mm/vmalloc: reclaim tail resources on large vrealloc() shrink
cover mm/vmalloc: reclaim tail resources on large vrealloc() shrink
Hi,
This RFC explores closing the resource retention gap in the vmalloc-backed shrink path of vrealloc().
Today, when a vmalloc-backed allocation is shrunk, vrealloc() updates the requested size but can keep most of the old vmalloc mapping and backing pages alive. For sufficiently large shrink operations, this can retain a large amount of tail resources even though the logical object became much smaller.
This first RFC keeps the scope intentionally conservative:
- only ordinary VM_ALLOC areas
- only page_order == 0 allocations
- skip more complex vmalloc object types
- only reclaim tail resources when the retained waste is at least PMD_SIZE
The current evidence supports this as a resource reclamation fix rather than a workload-tuned performance optimization. Local validation currently covers:
- synthetic large shrink correctness
- shrink-then-grow regression
- threshold boundary correctness for the current heuristic
- KASAN run-rootfs vmalloc_oob regression coverage
I would especially appreciate feedback on:
1. whether this shrink direction is desirable upstream at all
2. whether the initial object-type restrictions are reasonable
3. whether a conservative PMD_SIZE threshold is an acceptable first heuristic
4. what kind of in-tree regression test would be preferred
Thanks.
fujunjie (1):
mm/vmalloc: reclaim tail resources on large vrealloc() shrink
mm/vmalloc.c | 105 ++++++++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 100 insertions(+), 5 deletions(-)
Signature · 1 line
2.34.1
Quoted 25 lines · fujunjie
Hi,
This RFC explores closing the resource retention gap in the vmalloc-backed shrink path of
vrealloc().Today, when a vmalloc-backed allocation is shrunk,
vrealloc()updates the requested size but can keep most of the old vmalloc mapping and backing pages alive. For sufficiently large shrink operations, this can retain a large amount of tail resources even though the logical object became much smaller.This first RFC keeps the scope intentionally conservative:
- only ordinary
VM_ALLOCareas
- onlypage_order== 0 allocations
- skip more complex vmalloc object types
- only reclaim tail resources when the retained waste is at leastPMD_SIZEThe current evidence supports this as a resource reclamation fix rather than a workload-tuned performance optimization. Local validation currently covers:
- synthetic large shrink correctness
- shrink-then-grow regression
- threshold boundary correctness for the current heuristic
- KASAN run-rootfsvmalloc_oobregression coverage
Quoted 7 lines
I would especially appreciate feedback on:
1. whether this shrink direction is desirable upstream at all
2. whether the initial object-type restrictions are reasonable
3. whether a conservativePMD_SIZEthreshold is an acceptable first heuristic
4. what kind of in-tree regression test would be preferred
Could you please have a look at this work:
https://lore.kernel.org/all/[email protected]/
Shivam is working on the same feature. Could you please check?
--
Uladzislau Rezki
Quoted 33 lines · Uladzislau Rezki
Quoted 25 lines · fujunjie
Hi,
This RFC explores closing the resource retention gap in the vmalloc-backed shrink path of
vrealloc().Today, when a vmalloc-backed allocation is shrunk,
vrealloc()updates the requested size but can keep most of the old vmalloc mapping and backing pages alive. For sufficiently large shrink operations, this can retain a large amount of tail resources even though the logical object became much smaller.This first RFC keeps the scope intentionally conservative:
- only ordinary
VM_ALLOCareas
- onlypage_order== 0 allocations
- skip more complex vmalloc object types
- only reclaim tail resources when the retained waste is at leastPMD_SIZEThe current evidence supports this as a resource reclamation fix rather than a workload-tuned performance optimization. Local validation currently covers:
- synthetic large shrink correctness
- shrink-then-grow regression
- threshold boundary correctness for the current heuristic
- KASAN run-rootfsvmalloc_oobregression coverageQuoted 7 lines
I would especially appreciate feedback on:
1. whether this shrink direction is desirable upstream at all
2. whether the initial object-type restrictions are reasonable
3. whether a conservativePMD_SIZEthreshold is an acceptable first heuristic
4. what kind of in-tree regression test would be preferred
Quoted 7 lines
Could you please have a look at this work:
https://lore.kernel.org/all/[email protected]/Shivam is working on the same feature. Could you please check?
--
Uladzislau Rezki
Thanks for the pointer, and sorry for missing Shivam's series before sending this RFC.
I will read through it first and avoid duplicating the effort.
Thanks,
fujunjie
Quoted 39 lines · Fujunjie
Quoted 36 lines · Uladzislau Rezki
Quoted 25 lines · fujunjie
Hi,
This RFC explores closing the resource retention gap in the vmalloc-backed shrink path of
vrealloc().Today, when a vmalloc-backed allocation is shrunk,
vrealloc()updates the requested size but can keep most of the old vmalloc mapping and backing pages alive. For sufficiently large shrink operations, this can retain a large amount of tail resources even though the logical object became much smaller.This first RFC keeps the scope intentionally conservative:
- only ordinary
VM_ALLOCareas
- onlypage_order== 0 allocations
- skip more complex vmalloc object types
- only reclaim tail resources when the retained waste is at leastPMD_SIZEThe current evidence supports this as a resource reclamation fix rather than a workload-tuned performance optimization. Local validation currently covers:
- synthetic large shrink correctness
- shrink-then-grow regression
- threshold boundary correctness for the current heuristic
- KASAN run-rootfsvmalloc_oobregression coverageQuoted 7 lines
I would especially appreciate feedback on:
1. whether this shrink direction is desirable upstream at all
2. whether the initial object-type restrictions are reasonable
3. whether a conservativePMD_SIZEthreshold is an acceptable first heuristic
4. what kind of in-tree regression test would be preferredCould you please have a look at this work:
https://lore.kernel.org/all/[email protected]/
Quoted 9 lines
Quoted 4 lines
Shivam is working on the same feature. Could you please check?
--
Uladzislau RezkiThanks for the pointer, and sorry for missing Shivam's series before sending this RFC.
I will read through it first and avoid duplicating the effort.
Thank you!
Maybe Shivam can also have a look at your work. I put him into To.
--
Uladzislau Rezki
Hey Fujunjie
In my patch series,
1. I am freeing at least a single page.
2. Did not change the vmap area, only changed the nr_pages at vm_struct and update the pages array.
Rest of the questions are addressed in my patch series.
Feel free to review or test my patch series and let me know if you have any suggestions.
Shivam