rust: arc: remove unused PhantomData
PATCH 1/1 rust: arc: remove unused PhantomData
There's no need for this. The type had the same form when it was first introduced, so it seems this was never necessary.
Fixed: 9dc043655003 ("rust: sync: add Arc for ref-counted allocations")
--- rust/kernel/sync/arc.rs | 2 -- 1 file changed, 2 deletions(-)
rust/kernel/sync/arc.rs +0 -2
@@ -127,7 +127,6 @@/// ``` pub struct Arc <T : ?Sized > { ptr:NonNull <ArcInner <T >>, - _p:PhantomData <ArcInner <T >>, }#[pin_data]
@@ -219,7 +218,6 @@ unsafe fn from_inner(inner: NonNull<ArcInner<T>>) -> Self {// INVARIANT: By the safety requirements, the invariants hold. Arc { ptr: inner, - _p:PhantomData , } }
---
base-commit: ae7851c29747fa3765ecb722fe722117a346f988
change-id: 20241104-simplify-arc-70c3574b5fac
Best regards,
Signature · 1 line
Tamir Duberstein <[email protected]>
Hi Tamir,
Quoted 1 line · Tamir Duberstein
There's no need for this. The type had the same form when it was first
I believe we need this PhantomData to inform drop chec [1] that the drop of Arc may result into the drop of an ArcInner<T>. Rust std Arc has the similar definition [2], you can find more information about PhantomData usage on drop checking at [3].
[1]: https://doc.rust-lang.org/nightly/std/ops/trait.Drop.html#drop-check
[2]: https://doc.rust-lang.org/src/alloc/sync.rs.html#245
[3]: https://doc.rust-lang.org/nightly/std/marker/struct.PhantomData.html#ownership-and-the-drop-check
Regards,
Boqun
Quoted 37 lines
introduced, so it seems this was never necessary.
Fixed:
9dc043655003("rust: sync: addArcfor ref-counted allocations")Signed-off-by Tamir Duberstein <[email protected]>--- rust/kernel/sync/arc.rs | 2 -- 1 file changed, 2 deletions(-)rust/kernel/sync/arc.rs +0
-3@@ -127,7 +127,6 @@/// ``` pub struct Arc <T : ?Sized > { ptr:NonNull <ArcInner <T >>, - _p:PhantomData <ArcInner <T >>, }#[pin_data] @@ -219,7 +218,6 @@ unsafe fn from_inner(inner: NonNull<ArcInner<T>>) -> Self {// INVARIANT: By the safety requirements, the invariants hold. Arc { ptr: inner, - _p:PhantomData , } } ---base-commit:
ae7851c29747fa3765ecb722fe722117a346f988
change-id: 20241104-simplify-arc-70c3574b5facBest regards,
Signature · 1 line
Tamir Duberstein <[email protected]>
Quoted 11 lines · Boqun Feng
Hi Tamir,
Quoted 1 line · Tamir Duberstein
There's no need for this. The type had the same form when it was first
I believe we need this
PhantomDatato inform drop chec [1] that the drop ofArcmay result into the drop of anArcInner<T>. Rust stdArchas the similar definition [2], you can find more information about PhantomData usage on drop checking at [3].
Quoted 6 lines
[1]: https://doc.rust-lang.org/nightly/std/ops/trait.Drop.html#drop-check
[2]: https://doc.rust-lang.org/src/alloc/sync.rs.html#245
[3]: https://doc.rust-lang.org/nightly/std/marker/struct.PhantomData.html#ownership-and-the-drop-checkRegards,
Boqun
Thanks for explaining! Withdrawn.
Quoted 37 lines
Quoted 37 lines
introduced, so it seems this was never necessary.
Fixed:
9dc043655003("rust: sync: addArcfor ref-counted allocations")Signed-off-by Tamir Duberstein <[email protected]>--- rust/kernel/sync/arc.rs | 2 -- 1 file changed, 2 deletions(-)rust/kernel/sync/arc.rs +0
-3@@ -127,7 +127,6 @@/// ``` pub struct Arc <T : ?Sized > { ptr:NonNull <ArcInner <T >>, - _p:PhantomData <ArcInner <T >>, }#[pin_data] @@ -219,7 +218,6 @@ unsafe fn from_inner(inner: NonNull<ArcInner<T>>) -> Self {// INVARIANT: By the safety requirements, the invariants hold. Arc { ptr: inner, - _p:PhantomData , } } ---base-commit:
ae7851c29747fa3765ecb722fe722117a346f988
change-id: 20241104-simplify-arc-70c3574b5facBest regards,
--
Tamir Duberstein <[email protected]>
Quoted 5 lines · Boqun Feng
I believe we need this
PhantomDatato inform drop chec [1] that the drop ofArcmay result into the drop of anArcInner<T>. Rust stdArchas the similar definition [2], you can find more information about PhantomData usage on drop checking at [3].
Hmm... But they use may_dangle in their Drop and we don't (and we have a Drop unlike something like Unique), no? Or am I confused? i.e. if I understand correctly, that would have been needed in the past, but not anymore.
In any case, a comment here clarifying would be good -- the standard library does that in some of its types, which is a good idea. At the very least, the name of the field should indicate why we have the marker. I will add that to the guidelines patch I have to send...
Thanks!
Cheers,
Miguel
Quoted 8 lines · Miguel Ojeda
Quoted 5 lines · Boqun Feng
I believe we need this
PhantomDatato inform drop chec [1] that the drop ofArcmay result into the drop of anArcInner<T>. Rust stdArchas the similar definition [2], you can find more information about PhantomData usage on drop checking at [3].
Quoted 4 lines
Hmm... But they use
may_danglein theirDropand we don't (and we have aDropunlike something likeUnique), no? Or am I confused? i.e. if I understand correctly, that would have been needed in the past, but not anymore.
Doing a bit of archaeology I found the reasoning for the presence of PhantomData in std's Arc[0]. The TL;DR is that the presence of a type parameter T implies "owns T", but Arc owns ArcInner<T> rather than T. Thus in order to get correct dropck behavior it is necessary to opt out of "owns T" using may_dangle and opt into "owns ArcInner<T>" using PhantomData.
Please check my understanding; I couldn't find detailed documentation of the interaction between may_dangle and PhantomData. If this sounds correct, should we add may_dangle to our dropck? compile-fail tests would be useful here.
Cheers,
Tamir
Alice Ryhl
maintainerQuoted 23 lines · Tamir Duberstein
Quoted 8 lines · Miguel Ojeda
Quoted 5 lines · Boqun Feng
I believe we need this
PhantomDatato inform drop chec [1] that the drop ofArcmay result into the drop of anArcInner<T>. Rust stdArchas the similar definition [2], you can find more information about PhantomData usage on drop checking at [3].Quoted 4 lines
Hmm... But they use
may_danglein theirDropand we don't (and we have aDropunlike something likeUnique), no? Or am I confused? i.e. if I understand correctly, that would have been needed in the past, but not anymore.Doing a bit of archaeology I found the reasoning for the presence of
PhantomDatain std'sArc[0]. The TL;DR is that the presence of a type parameterTimplies "owns T", butArcownsArcInner<T>rather thanT. Thus in order to get correct dropck behavior it is necessary to opt out of "owns T" usingmay_dangleand opt into "owns ArcInner<T>" usingPhantomData.
Quoted 4 lines
Please check my understanding; I couldn't find detailed documentation of the interaction between
may_dangleandPhantomData. If this sounds correct, should we addmay_dangleto our dropck? compile-fail tests would be useful here.
We don't *have* to use #[may_dangle]. Using it may allow more stuff, but it's not a problem for it to be missing. We probably don't want to use it since it's unstable.
Since we don't use #[may_dangle], we don't *need* the PhantomData field. Having it doesn't change anything.
Alice
Quoted 30 lines · Alice Ryhl
Quoted 23 lines · Tamir Duberstein
Quoted 8 lines · Miguel Ojeda
Quoted 5 lines · Boqun Feng
I believe we need this
PhantomDatato inform drop chec [1] that the drop ofArcmay result into the drop of anArcInner<T>. Rust stdArchas the similar definition [2], you can find more information about PhantomData usage on drop checking at [3].Quoted 4 lines
Hmm... But they use
may_danglein theirDropand we don't (and we have aDropunlike something likeUnique), no? Or am I confused? i.e. if I understand correctly, that would have been needed in the past, but not anymore.Doing a bit of archaeology I found the reasoning for the presence of
PhantomDatain std'sArc[0]. The TL;DR is that the presence of a type parameterTimplies "owns T", butArcownsArcInner<T>rather thanT. Thus in order to get correct dropck behavior it is necessary to opt out of "owns T" usingmay_dangleand opt into "owns ArcInner<T>" usingPhantomData.Quoted 4 lines
Please check my understanding; I couldn't find detailed documentation of the interaction between
may_dangleandPhantomData. If this sounds correct, should we addmay_dangleto our dropck? compile-fail tests would be useful here.
Quoted 6 lines
We don't *have* to use #[may_dangle]. Using it may allow more stuff, but it's not a problem for it to be missing. We probably don't want to use it since it's unstable.
Since we don't use #[may_dangle], we don't *need* the PhantomData field. Having it doesn't change anything.
In that case, should we reconsider this patch?
Quoted 2 lines · Tamir Duberstein
In that case, should we reconsider this patch?
Either that [*] or we could at least add a comment explaining it is not required for drop check purposes but that we have it anyway for clarity.
Starting to use may_dangle is a third option, but I agree we should avoid it unless we got at least an indication from upstream Rust that they want to stabilize it soon in that form (and probably only if we feel an actual need for it, since it is yet another unsafe use).
[*] Well, not this patch exactly -- the commit message should be fixed to explain things properly (and likely the "Fixes" tag removed) and it should also mention it double-checked the effect on variance and auto traits.
Cheers,
Miguel
Quoted 14 lines · Miguel Ojeda
Quoted 2 lines · Tamir Duberstein
In that case, should we reconsider this patch?
Either that [*] or we could at least add a comment explaining it is not required for drop check purposes but that we have it anyway for clarity.
Starting to use
may_dangleis a third option, but I agree we should avoid it unless we got at least an indication from upstream Rust that they want to stabilize it soon in that form (and probably only if we feel an actual need for it, since it is yet anotherunsafeuse).
Quoted 4 lines
[*] Well, not this patch exactly -- the commit message should be fixed to explain things properly (and likely the "Fixes" tag removed) and it should also mention it double-checked the effect on variance and auto traits.
The upstream changes to dropck predate the PR I linked up-thread which landed in 2017. Since this Arc code was written in 2022, it never had any effect. Isn't it proper to keep the "Fixes" tag?
Quoted 4 lines · Tamir Duberstein
The upstream changes to dropck predate the PR I linked up-thread which landed in 2017. Since this Arc code was written in 2022, it never had any effect. Isn't it proper to keep the "Fixes" tag?
If there is a bug, definitely yes, but I don't think that applies is -- this is more of a cleanup, no?
Sometimes things are marked as "fixes" that are perhaps a stretch (e.g. a typo in a comment). It depends a bit on the maintainer and how we define "bug" (e.g. does it count in docs, or just actual end users). But here it just seems something is superfluous, at worst, and it does not need to be backported either. Even if we kept the tag for some reason, I think this belongs in rust-next.
But if I am missing something, and this does indeed fix something that we should prioritize, please let me know!
What looks more important, to me, is to clarify/document why (or why not) we have it, regardless of whether we keep it or not, i.e. having thought about it, I think it wouldn't hurt having a line/comment even if we remove it.
Thanks!
Cheers,
Miguel
Quoted 10 lines · Miguel Ojeda
Quoted 4 lines · Tamir Duberstein
The upstream changes to dropck predate the PR I linked up-thread which landed in 2017. Since this Arc code was written in 2022, it never had any effect. Isn't it proper to keep the "Fixes" tag?
If there is a bug, definitely yes, but I don't think that applies is -- this is more of a cleanup, no?
Quoted 6 lines
Sometimes things are marked as "fixes" that are perhaps a stretch (e.g. a typo in a comment). It depends a bit on the maintainer and how we define "bug" (e.g. does it count in docs, or just actual end users). But here it just seems something is superfluous, at worst, and it does not need to be backported either. Even if we kept the tag for some reason, I think this belongs in
rust-next.
Thanks, I forgot that "Fixes" changes are backported. This makes sense!
Quoted 7 lines
But if I am missing something, and this does indeed fix something that we should prioritize, please let me know!
What looks more important, to me, is to clarify/document why (or why not) we have it, regardless of whether we keep it or not, i.e. having thought about it, I think it wouldn't hurt having a line/comment even if we remove it.
Will do.
Quoted 4 lines
Thanks!
Cheers,
Miguel
Quoted 2 lines · Tamir Duberstein
Thanks, I forgot that "Fixes" changes are backported. This makes sense!
Sometimes they are, yeah, but things that should go into stable should be marked explicitly nevertheless.
Cheers,
Miguel
Quoted 2 lines · Tamir Duberstein
Doing a bit of archaeology I found the reasoning for the presence of
What I meant by "in the past" is that, before the Rust RFC that changed it, the PhantomData would have been needed as far as I understand it, since having the Drop wouldn't have been enough.
Cheers,
Miguel
Quoted 8 lines · Miguel Ojeda
Quoted 2 lines · Tamir Duberstein
Doing a bit of archaeology I found the reasoning for the presence of
What I meant by "in the past" is that, before the Rust RFC that changed it, the
PhantomDatawould have been needed as far as I understand it, since having theDropwouldn't have been enough.
I would be happy to add the relevant details to the commit message but this is one citation that I haven't been able to locate. The closest mention I could find[0] only vaguely mentions that this change was made, but does not reference a commit (and certainly not an RFC).
Quoted 5 lines · Tamir Duberstein
I would be happy to add the relevant details to the commit message but this is one citation that I haven't been able to locate. The closest mention I could find[0] only vaguely mentions that this change was made, but does not reference a commit (and certainly not an RFC).
In Boqun's first link, there is a reference to the nomicon with details, and the section:
https://doc.rust-lang.org/nomicon/phantom-data.html#generic-parameters-and-drop-checking
explains the change, including:
"But ever since RFC 1238, this is no longer true nor necessary."
There was another RFC (1327) after that, for a finer-grained approach (may_dangle). The name of the feature gate was also changed.
Anyway, I don't think we need to add any of that to the commit message though. Perhaps linking the latest RFC is good for context, so if you think it is a good idea, of course please go for it -- but in case you are referring to what I said, I didn't say that we should add the RFC bits into the commit message.
Cheers,
Miguel
Quoted 20 lines · Miguel Ojeda
Quoted 5 lines · Tamir Duberstein
I would be happy to add the relevant details to the commit message but this is one citation that I haven't been able to locate. The closest mention I could find[0] only vaguely mentions that this change was made, but does not reference a commit (and certainly not an RFC).
In Boqun's first link, there is a reference to the nomicon with details, and the section:
https://doc.rust-lang.org/nomicon/phantom-data.html#generic-parameters-and-drop-checking
explains the change, including:
"But ever since RFC 1238, this is no longer true nor necessary."
There was another RFC (1327) after that, for a finer-grained approach (
may_dangle). The name of the feature gate was also changed.
Quoted 8 lines
Anyway, I don't think we need to add any of that to the commit message though. Perhaps linking the latest RFC is good for context, so if you think it is a good idea, of course please go for it -- but in case you are referring to what I said, I didn't say that we should add the RFC bits into the commit message.
Cheers,
Miguel
Thanks a lot for all the hand-holding here. I think the sensible thing to do here is to add a comment rather than remove the PhantomData. I'll send that as v2 shortly if no-one objects.
Alice Ryhl
maintainerQuoted 29 lines · Tamir Duberstein
Quoted 26 lines · Miguel Ojeda
Quoted 5 lines · Tamir Duberstein
I would be happy to add the relevant details to the commit message but this is one citation that I haven't been able to locate. The closest mention I could find[0] only vaguely mentions that this change was made, but does not reference a commit (and certainly not an RFC).
In Boqun's first link, there is a reference to the nomicon with details, and the section:
https://doc.rust-lang.org/nomicon/phantom-data.html#generic-parameters-and-drop-checking
explains the change, including:
"But ever since RFC 1238, this is no longer true nor necessary."
There was another RFC (1327) after that, for a finer-grained approach (
may_dangle). The name of the feature gate was also changed.Anyway, I don't think we need to add any of that to the commit message though. Perhaps linking the latest RFC is good for context, so if you think it is a good idea, of course please go for it -- but in case you are referring to what I said, I didn't say that we should add the RFC bits into the commit message.
Quoted 6 lines
Quoted 2 lines
Cheers,
MiguelThanks a lot for all the hand-holding here. I think the sensible thing to do here is to add a comment rather than remove the PhantomData. I'll send that as v2 shortly if no-one objects.
Adding a comment and keeping PhantomData SGTM.
Alice