1980056 - Limit mozjemalloc "logical" page size to 4KB

Status: RESOLVED FIXED

(Core :: Memory Allocator, task, P1)

Description

[Paul Bone [:pbone]]

To ensure that regions can be byte-addressed within runs (Bug 1980047) we need to limit mozjemalloc's page size to 4096 bytes, even on systems where the physical/OS page size is larger. This means separating mozjemalloc's concept of page size into two parts, a "physical" page size and a "logical" page size. OS interfaces like decommitting memory will now have many edge cases that must be handled. This depends on Bug 1921908 which provides a refactoring that will make this change much simpler.

Comment 1

[Paul Bone [:pbone]]

Attachment: WIP: Bug 1980056 - pt 5. Mozjemalloc can now use page sizes smaller than the real size

Comment 2

[Paul Bone [:pbone]]

Attachment: Bug 1980056 - pt 6. Remove the sub-page size class r=glandium

Sub-page size classes will be unsed so remove their logic.

Comment 3

[Paul Bone [:pbone]]

Attachment: WIP: Bug 1980056 - pt 1. Correct a comment

Comment 4

[Paul Bone [:pbone]]

Attachment: WIP: Bug 1980056 - pt 2. Merge purge results, ReachedThreshold and Busy

Comment 5

[Paul Bone [:pbone]]

Attachment: WIP: Bug 1980056 - pt 3. Change the dirty chunks tree to a list

Arenas track dirty chunks in a red-black tree ordered by memory address.
Switching this to a list will make it faster to test if an item is in
the list, which is required by the next patch.

Additionally there's no reason this collection needs to be ordered or
support lookups. The ordering by memory address may have some
fragmentation benefits, the new list is operated as a FIFO. Items are
appended at the tail when they get their first dirty page, and removed
from the head as they're purged.

Comment 6

[Paul Bone [:pbone]]

Attachment: WIP: Bug 1980056 - pt 4. mChunksDirty doesn't always track all dirty chunks

Currently a dirty chunk (that's a chunk with at least one dirty page)
must:

• Be in mChunksDirty
• Be busy pruging by another thread.

The next change will allow it so that chunks with a non-zero mNumDirty
might not be in the dirty chunks list because its impossible to purge
their pages. Therefore in this change we relax the above assumption.

Comment 7

[Phabricator Automation]

Comment on attachment 9510759 [details]
WIP: Bug 1980056 - pt 1. Correct a comment

Revision D263334 was moved to bug 1987055. Setting attachment 9510759 [details] to obsolete.

Comment 8

[Phabricator Automation]

Comment on attachment 9510760 [details]
WIP: Bug 1980056 - pt 2. Merge purge results, ReachedThreshold and Busy

Revision D263335 was moved to bug 1987055. Setting attachment 9510760 [details] to obsolete.

Comment 9

[Phabricator Automation]

Comment on attachment 9510761 [details]
WIP: Bug 1980056 - pt 3. Change the dirty chunks tree to a list

Revision D263336 was moved to bug 1987055. Setting attachment 9510761 [details] to obsolete.

Comment 10

[Phabricator Automation]

Comment on attachment 9510762 [details]
WIP: Bug 1980056 - pt 4. mChunksDirty doesn't always track all dirty chunks

Revision D263337 was moved to bug 1987055. Setting attachment 9510762 [details] to obsolete.

Comment 11

[Phabricator Automation]

Comment on attachment 9503950 [details]
WIP: Bug 1980056 - pt 5. Mozjemalloc can now use page sizes smaller than the real size

Revision D259219 was moved to bug 1987055. Setting attachment 9503950 [details] to obsolete.

Comment 12

[Paul Bone [:pbone]]

Attachment: Bug 1980056 - pt 5. Limit logical page size to 4KiB r=glandium

Comment 13

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/01d083e3c60f
https://hg.mozilla.org/integration/autoland/rev/4b4fb5fd1f38
pt 1. Limit logical page size to 4KiB r=glandium
https://github.com/mozilla-firefox/firefox/commit/784e006a344c
https://hg.mozilla.org/integration/autoland/rev/007298d99ba4
pt 2. Remove the sub-page size class r=glandium

Comment 14

[amarc]

https://hg.mozilla.org/mozilla-central/rev/4b4fb5fd1f38
https://hg.mozilla.org/mozilla-central/rev/007298d99ba4

Comment 15

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/15e87a34961f
https://hg.mozilla.org/integration/autoland/rev/cdf9cf902cdd
Revert "Bug 1980056 - pt 2. Remove the sub-page size class r=glandium" for causing Bug 2010505.

Comment 16

[Cosmin Sabou [:CosminS]]

Backout merged to central: https://hg.mozilla.org/mozilla-central/rev/cdf9cf902cdd

Comment 17

[Paul Bone [:pbone]]

Attachment: Bug 1980056 - pt 1. Use REAL_PAGE_CEILING in malloc_good_size r=glandium

Huge allocations will always be rounded to real pages.

Comment 18

[Paul Bone [:pbone]]

Attachment: Bug 1980056 - pt 2. Use real page size in one of the jemalloc tests r=glandium

The test uses page size to calculate the size of the chunk header for
testing jemalloc_ptr_info. Using real page size is correct for finding
the size of the chunk header.

Comment 19

[Paul Bone [:pbone]]

Attachment: Bug 1980056 - pt 3. Use gRealPageSize to qualify huge in-place realloc r=glandium

Realloc of huge allocations used the wrong page size to determine if
they're safe to do in-place. This patch fixes this.

Comment 20

[Paul Bone [:pbone]]

Attachment: Bug 1980056 - pt 4. Make sure that madvised pages are aligned with real pages r=glandium

When the allocator touches previously-madvised memory it must clear
madvised bits of trailing pages to keep them aligned with
real pages.

Comment 21

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/c06f4fd6abc0
https://hg.mozilla.org/integration/autoland/rev/94615f8bf640
pt 1. Use REAL_PAGE_CEILING in malloc_good_size r=glandium
https://github.com/mozilla-firefox/firefox/commit/f3df305fd228
https://hg.mozilla.org/integration/autoland/rev/4d3ec02a3564
pt 2. Use real page size in one of the jemalloc tests r=glandium
https://github.com/mozilla-firefox/firefox/commit/d9a447eb775f
https://hg.mozilla.org/integration/autoland/rev/48be5ff02dcf
pt 3. Use gRealPageSize to qualify huge in-place realloc r=glandium
https://github.com/mozilla-firefox/firefox/commit/b54c56c1a2df
https://hg.mozilla.org/integration/autoland/rev/c88c117b2663
pt 4. Make sure that madvised pages are aligned with real pages r=glandium
https://github.com/mozilla-firefox/firefox/commit/ea388bd5508b
https://hg.mozilla.org/integration/autoland/rev/97bf4a05fb87
pt 5. Limit logical page size to 4KiB r=glandium
https://github.com/mozilla-firefox/firefox/commit/2ec92a906edb
https://hg.mozilla.org/integration/autoland/rev/664ac9f557d1
pt 6. Remove the sub-page size class r=glandium

Comment 22

[agoloman]

https://hg.mozilla.org/mozilla-central/rev/94615f8bf640
https://hg.mozilla.org/mozilla-central/rev/4d3ec02a3564
https://hg.mozilla.org/mozilla-central/rev/48be5ff02dcf
https://hg.mozilla.org/mozilla-central/rev/c88c117b2663
https://hg.mozilla.org/mozilla-central/rev/97bf4a05fb87
https://hg.mozilla.org/mozilla-central/rev/664ac9f557d1