Search native code in all R2R ni.dll in version bubble #57277

y-yamshchikov · 2021-08-12T11:53:13Z

This PR fixes part of #44948 and fixes #46160.

This code simply traverses through assemblies in the application
domain. For each assembly (module) it realizes is it Ready To Run and is
it in the same bubble with (does it deliberately bubbling the) module
from which generic function originates. If so, it makes request for code
is Ready To Run and hopes there is some in the module. If the request
succeeds it proceeds with found pointer to the bare native code.

Generic instantiations like System.Collections.Concurrent.ConcurrentDictionary`2[System.Int32,System.__Canon] contained within same version bubble are either a bug in crossgen2 or should be taken care of by having PGO data.

Now such methods use their Ready To Run code.

We have got significant performance gain on startup: 7% average on our representative set on Tizen.

This PR worked out notices in PR below:
#47269
about linear search through the set of RangeSections. In this new PR we propose storing of RangeSections in sorted array (with number of optimizations inspired by prior linked list based solution).

We have extensively tested this PR on armel/Tizen platform so in this case we are confident in reliability and profitability of the solution.

Dear colleagues @jkotas @alpencolt @gbalykov @t-mustafin, please take a look.

src/coreclr/vm/codeman.cpp

src/coreclr/vm/codeman.h

src/coreclr/vm/prestub.cpp

src/coreclr/vm/codeman.h

src/coreclr/vm/codeman.cpp

alpencolt · 2021-09-03T14:34:35Z

@jkotas had you time to check PR?

jkotas · 2021-09-03T16:00:31Z

src/coreclr/vm/codeman.h

This needs comment on what this returns. It does not return a simple index...

Also, can it be in the private section (before public:)?

jkotas · 2021-09-03T16:01:32Z

src/coreclr/vm/codeman.h

In debug/checked builds, we should make the initial size small and make the array grow on element at a time to ensure that the array resize algorithm and the potential race conditions are exercised.

jkotas · 2021-09-03T16:02:18Z

src/coreclr/vm/codeman.h

Why does this need to be volatile?

jkotas · 2021-09-03T16:05:57Z

src/coreclr/vm/codeman.h

I do not think that this needs to store HighAddress. It can just store the LowAddress and the method that is binary searching inside the array can just return the candidate RS index. We can then validate that the candidate RS index actually fits by comparing the address with RangeSection::HighAddress.

Side-effect of this change is going to be that the size of this struct will be power of two that will make the binary search faster.

jkotas · 2021-09-03T16:17:18Z

src/coreclr/vm/prestub.cpp

BEGIN_PRESERVE_LAST_ERROR should be before the newly added code.

jkotas · 2021-09-03T16:21:38Z

src/coreclr/vm/prestub.cpp

It does not look right to create a dummy ExternalMethodFrame just to get the return address.

The fixups from R2R modules should come via ExternalMethodFixupWorker. Can we just use the module that is computed there?

jkotas · 2021-09-03T16:23:00Z

RangeSections now stored in sorted array

BTW: I do like this change - it should be general performance improvement for any scenario with large number of modules.

and make it private

Binary search algorighm in FindRangeSectionHandleHelper changes accordingly.

jkotas · 2021-10-18T13:40:42Z

src/coreclr/vm/codeman.cpp

It is important for this path to be lock-free. Adding a lock here is very likely to regress GC stackwalking performance on machines with a lot of cores.

If it is so important, we will investigate another way to isolate readers from writers. We think we could find one with two interchanging arrays.

jkotas · 2021-10-18T13:41:08Z

src/coreclr/vm/codeman.cpp

I am not sure what this is trying to achieve, but it does not look right.

We are obliged to ask if the lock has been aquired after NoHostCalls constructor of ReaderLockHolder. NoHostCalls is necessary because of contract requirement from very above on call stack. We have encountered contract violation if we use default constructor (and the source of violation is not GetRangeSection, it is much higher).

mangod9 · 2021-11-15T17:07:09Z

hi @y-yamshchikov, assume you are still working on this PR feedback?

mangod9 · 2022-01-10T16:08:41Z

hi @y-yamshchikov, checking if this PR still needs to be kept open since it hasnt been updated in a couple of months. Thx

y-yamshchikov · 2022-01-13T12:18:42Z

hi @y-yamshchikov, checking if this PR still needs to be kept open since it hasnt been updated in a couple of months. Thx

Dear @mangod9, we are working on new approach granting lockfree reading of RangeSections and still benefiting from binary search capabilities of sorted array, fighting all the contradictions dictating by the multiprocessor environment. We end up with new version of synchronization mechanics and for now working on comprehensive testing environment which grants extensive coverage of all possible synchronization issues. To be presented soon, both the code and the testing model.

mangod9 · 2022-01-13T15:39:03Z

thanks for the update, should we close this PR for now and open a new one once ready? Thx!

y-yamshchikov · 2022-01-15T16:14:07Z

thanks for the update, should we close this PR for now and open a new one once ready? Thx!

There is lot of useful feedback conversation still not outdated, I think the change is made in a successive way to prolong current PR.

mangod9 · 2022-02-07T16:23:54Z

I would suggest we close the PR and reopen when more changes are added. This helps to keep the "Open PRs" list manageable.

mangod9 · 2022-03-14T18:04:30Z

Closing for now so it doesnt get flagged in stale PRs. Please reopen when ready to merge again. Thx.

ghost added area-crossgen2-coreclr community-contribution Indicates that the PR has been added by a community member labels Aug 12, 2021

y-yamshchikov changed the title ~~Partial fix 44948~~ Search native code in all R2R ni.dll in version bubble Aug 12, 2021

y-yamshchikov force-pushed the partial-fix-44948 branch 7 times, most recently from c2d2f5d to 91e753a Compare August 13, 2021 17:50

gbalykov suggested changes Aug 17, 2021

View reviewed changes

jaredpar mentioned this pull request Aug 17, 2021

bundled_self_contained_nocompression_app_run_succeeds blocking CI #53551

Closed

gbalykov suggested changes Aug 17, 2021

View reviewed changes

y-yamshchikov force-pushed the partial-fix-44948 branch 3 times, most recently from da2f1cd to 8dc0aa9 Compare September 3, 2021 12:50

jkotas reviewed Sep 3, 2021

View reviewed changes

src/coreclr/vm/codeman.h Outdated

Copy link

Member

jkotas Sep 3, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why does this need to be volatile?

jkotas reviewed Sep 3, 2021

View reviewed changes

src/coreclr/vm/prestub.cpp Outdated

Copy link

Member

jkotas Sep 3, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

BEGIN_PRESERVE_LAST_ERROR should be before the newly added code.

jkotas reviewed Sep 3, 2021

View reviewed changes

y-yamshchikov force-pushed the partial-fix-44948 branch 5 times, most recently from 3d7cad1 to fe70768 Compare September 16, 2021 11:24

Probing R2R code in caller modules

2e0547f

y-yamshchikov added 12 commits September 27, 2021 16:28

RangeSections now stored in sorted array

998c46a

debug increments of RangeSectionHandleArrayCapacity

3caac2e

comment on FindRangeSectionHandleHelper

e0806bf

and make it private

fix build codeman.cpp

96b9671

RangeSectionHandle.pRS is now PTR_RangeSection

23ddc25

debug increments: initial array size is now 8 for _DEBUG

ea27c1d

fix error in AddRangeSection

5d4654d

Remove HighAddress from RangeSectionHandle

4af5665

Binary search algorighm in FindRangeSectionHandleHelper changes accordingly.

Refactoring getting caller module

477535c

LastUsedRSIndex negative case bug fix

486ffdb

GetRangeSection now work only with reader lock

78444ee

rearranging reader lock in FindCodeRange and GetRangeSection

b3cebf2

y-yamshchikov force-pushed the partial-fix-44948 branch from fe70768 to b3cebf2 Compare October 13, 2021 13:09

runfoapp bot mentioned this pull request Oct 13, 2021

System.Runtime.Serialization.Xml.Tests failed #59926

Closed

y-yamshchikov force-pushed the partial-fix-44948 branch from 0a80132 to 0d663c4 Compare October 15, 2021 09:10

rlh with NoHostCalls in GetRangeSection

84187af

y-yamshchikov force-pushed the partial-fix-44948 branch from 0d663c4 to 84187af Compare October 18, 2021 09:23

jkotas reviewed Oct 18, 2021

View reviewed changes

active wait with rlh.Acquired()

80f8f73

y-yamshchikov force-pushed the partial-fix-44948 branch from 6df483d to 80f8f73 Compare October 18, 2021 13:44

mangod9 closed this Mar 14, 2022

ghost locked as resolved and limited conversation to collaborators Apr 13, 2022

Search native code in all R2R ni.dll in version bubble #57277

Search native code in all R2R ni.dll in version bubble #57277

Uh oh!

Conversation

y-yamshchikov commented Aug 12, 2021

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

alpencolt commented Sep 3, 2021

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jkotas Sep 3, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jkotas commented Sep 3, 2021

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

y-yamshchikov Oct 19, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mangod9 commented Nov 15, 2021

Uh oh!

mangod9 commented Jan 10, 2022

Uh oh!

y-yamshchikov commented Jan 13, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mangod9 commented Jan 13, 2022

Uh oh!

y-yamshchikov commented Jan 15, 2022

Uh oh!

mangod9 commented Feb 7, 2022

Uh oh!

mangod9 commented Mar 14, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

jkotas Sep 3, 2021 •

edited

Loading

y-yamshchikov Oct 19, 2021 •

edited

Loading

y-yamshchikov commented Jan 13, 2022 •

edited

Loading