: set_next(mid, *next_p); // unmark mid om_unlock(mid);
L12L11: continue;
L12: }
L13: return mid;
L14: }
L15: }
L14L16: // We marked next field to guard against races.
L15: *mid_p = mid;
L16: return true;
L17: }
L18: }
L19: }
The above function tries to mark the next field in the list head's ObjectMonitor:
- If the list is empty, false is returned.
- Otherwise, the list head's ObjectMonitor* is returned via parameter (mid_p), the unmarked next field is returned via parameter (next_p) and true is returned.
The function can be called by more than one thread on the same 'list_p' at a time. False is only returned when 'list_p' refers to an empty list. Otherwise only one thread will return true at a time with the 'mid_p' and 'next_p' return parameters set. Since the next field in 'mid_p' is marked, any parallel callers to mark_list_head() will loop until the next field in the list head's ObjectMonitor is no longer marked. That typically happens when the list head's ObjectMonitor is taken off the list and 'list_p' is advanced to the next ObjectMonitor on the list. Of course, making sure that "only one thread will return true at a time" is where all the details come in:
- L04 load-acquires the current 'list_p' value into 'mid'; the use of load-acquire is necessary to get the latest value release-stored by another thread; the current 'list_p' is updated by either a release-store or a cmpxchg depending on the algorithm that made the update; only the release-store needs to match up with a load-acquire, but this code doesn't know whether release-store or cmpxchg was used.
- L0[56] is the empty list check and the only time that false is returned by this function.
- L08 tries to mark the next field in 'mid':
- If marking is not successful, we loop around to try it all again.
- If marking is successful, then 'next_p' contains mid's unmarked next field value.
- L09 load-acquires the current 'list_p' value to see if it still matches 'mid':
- If the list head has changed, then we unmark mid on L11 and try it all again.
- Otherwise, 'mid' is returned via 'mid_p' and we return true.
When this function returns true, the next field in 'mid_p' is marked and any parallel callers of mark_list_head() on the same list will be looping until the next field in the list head's ObjectMonitor is no longer marked. The caller that just got the 'true' return needs to finish up its work with 'mid_p' quickly.
Debugging Tip: If there's a bug where the list head ObjectMonitor's next field is not properly unmarked, then this function will loop forever and the caller will be stuck.
Taking From The Start Of A List
The next case to consider for lock-free list management with the Java Monitor subsystem is taking an ObjectMonitor from the start of a list. Taking an ObjectMonitor from the start of a list is a specialized form of delete that is guaranteed to interact with a thread that is prepending to the same list at the same time. Again, the core of the solution is to "mark" the next field in the ObjectMonitor at the head of the list we're trying to take the ObjectMonitor from, but we use slightly different code because we have less linkages to make than a prepend.
L01: static ObjectMonitor* take_from_start_of_common(ObjectMonitor* volatile * list_p,
L02: int volatile * count_p) {
L03: ObjectMonitor* next = NULL;
L04: ObjectMonitor* take = NULL;
L05: // Mark the list head to guard against A-B-A race:
L06: if (!mark_list_head(list_p, &take, &next)) {
L07: return NULL; // None are available.
L08: }
L09: // Switch marked list head to next (which unmarks the list head, but
L10: // leaves take marked):
L11: OrderAccess::release_store(list_p, next);
L12: Atomic::dec(count_p);
L13: // Unmark take, but leave the next value for any lagging list
L14: // walkers. It will get cleaned up when take is prepended to
L15: // the in-use list:
L16: set_next(take, next);
L17: return take;
L18: }
What the above function does is:
- Tries to mark the ObjectMonitor at the head of the list:
- Marking will only fail if the list is empty so that NULL can be returned.
- Otherwise mark_list_head() will loop until the ObjectMonitor at the list head has been marked.
- Updates the list head to refer to the next ObjectMonitor.
- Decrements the counter referred to by 'count_p'.
- Unmarks the next field in the taken ObjectMonitor.
- Tries to mark the ObjectMonitor at the head of the list:
The function can be called by more than one thread at a time and each thread will take a unique ObjectMonitor from the start of the list (if one is available) without losing any other ObjectMonitors on the list. Of course, the "take a unique ObjectMonitor" and "without losing any other ObjectMonitors" parts are where all the details come in:
- L06 tries to mark the list head:
- mark_list_head() returns false if the list is empty so we return NULL on L07.
- Otherwise, 'take' is a pointer to the marked list head and 'next' is the unmarked next field in the list head.
- L11 release-stores 'next' into 'list_p'.
You might be asking yourself: Why release-store instead of cmpxchg?- mark_list_head() only returns to the caller when it has marked the next field in the ObjectMonitor at the head of the list.
- Because of that guarantee, any prepender or deleter thread that is running in parallel must loop until we have release-stored 'next' into 'list_p' which unmarks the list head.
- L12 decrements the counter referred to by 'count_p'.
- L16 unmarks 'take' using the unmarked 'next' value we got from mark_list_head():
- Keeping the 'next' value in take's next field allows any lagging list walker to get to the next ObjectMonitor on that list.
- take's next field will get cleaned up when take is prepended to its target in-use list.
- L17 returns 'take' to the caller.
- L06 tries to mark the list head:
...
}The above function tries to lock the list head's ObjectMonitor:
- If the list is empty, NULL is returned.
- Otherwise, the list head's ObjectMonitor* is returned.
The function can be called by more than one thread on the same 'list_p' at a time. False is only returned when 'list_p' refers to an empty list. Otherwise only one thread will return an ObjectMonitor* at a time. Since the ObjectMonitor is locked, any parallel callers to get_list_head_locked() will loop until the list head's ObjectMonitor is no longer locked. That typically happens when the list head's ObjectMonitor is taken off the list and 'list_p' is advanced to the next ObjectMonitor on the list. Of course, making sure that "only one thread will return true at a time" is where all the details come in:
- L03 loads the current 'list_p' value into 'mid'.
- L0[45] is the empty list check and the only time that NULL is returned by this function.
- L07 tries to lock 'mid':
- If locking is not successful, we loop around to try it all again (the "spin" part of spin-lock).
- L08 loads the current 'list_p' value to see if it still matches 'mid':
- If the list head has changed, then we unlock mid on L10 and try it all again.
- Otherwise, 'mid' is returned.
When this function returns a non-NULL ObjectMonitor*, the ObjectMonitor is locked and any parallel callers of get_list_head_locked() on the same list will be looping until the list head's ObjectMonitor is no longer locked. The caller that just got the ObjectMonintor* needs to finish up its work quickly.
Debugging Tip: If there's a bug where the list head ObjectMonitor is not properly unlocked, then this function will loop forever and the caller will be stuck.
Taking From The Start Of A List
The next case to consider for spin-lock list management with the Java Monitor subsystem is taking an ObjectMonitor from the start of a list. Taking an ObjectMonitor from the start of a list is a specialized form of delete that is guaranteed to interact with a thread that is prepending to the same list at the same time. Again, the core of the solution is to lock the ObjectMonitor at the head of the list we're trying to take the ObjectMonitor from, but we use slightly different code because we have less linkages to make than a prepend.
L01: static ObjectMonitor* take_from_start_of_common(ObjectMonitor** list_p,
L02: int* count_p) {
L03: ObjectMonitor* take = NULL;
L04: // Lock the list head to guard against A-B-A race:
L05: if ((take = get_list_head_locked(list_p)) == NULL) {
L06: return NULL; // None are available.
L07: }
L08: ObjectMonitor* next = unmarked_next(take);
L09: // Switch locked list head to next (which unlocks the list head, but
L10: // leaves take locked):
L11: Atomic::store(list_p, next);
L12: Atomic::dec(count_p);
L13: // Unlock take, but leave the next value for any lagging list
L14: // walkers. It will get cleaned up when take is prepended to
L15: // the in-use list:
L16: om_unlock(take);
L17: return take;
L18: }
What the above function does is:
- Tries to lock the ObjectMonitor at the head of the list:
- Locking will only fail if the list is empty so that NULL can be returned.
- Otherwise get_list_head_locked() will loop until the ObjectMonitor at the list head has been locked.
- Get the next pointer from the taken ObjectMonitor.
- Updates the list head to refer to the next ObjectMonitor.
- Decrements the counter referred to by 'count_p'.
- Unlocks the taken ObjectMonitor.
- Tries to lock the ObjectMonitor at the head of the list:
The function can be called by more than one thread at a time and each thread will take a unique ObjectMonitor from the start of the list (if one is available) without losing any other ObjectMonitors on the list. Of course, the "take a unique ObjectMonitor" and "without losing any other ObjectMonitors" parts are where all the details come in:
- L05 tries to lock the list head:
- get_list_head_locked() returns NULL if the list is empty so we return NULL on L06.
- Otherwise, 'take' is a pointer to the locked list head.
- L08 gets the next pointer from take.
- L11 stores 'next' into 'list_p'.
You might be asking yourself: Why store instead of cmpxchg?- get_list_head_locked() only returns to the caller when it has locked the ObjectMonitor at the head of the list.
- Because of that guarantee, any prepender or deleter thread that is running in parallel must loop until we have stored 'next' into 'list_p' which unlocks the list head.
- L12 decrements the counter referred to by 'count_p'.
- L16 unlocks 'take':
- Keeping the 'next' value in take's next field allows any lagging list walker to get to the next ObjectMonitor on that list.
- take's next field will get cleaned up when take is prepended to its target in-use list.
- L17 returns 'take' to the caller.
- L05 tries to lock the list head:
lock_next_for_traversal() Helper Function
This last helper function exists for making life easier for list walker code. List walker code calls get_list_head_locked() to get the locked list head and then walks the list applying its particular logic to elements in the list. In order to safely walk to the 'next' ObjectMonitor in a list, the list walker code must lock the 'next' ObjectMonitor before unlocking the 'current' ObjectMonitor that it has locked. If a list walker unlocks 'current' before locking 'next', then there is race where 'current' could be modified to refer to something other than the 'next' value that was in place when 'current' was locked. By locking 'next' first and then unlocking 'current', the list walker can safely advance to 'next'.
L01: static ObjectMonitor* lock_next_for_traversal(ObjectMonitor* cur) {
L02: assert(is_locked(cur), "cur=" INTPTR_FORMAT " must be locked", p2i(cur));
L03: ObjectMonitor* next = unmarked_next(cur);
L04: if (next == NULL) { // Reached the end of the list.
L05: om_unlock(cur);
L06: return NULL;
L07: }
L08: om_lock(next); // Lock next before unlocking current to keep
L09: om_unlock(cur); // from being by-passed by another thread.
L10: return next;
L11: }
This function is pretty straight forward so there are no detailed notes for it.
Using The New Spin-Lock Monitor List Functions
ObjectSynchronizer::om_alloc(Thread* self, ...)
...
2) Try to allocate from the global free list (up to self→om_free_provision times):
- take_from_start_of_gglobal_free_list() takes an ObjectMonitor from the global free list (if possible).
- om_release(self, take, false) prepends the newly taken ObjectMonitor to self's free list.
- Retry the allocation from step 1.
3) Allocate a block of new ObjectMonitors:
- prepend_block_to_lists() prepends the newly allocated block to 'g_block_list' and to the global free list.
- Retry the allocation from step 1.
ObjectSynchronizer::om_release(Thread* self, ObjectMonitor* m, bool from_per_thread_alloc)
ObjectSynchronizer::om_release() is responsible for putting an ObjectMonitor on self's free list. If 'from_per_thread_alloc' is true, then om_release() is also responsible for extracting the ObjectMonitor from self's in-use list. The extraction from self's in-use list must happen first:
- , false) prepends the newly taken ObjectMonitor to self's free list.
- Retry the allocation from step 1.
3) Allocate a block of new ObjectMonitors:
- prepend_block_to_lists() prepends the newly allocated block to 'g_block_list' and to the global free list.
- Retry the allocation from step 1.
ObjectSynchronizer::om_release(Thread* self, ObjectMonitor* m, bool from_per_thread_alloc)
ObjectSynchronizer::om_release() is responsible for putting an ObjectMonitor on self's free list. If 'from_per_thread_alloc' is true, then om_release() is also responsible for extracting the ObjectMonitor from self's in-use list. The extraction from self's in-use list must happen first:
L01: if (from_per_thread_alloc) {
L02: if ((mid = get_list_head_locked(&self->om_in_use_list)) == NULL) {
L03: fatal("thread=" INTPTR_FORMAT " in-use list must not be empty.", p2i(self));
L04: }
L05: next = unmarked_next(mid);
L06: if (m == mid) {
L07: Atomic::store(&self->om_in_use_list, next);
L08: } else if (m == next) {
L09: mid = next;
L10: om_lock(mid);
L11: next = unmarked_next(mid);
L12: L01: if (from_per_thread_alloc) {
L02: mark_list_head(&self->om_in_use_list, &mid, &->set_next_om(next);
L03L13: while} (true)else {
L04: if (m == mid) {
L05: if (cur_mid_in_use =L14: ObjectMonitor* anchor = next;
L15: om_lock(anchor);
L16: om_unlock(mid);
L17: while ((mid = unmarked_next(anchor)) != NULL) {
L06L18: OrderAccess::release_store(&self->om_in_use_list, next);
L07: } else if (m == mid) {
L08L19: next = OrderAccess::release_store(&cur_mid_in_use->unmarked_next(mid);
L20: anchor->set_next_om, (next);
L21: L09: break;
L22: } else }{
L10L23: extracted = true;
L11: Atomic::dec(&self->om_in_use_count om_lock(mid);
L12L24: setom_next(mid, nextunlock(anchor);
L13L25: break;
L14: } anchor = mid;
L15L26: if (cur_mid_in_use != NULL) { }
L16L27: set_next(cur_mid_in_use, mid); // umark cur_mid_in_use}
L17L28: }
L18L29: cur_mid_in_use = mid;
L19: mid = nextAtomic::dec(&self->om_in_use_count);
L20L30: next = markom_next_loopunlock(mid);
L21L31: }
L22: }
L23L32: prepend_to_om_free_list(self, m);
Most of the above code block extracts 'm' from self's in-use list; it is not an exact quote from om_release(), but it is the highlights:
- L02 is used to mark lock self's in-use list head:
- 'mid' is self's in-use list head and its next field it is markedlocked.
- L05 'next' is the unmarked next field from 'mid'.
- L06 → L07: handle first special case where the target ObjectMonitor 'm' matches the list head.
- L08 → L12: handle second special case where the target ObjectMonitor 'm' matches next after the list head.
- L03 L14 → L21L30: self's in-use list is traversed looking for the target ObjectMonitor 'm':
- L04L18: if the current 'mid' matches 'm':
- L05: if cur_mid_in_use is NULL, we're still processing the head of the thread's in-use list so...
- L06: we cmpxchg the list head's next field from 'marked_mid' to 'next'.
- else
- L08: we cmpxchg the cur_mid_in_use's next field from 'marked_mid' to 'next'.
L10 → L13 - L19: get the next after 'm'
- L20: update the anchor to refer to the next after 'm'
- L21: break out since we found a match
- L05: if cur_mid_in_use is NULL, we're still processing the head of the thread's in-use list so...
- else
- L23: lock the current 'mid'
- L24-5: unlock the current anchor and advance to the new anchor
- loop around and try again
- L04L18: if the current 'mid' matches 'm':
- L29 → L30: we've successfully extracted 'm' from self's in-use list so we decrement self's in-use counter, unmark the next field in unlock 'mid' and we're done.
- L1[56]: if cur_mid_in_use != NULL, then unmark its next field.
- L18: set 'cur_mid_in_use' to 'mid'
Note: cur_mid_in_use keeps the marked next field so that it remains stable for a possible next field change. It cannot be deflated while it is marked. - L19: set 'mid' to 'next'. L20: mark next field in the new 'mid' and update 'next'; loop around and do it all again.
- L02 is used to mark lock self's in-use list head:
The last line of the code block (L23L32) prepends 'm' to self's free list.
...
ObjectSynchronizer::om_flush() is reponsible for flushing self's in-use list to the global in-use list and self's free list to the global free list during self's thread exit processing. om_flush() starts with self's in-use list:
L01: if (mark if ((in_use_list = get_list_head_locked(&self->om_in_use_list, &in_use_list, &next))) != NULL) {
L02: in_use_tail = in_use_list;
L03: in_use_count++;
L04: for (ObjectMonitor* cur_om = unmarked_next(in_use_list); cur_om != NULL;) {
L05: if (is_next_markedlocked(cur_om)) {
L06: while (is_locked(cur_om)) {
L07: os::naked_short_sleep(1);
L08: }
L09: cur_om = unmarked_next_marked(in_use_tail);
L10: continue;
L11: }
L12: if (cur_om->is_free()) {
L13: cur_om = unmarked_next(in_use_tail);
L07L14: continue;
os::naked_short_sleep(1)L15: }
L16: in_use_tail = cur_om;
L08L17: }in_use_count++;
L09L18: cur_om = unmarked_next(incur_use_tailom);
L10L19: continue;}
L11L20: } guarantee(in_use_tail != NULL, "invariant");
L12L21: ifint (!curl_om->is_active()) {_in_use_count = Atomic::load(&self->om_in_use_count);
L13L22: cur_om = unmarked_next(in_use_tail);
L14: continue;
L15: }
L16: in_use_tail = cur_om;
L17: in_use_count++;
L18: cur_om = unmarked_next(cur_om);
L19: }
L20: OrderAccess::release_ ADIM_guarantee(l_om_in_use_count == in_use_count, "in-use counts don't match: "
L23: "l_om_in_use_count=%d, in_use_count=%d", l_om_in_use_count, in_use_count);
L24: Atomic::store(&self->om_in_use_count, 0);
L21L25: OrderAccessAtomic::release_store(&self->om_in_use_list, (ObjectMonitor*)NULL);
L22L26: setom_nextunlock(in_use_list, next);
L23L27: }
The above is not an exact copy of the code block from om_flush(), but it is the highlights. What the above code block needs to do is pretty simple:
...
However, in this case, there are a lot of details:
- L01 marks locks the in-use list head (if it is not empty):
- 'in_use_list' is self's in-use list head and its next field is marked. 'next' is the unmarked next field from 'in_use_list'.
- The in-use list head is kept marked locked to prevent an async deflation thread from entering the list behind this thread.
Note: An async deflation thread does check to see if the target thread is exiting, but if it has made it past that check before this thread started exiting, then we're racing.
- L04-L18L19: loops over the in-use list counting and advancing 'in_use_tail'.
- L05-L10: 'cur_om' s next field is marked locked so there must be an async deflater thread or a list walker thread ahead of us so we delay to give it a chance to finish and refetch 'in_use_tail's (possibly changed) next field and try again.
- L12-L14: 'cur_om' was deflated and its allocation state was changed to Free while it was markedlocked. We just happened to be lucky enough to see it just after it was unmarked unlocked (and added to the free list). We refetch 'in_use_tail's (possibly changed) next field and try again.
- L1[67]: finally 'cur_om' has been completely vetted so we can update 'in_use_tail' and increment 'in_use_count'.
- L18: advance 'cur_om' to the next ObjectMonitor and do it all again.
- L20L24: release-store self's in-use count to zero.
Note: We clear self's in-use count before umarking unlocking self's in-use list head to avoid races. - L21L25: release-store self's in-use list head to NULL.
- L22L26: unmark unlock the disconnected list head.
Note: Yes, the next field in self's in-use list head was kept marked locked for the whole loop to keep any racing async deflater thread or list walker thread out of the in-use list. After L21L26, the racing async deflater thread will loop around and see self's in-use list is empty and bail out. Similarly, a racing list walker thread will retry and see self's in-use list is empty and bail out.
The code to process self's free list is much, much simpler because we don't have any races with an async deflater thread like self's in-use list. The only interesting bits:
- load-acquire self's free list head.
- release-store self's free list head count to NULLzero.
- release-store self's free list count head to zeroNULL.
The last interesting bits for this function are prepending the local lists to the right global places:
- prepend_list_to_gglobal_free_list(free_list, free_tail, free_count);
- prepend_list_to_gglobal_om_in_use_list(in_use_list, in_use_tail, in_use_count);
...
ObjectSynchronizer::deflate_monitor_list() is responsible for deflating idle ObjectMonitors at a safepoint. This function can use the simpler marklock-mid-as-we-go protocol since there can be no parallel list deletions due to the safepoint:
L01: int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor* volatile * list_p,
L02: int volatile * count_p,
L03: ObjectMonitor** free_head_p,
L04: ObjectMonitor** free_tail_p) {
L05: ObjectMonitor* cur_mid_in_use = NULL;
L06: ObjectMonitor* mid = NULL;
L07: ObjectMonitor* next = NULL;
L08: int deflated_count = 0;
L09: if (!mark((mid = get_list_head_locked(list_p, &mid, &next))) == NULL) {
L10: return 0; // The list is empty so nothing to deflate.
L11: }
L12: next = unmarked_next(mid);
L13: while (true) {
L13L14: oop obj = (oop) mid->object();
L14L15: if (obj != NULL && deflate_monitor(mid, obj, free_head_p, free_tail_p)) {
L15L16: if (cur_mid_in_use == NULL) {
L16L17: OrderAccessAtomic::release_store(list_p, next);
L17: L18: } else {
L18: OrderAccess::release_store(&L19: cur_mid_in_use->>set_next_om, (next);
L19L20: }
L20L21: deflated_count++;
L21L22: Atomic::dec(count_p);
L22L23: setmid->set_next_om(mid, NULL);
L23: mid = next;
L24: } else {
L25: setom_nextunlock(mid, next); // unmark next field
L26: cur_mid_in_use = mid; mid;
L27: }
L27L28: mid = next;
L28: }
L29: if (mid == NULL) {
L30: break; // Reached end of the list so nothing more to deflate.
L31: }
L32: om_lock(mid);
L33: next = markunmarked_next_loop(mid);
L33L34: }
L34L35: return deflated_count;
L35L36: }
Note: The above version of deflate_monitor_list() uses locking, but those changes were dropped during the code review cycle for JDK-8235795. The locking is only needed when additional calls to audit_and_print_stats() are used during debugging so it was decided that the pushed version would be simpler.
The above is not an exact copy of the code block from deflate_monitor_list(), but it is the highlights. What the above code block needs to do is pretty simple:
...
Since we're using the simpler mark-mid-as-we-go protocol, there are not too many details:
- L09 marks : locks the 'list_p' head (if it is not empty)
- L12: :
- 'mid' is 'list_p's head and its next field is marked. 'next' is the unmarked next field from 'mid'.
- L12L13-L34L33: We walk each 'mid' in the list and determine if it can be deflated:
- L14L15: if 'mid' is associated with an object and can be deflated:
- L15L16: if cur_mid_in_use is NULL, we're still processing the head of the in-use list so...
- L16L17: we cmpxchg store the list head 's next field from 'marked_mid' to 'next'.
- else
- L18L19: we cmpxchg the curset cur_mid_in_use's next field from 'marked_mid' to 'next'.
- L20 L21 → L23: we've successfully extracted 'mid' from 'list_p's list so we increment 'deflated_count', decrement the counter referred to by 'count_p', set 'mid's next field to NULL and we're done.
Note: 'mid' is the current tail in the 'free_head_p' list so we have to NULL terminate it (which also unmarks unlocks it).
- L15L16: if cur_mid_in_use is NULL, we're still processing the head of the in-use list so...
- L14L15: if 'mid' is associated with an object and can be deflated:
- L2[4-76]: else 'mid' can't be deflated so unmark unlock mid 's next field and advance both 'cur_mid_in_use' and .
- L28: advance 'mid'.
- L29 → L30]: we reached the end of the list so break out of the loop.
- L32: mark next field in lock the new 'mid'
- L33: and update 'next'; loop around and do it all again.
- L34L35: all done so return 'deflated_count'.
...
ObjectSynchronizer::deflate_monitor_list_using_JT() is responsible for asynchronously deflating idle ObjectMonitors using a JavaThread. This function uses the more complicated marklock-cur_mid_in_use-and-mid-as-we-go protocol because om_release() can do list deletions in parallel. We also marklock-next-next-as-we-go to prevent an om_flush() that is behind this thread from passing us. Because this function can asynchronously interact with so many other functions, this is the largest clip of code:
L01: int ObjectSynchronizer::deflate_monitor_list_using_JT(ObjectMonitor* volatile * list_p,
L02: int volatile * count_p,
L03: ObjectMonitor** free_head_p,
L04: ObjectMonitor** free_tail_p,
L05: ObjectMonitor** saved_mid_in_use_p) {
L06: JavaThread* self = JavaThread::current();
L07: ObjectMonitor* cur_mid_in_use = NULL;
L07L08: ObjectMonitor* mid = NULL;
L08L09: ObjectMonitor* next = NULL;
L09L10: ObjectMonitor* next_next = NULL;
L10L11: int deflated_count = 0;
L11: L12: NoSafepointVerifier nsv;
L13: if (*saved_mid_in_use_p == NULL) {
L12L14: if (!mark(mid = get_list_head_locked(list_p, &mid, &next)))) == NULL) {
L13L15: return 0; // The list is empty so nothing to deflate.
L14L16: }
L15: L17: next = unmarked_next(mid);
L18: } else {
L16L19: cur_mid_in_use = *saved_mid_in_use_p;
L20: om_lock(cur_mid_in_use);
L17L21: mid = markunmarked_next_loop(cur_mid_in_use);
L18L22: if (mid == NULL) {
L19L23: setom_nextunlock(cur_mid_in_use, NULL); // unmark next field
L20L24: *saved_mid_in_use_p = NULL;
L21L25: return 0; // The remainder is empty so nothing more to deflate.
L22L26: }
L27: om_lock(mid);
L23L28: next = markunmarked_next_loop(mid);
L24L29: }
L25L30: while (true) {
L26L31: if (next != NULL) {
L32: om_lock(next);
L27L33: next_next = markunmarked_next_loop(next);
L28L34: }
L29L35: if (mid->object() != NULL && mid->is_old() &&
L30L36: deflate_monitor_using_JT(mid, free_head_p, free_tail_p)) {
L31L37: if (cur_mid_in_use == NULL) {
L32L38: OrderAccessAtomic::release_store(list_p, next);
L33L39: } else {
L34L40: ObjectMonitor* markedlocked_next = mark_om_ptr(next);
L35L41: OrderAccess::release_store(& cur_mid_in_use->>set_next_om, marked(locked_next);
L36: L42: }
L37L43: deflated_count++;
L38L44: Atomic::dec(count_p);
L39L45: setmid->set_next_om(mid, NULL);
L40L46: mid = next; // mid keeps non-NULL next's markedlocked next fieldstate
L41L47: next = next_next;
L42L48: } else {
L43L49: if (cur_mid_in_use != NULL) {
L44L50: setom_nextunlock(cur_mid_in_use, mid); // umark cur_mid_in_use
L45L51: }
L46L52: cur_mid_in_use = mid;
L47L53: mid = next; // mid keeps non-NULL next's markedlocked next fieldstate
L48L54: next = next_next;
L49L55: if (SafepointSynchronizeSafepointMechanism::isshould_synchronizingblock(self) &&
L50L56: cur_mid_in_use != OrderAccessAtomic::load_acquire(list_p) &&
L51: cur_mid_in_use->is_old()) {
L52L57: *saved_mid_in_use_p = cur_mid_in_use;
L53L58: setom_nextunlock(cur_mid_in_use, mid); // umark cur_mid_in_use
L54L59: if (mid != NULL) {
L55L60: setom_nextunlock(mid, next); // umark mid
L56L61: }
L57L62: return deflated_count;
L58L63: }
L59L64: }
L60L65: if (mid == NULL) {
L61L66: if (cur_mid_in_use != NULL) {
L62L67: setom_nextunlock(cur_mid_in_use, mid); // umark cur_mid_in_use
L63L68: }
L64L69: break; // Reached end of the list so nothing more to deflate.
L65L70: }
L66L71: }
L67L72: *saved_mid_in_use_p = NULL;
L68L73: return deflated_count;
L69L74: }
...
The above is not an exact copy of the code block from deflate_monitor_list_using_JT(), but it is the highlights. What the above code block needs to do is pretty simple:
...
Since we're using the more complicated marklock-cur_mid_in_use-and-mid-as-we-go protocol and also the marklock-next-next-as-we-go protocol, there is a mind numbing amount of detail:
- L1[13-37]: Handle the initial setup if we are not resuming after a safepoint or a handshake:
- L12 marks L14: locks the 'list_p' head (if it is not empty):
- 'mid' is 'list_p's head and its next field is marked. L17: 'next' is the unmarked next field from 'mid'.
- L1[13-37]: Handle the initial setup if we are not resuming after a safepoint or a handshake:
- L15L18-L23L28: Handle the initial setup if we are resuming after a safepoint or a handshake:
- L17: mark next field in L20: lock 'cur_mid_in_use' and
- L21: update 'mid'
- L18L22-L21L25: If 'mid' == NULL, then we've resumed context at the end of the list so we're done.
- L23: mark next field in L27: lock 'mid' and
- L28: update 'next'
- L25L30-L64L71: We walk each 'mid' in the list and determine if it can be deflated:
- L2L3[671-3]: if next != NULL, then mark next field in lock 'next' and update 'next_next'
- L29L35-L41L47: if 'mid' is associated with an object, 'mid' is old, and can be deflated:
- L31L37: if cur_mid_in_use is NULL, we're still processing the head of the in-use list so...
- L32L38: we cmpxchg store the list head 's next field from 'marked_mid' to 'next'.
- else
- L34L40: make a marked locked copy of 'next'
- L35L41: we cmpxchg the curset cur_mid_in_use's next field from to 'markedlocked_mid' to 'next'.
- L37 L43 → L39L45: we've successfully extracted 'mid' from 'list_p's list so we increment 'deflated_count', decrement the counter referred to by 'count_p', set 'mid's next field to NULL and we're done.
Note: 'mid' is the current tail in the 'free_head_p' list so we have to NULL terminate it (which also unmarks unlocks it). - L40L46: advance 'mid' to 'next'.
Note: 'mid' keeps non-NULL 'next's marked next field.locked state - L47L41: advance 'next' to 'next_next'.
- L31L37: if cur_mid_in_use is NULL, we're still processing the head of the in-use list so...
- L42L48-L57L63: 'mid' can't be deflated so we have to carefully advance the list pointers:
- L4[34]L49,50: if cur_mid_in_use != NULL, then unmark next field in unlock 'cur_mid_in_use'.
- L46L52: advance 'cur_mid_in_use' to 'mid'.
Note: The next field in 'mid' is still marked locked and 'cur_mid_in_use' keeps that state. - L47L53: advance 'mid' to 'next'.
Note: The next field in a A non-NULL 'next' is still marked locked and 'mid' keeps that state. - L48L54: advance 'next' to 'next_next'.
- L49L55-L57L62: Handle a safepoint or a handshake if one has started and it is safe to do so.
- L60L65-L64L69: we reached the end of the list:
- L6[1267]: if cur_mid_in_use != NULL, then unmark next field in unlock 'cur_mid_in_use'.
- L64L69: break out of the loop because we are done
- L67L72: not pausing for a safepoint or handshake so clear saved state.
- L68L73: all done so return 'deflated_count'.
...
ObjectSynchronizer::deflate_idle_monitors() handles deflating idle monitors at a safepoint from the global in-use list using ObjectSynchronizer::deflate_monitor_list(). There are only a few things that are worth mentioning:
- OrderAccessAtomic::load_acquire(&gom_omlist_globals.in_use_list) is used to get the latest global in-use list.OrderAccess
- Atomic::load_acquire(&gom_omlist_globals.in_use_count) is used to get the latest global in-use count.
- prepend_list_to_gglobal_free_list(free_head_p, free_tail_p, deflated_count) is used to prepend the deflated ObjectMonitors on the global free list.
...
ObjectSynchronizer::deflate_common_idle_monitors_using_JT() handles asynchronously deflating idle monitors from either the global in-use list or a per-thread in-use list using ObjectSynchronizer::deflate_monitor_list_using_JT(). There are only a few things that are worth mentioning:
- OrderAccessAtomic::load_acquire(&gom_omlist_globals.in_use_count) is used to get the latest global in-use count.
- OrderAccessAtomic::load_acquire(&target→om_in_use_count) is used to get the latest per-thread in-use count.
- prepend_list_to_gglobal_free_list(free_head_p, free_tail_p, local_deflated_count) is used to prepend the deflated ObjectMonitors on the global free list.
...
- New diagnostic option '-XX:AsyncDeflateIdleMonitors' that is default 'true' so that the new mechanism is used by default, but it can be disabled for potential failure diagnosis.
- ObjectMonitor deflation is still initiated or signaled as needed at a safepoint. When Async Monitor Deflation is in use, flags are set so that the work is done by the ServiceThread which offloads the safepoint cleanup mechanism.
- Having the ServiceThread deflate a potentially long list of in-use monitors could potentially delay the start of a safepoint. This is detected in ObjectSynchronizer::deflate_monitor_list_using_JT() which will save the current state when it is safe to do so and return to its caller to drop locks as needed before honoring the safepoint request.
- New diagnostic option '-XX:AsyncDeflationInterval' that is default 250 millis; this this option controls how frequently we async default idle monitors when MonitorUsedDeflationThreshold is exceeded.New diagnostic option '-XX:HandshakeAfterDeflateIdleMonitors' that is default false on the LP64 X64 platform and default true on other platforms that implement C2is exceeded.
- Everything else is just monitor list management, infrastructure, logging, debugging and the like. :-)
...
- For this option, exceeded means:
((gom_omlist_globals.population - gom_omlist_globals.free_count) / gom_omlist_globals.population) > NN%
- If MonitorBound is exceeded (default is 0 which means off), cleanup safepoint will be induced.
- For this option, exceeded means:
(g_om_population - g_om_free_count) > MonitorBound
...
- Changes to the safepoint deflation mechanism by the Async Monitor Deflation project (when async deflation is enabled):
- If System.gc() is called, then a special deflation request is made which invokes the safepoint deflation mechanism.
- Added the AsyncDeflationInterval diagnostic option (default 250 millis, 0 means off) to prevent MonitorUsedDeflationThreshold requests from swamping the ServiceThread.
- Description: Async deflate idle monitors every so many milliseconds when MonitorUsedDeflationThreshold is exceeded (0 is off).
- A special deflation request can cause an async deflation to happen sooner than AsyncDeflationInterval.
- SafepointSynchronize::dois_cleanup_tasksneeded() now calls:
- ObjectSynchronizer::is_safepoint_deflation_needed() instead of ObjectSynchronizer::is_cleanup_needed().
- is_safepoint_deflation_needed() returns true only if a special deflation request is made (see abovea special deflation request is made (see above).
- SafepointSynchronize::do_cleanup_tasks() now (indirectly) calls:
- ObjectSynchronizer::do_safepoint_work() instead of ObjectSynchronizer::deflate_idle_monitors().
- do_cleanup_tasks() can be called for non deflation related cleanup reasons and that will still result in a call to do_safepoint_work().
- ObjectSynchronizer::do_safepoint_work() only does the safepoint cleanup tasks if there is a special deflation request. Otherwise it just sets the is_async_deflation_requested flag and notifies the ServiceThread.
- ObjectSynchronizer::deflate_idle_monitors() and ObjectSynchronizer::deflate_thread_local_monitors() do nothing unless there is a special deflation request.
...
Other invocation changes by the Async Monitor Deflation project (when async deflation is enabled):
VM_Exit::doit_prologue() will request a special cleanup to reduce the noise in 'monitorinflation' logging at VM exit time.
Before the final safepoint in a non-System.exit() end to the VM, we will request a special cleanup to reduce the noise in 'monitorinflation' logging at VM exit time.
- The following whitebox test functions will request a special cleanup:
WB_G1StartMarkCycle()
- WB_FullGC()
- WB_ForceSafepoint()
Gory Details
- Counterpart function mapping for those that know the existing code:
- ObjectSynchronizer class:
- deflate_idle_monitors() has deflate_idle_monitors_using_JT(), deflate_global_idle_monitors_using_JT(), deflate_per_thread_idle_monitors_using_JT(), and deflate_common_idle_monitors_using_JT().
- deflate_monitor_list() has deflate_monitor_list_using_JT()
- deflate_monitor() has deflate_monitor_using_JT()
- ObjectMonitor class:
- clear() has clear_using_JT()
- ObjectSynchronizer class:
- These functions recognize the Async Monitor Deflation protocol and adapt their operations:
- ObjectMonitor::enter()
- ObjectMonitor::EnterI()ObjectMonitor::ReenterI()
- ObjectSynchronizer::quick_enter()
- ObjectSynchronizer::deflate_monitor()
- Note: These changes include handling the lingering owner == DEFLATER_MARKER value.
- Also these functions had to adapt and retry their operations:
- ObjectSynchronizer::FastHashCode()
- ObjectSynchronizer::current_thread_holds_lock()
- ObjectSynchronizer::query_lock_ownership()
- ObjectSynchronizer::get_lock_owner()
- ObjectSynchronizer::monitors_iterate()
- ObjectSynchronizer::inflate_helper()
- ObjectSynchronizer::inflate()
- Various assertions had to be modified to pass without their real check when AsyncDeflateIdleMonitors is true; this is due to the change in semantics for the ObjectMonitor owner field.
- ObjectMonitor has a new allocation_state field that supports three states: 'Free', 'New', 'Old'. Async Monitor Deflation is only applied to ObjectMonitors that have reached the 'Old' state.
- Note: Prior to CR1/v2.01/4-for-jdk13, the allocation state was transitioned from 'New' to 'Old' in deflate_monitor_via_JT(). This meant that deflate_monitor_via_JT() had to see an ObjectMonitor twice before deflating it. This policy was intended to prevent oscillation from 'New' → 'Old' and back again.
- In CR1/v2.01/4-for-jdk13, the allocation state is transitioned from 'New' -> "Old" in inflate(). This makes ObjectMonitors available for deflation earlier. So far there has been no signs of oscillation from 'New' → 'Old' and back again.
- .
- The ObjectMonitor::owner() accessor detects DEFLATER_MARKER and returns NULL in that case to minimize the places that need to understand the new DEFLATER_MARKER value.
- System.gc()/JVM_GC() causes a special monitor list cleanup request which uses the safepoint based monitor list mechanism. So even if AsyncDeflateIdleMonitors is enabled, the safepoint based mechanism is still used by this special case.
- This is necessary for those tests that do something to cause an object's monitor to be inflated, clear the only reference to the object and then expect that enough System.gc() calls will eventually cause the object to be GC'ed even when the thread never inflates another object's monitor. Yes, we have several tests like that. :-)
Overview
Content Tools