mm/hmm: pass mmu_notifier_range to sync_cpu_device_pagetables
Subsystems (6)
RFC 1/1 mm/hmm: pass mmu_notifier_range to sync_cpu_device_pagetables
HMM defines its own struct hmm_update which is passed to the sync_cpu_device_pagetables() callback function. This is sufficient when the only action is to invalidate. However, a device may want to know the reason for the invalidation and be able to see the new permissions on a range, update device access rights or range statistics. Since sync_cpu_device_pagetables() can be called from try_to_unmap(), the mmap_sem may not be held and find_vma() is not safe to be called.
Pass the struct mmu_notifier_range to sync_cpu_device_pagetables() to allow the full invalidation information to be used.
---
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
drivers/gpu/drm/nouveau/nouveau_svm.c | 4 ++-- include/linux/hmm.h | 27 ++------------------------- mm/hmm.c | 13 ++++--------- 3 files changed, 8 insertions(+), 36 deletions(-)
drivers/gpu/drm/nouveau/nouveau_svm.c +2 -2
@@ -252,13 +252,13 @@ nouveau_svmm_invalidate(struct nouveau_svmm *svmm, u64 start, u64 limit)static int nouveau_svmm_sync_cpu_device_pagetables (struct hmm_mirror *mirror, -const struct hmm_update *update) +const struct mmu_notifier_range *update) {struct nouveau_svmm *svmm =container_of (mirror,typeof (*svmm), mirror);unsigned long start = update->start;unsigned long limit = update->end; -if (!update->blockable) +if (!mmu_notifier_range_blockable (update))return -EAGAIN ;SVMM_DBG (svmm,"invalidate %016lx-%016lx" , start, limit);
include/linux/hmm.h +2 -25
@@ -377,29 +377,6 @@ static inline uint64_t hmm_pfn_from_pfn(const struct hmm_range *range,struct hmm_mirror ; -/* -* enum hmm_update_event - type of update -* @HMM_UPDATE_INVALIDATE: invalidate range (no indication as to why) -*/ -enum hmm_update_event { -HMM_UPDATE_INVALIDATE , -}; - -/* -* struct hmm_update - HMM update information for callback -* -* @start: virtual start address of the range to update -* @end: virtual end address of the range to update -* @event: event triggering the update (what is happening) -* @blockable: can the callback block/sleep ? -*/ -struct hmm_update { -unsigned long start; -unsigned long end; -enum hmm_update_event event; -bool blockable; -}; -/* * struct hmm_mirror_ops - HMM mirror device operations callback *
@@ -420,7 +397,7 @@ struct hmm_mirror_ops {/* sync_cpu_device_pagetables() - synchronize page tables * * @mirror: pointer to struct hmm_mirror -* @update: update information (see struct hmm_update) +* @update: update information (see struct mmu_notifier_range) * Return: -EAGAIN if update.blockable false and callback need to * block, 0 otherwise. *
@@ -434,7 +411,7 @@ struct hmm_mirror_ops {* synchronous call. */ int (*sync_cpu_device_pagetables)(struct hmm_mirror *mirror, -const struct hmm_update *update); +const struct mmu_notifier_range *update); };/*
mm/hmm.c +4 -9
@@ -164,7 +164,6 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn, { struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);struct hmm_mirror *mirror; -struct hmm_update update;struct hmm_range *range;unsigned long flags;int ret =0 ;
@@ -173,15 +172,10 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn,if (!kref_get_unless_zero (&hmm->kref))return 0 ; - update.start = nrange->start; - update.end = nrange->end; - update.event =HMM_UPDATE_INVALIDATE ; - update.blockable =mmu_notifier_range_blockable (nrange); -spin_lock_irqsave (&hmm->ranges_lock, flags); hmm->notifiers++;list_for_each_entry (range, &hmm->ranges, list) { -if (update.end < range->start || update.start >= range->end) +if (nrange->end < range->start || nrange->start >= range->end)continue ; range->valid = false;
@@ -198,9 +192,10 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn,list_for_each_entry (mirror, &hmm->mirrors, list) {int rc; - rc = mirror->ops->sync_cpu_device_pagetables (mirror, &update); + rc = mirror->ops->sync_cpu_device_pagetables (mirror, nrange);if (rc) { -if (WARN_ON (update.blockable || rc != -EAGAIN )) +if (WARN_ON (mmu_notifier_range_blockable (nrange) || + rc != -EAGAIN ))continue ; ret = -EAGAIN ;break ;
Signature · 6 lines
2.20.1 _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 11 lines · Ralph Campbell
HMM defines its own struct
hmm_updatewhich is passed to thesync_cpu_device_pagetables()callback function. This is sufficient when the only action is to invalidate. However, a device may want to know the reason for the invalidation and be able to see the new permissions on a range, update device access rights or range statistics. Sincesync_cpu_device_pagetables()can be called fromtry_to_unmap(), themmap_semmay not be held andfind_vma()is not safe to be called.
Pass the structmmu_notifier_rangetosync_cpu_device_pagetables()to allow the full invalidation information to be used.
Quoted 5 lines
Signed-off-by Ralph Campbell <[email protected]>---
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
This is the right thing to do. But the really right thing is to just kill the hmm_mirror API entirely and move to mmu_notifiers. At least for noveau this already is way simpler, although right now it defeats Jasons patch to avoid allocating the struct hmm in the fault path. But as said before that can be avoided by just killing struct hmm, which for many reasons is the right thing to do anyway.
I've got a series here, which is a bit broken (epecially the last patch can't work as-is), but should explain where I'm trying to head:
http://git.infradead.org/users/hch/misc.git/shortlog/refs/heads/hmm-mirror-simplification
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 24 lines · Christoph Hellwig
Quoted 11 lines · Ralph Campbell
HMM defines its own struct
hmm_updatewhich is passed to thesync_cpu_device_pagetables()callback function. This is sufficient when the only action is to invalidate. However, a device may want to know the reason for the invalidation and be able to see the new permissions on a range, update device access rights or range statistics. Sincesync_cpu_device_pagetables()can be called fromtry_to_unmap(), themmap_semmay not be held andfind_vma()is not safe to be called.
Pass the structmmu_notifier_rangetosync_cpu_device_pagetables()to allow the full invalidation information to be used.Quoted 4 lines
Signed-off-by Ralph Campbell <[email protected]>I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
This is the right thing to do. But the really right thing is to just kill the
hmm_mirrorAPI entirely and move tommu_notifiers. At least for noveau this already is way simpler, although right now it defeats Jasons patch to avoid allocating the struct hmm in the fault path. But as said before that can be avoided by just killing struct hmm, which for many reasons is the right thing to do anyway.
Quoted 4 lines
I've got a series here, which is a bit broken (epecially the last patch can't work as-is), but should explain where I'm trying to head:
http://git.infradead.org/users/hch/misc.git/shortlog/refs/heads/hmm-mirror-simplification
At least the current hmm approach does rely on the collision retry locking scheme in struct hmm/struct hmm_range for the pagefault side to work right.
So, before we can apply patch one in this series we need to fix hmm_vma_fault() and all its varients. Otherwise the driver will be broken.
I'm hoping to first define what this locking should be (see other emails to Ralph) then, ideally, see if we can extend mmu notifiers to get it directly withouth hmm stuff.
Then we apply your patch one and the hmm ops wrapper dies.
Jason
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 38 lines · Jason Gunthorpe
Quoted 24 lines · Christoph Hellwig
Quoted 11 lines · Ralph Campbell
HMM defines its own struct
hmm_updatewhich is passed to thesync_cpu_device_pagetables()callback function. This is sufficient when the only action is to invalidate. However, a device may want to know the reason for the invalidation and be able to see the new permissions on a range, update device access rights or range statistics. Sincesync_cpu_device_pagetables()can be called fromtry_to_unmap(), themmap_semmay not be held andfind_vma()is not safe to be called.
Pass the structmmu_notifier_rangetosync_cpu_device_pagetables()to allow the full invalidation information to be used.Quoted 4 lines
Signed-off-by Ralph Campbell <[email protected]>I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
This is the right thing to do. But the really right thing is to just kill the
hmm_mirrorAPI entirely and move tommu_notifiers. At least for noveau this already is way simpler, although right now it defeats Jasons patch to avoid allocating the struct hmm in the fault path. But as said before that can be avoided by just killing struct hmm, which for many reasons is the right thing to do anyway.Quoted 4 lines
I've got a series here, which is a bit broken (epecially the last patch can't work as-is), but should explain where I'm trying to head:
http://git.infradead.org/users/hch/misc.git/shortlog/refs/heads/hmm-mirror-simplification
At least the current hmm approach does rely on the collision retry locking scheme in struct hmm/struct
hmm_rangefor the pagefault side to work right.So, before we can apply patch one in this series we need to fix
hmm_vma_fault()and all its varients. Otherwise the driver will be broken.
Quoted 6 lines
I'm hoping to first define what this locking should be (see other emails to Ralph) then, ideally, see if we can extend mmu notifiers to get it directly withouth hmm stuff.
Then we apply your patch one and the hmm ops wrapper dies.
This all makes sense, and thanks for all this work to simplify and clarify HMM. It's going to make it a lot easier to work with, when the dust settles.
thanks,
Signature · 6 lines
John Hubbard NVIDIA _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 11 lines · Ralph Campbell
HMM defines its own struct
hmm_updatewhich is passed to thesync_cpu_device_pagetables()callback function. This is sufficient when the only action is to invalidate. However, a device may want to know the reason for the invalidation and be able to see the new permissions on a range, update device access rights or range statistics. Sincesync_cpu_device_pagetables()can be called fromtry_to_unmap(), themmap_semmay not be held andfind_vma()is not safe to be called.
Pass the structmmu_notifier_rangetosync_cpu_device_pagetables()to allow the full invalidation information to be used.
Quoted 5 lines
Signed-off-by Ralph Campbell <[email protected]>---
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
I agree with CH that struct hmm_update seems particularly pointless and we really should just use mmu_notifier_range directly.
We need to find out from the DRM folks if we can merge this kind of stuff through hmm.git and then resolve any conflicts that might arise in DRM tree or in nouveau tree?
But I would like to see this patch go in this cycle, thanks
Jason
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 11 lines · Ralph Campbell
HMM defines its own struct
hmm_updatewhich is passed to thesync_cpu_device_pagetables()callback function. This is sufficient when the only action is to invalidate. However, a device may want to know the reason for the invalidation and be able to see the new permissions on a range, update device access rights or range statistics. Sincesync_cpu_device_pagetables()can be called fromtry_to_unmap(), themmap_semmay not be held andfind_vma()is not safe to be called.
Pass the structmmu_notifier_rangetosync_cpu_device_pagetables()to allow the full invalidation information to be used.
Quoted 1 line
Signed-off-by Ralph Campbell <[email protected]>
I don't disagree with Christoph or Jason but since I've been trying to sort out where hmm does and does not fit any chance to remove a custom structure is a good simplification IMO. So...
Quoted 128 lines
---
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
drivers/gpu/drm/nouveau/nouveau_svm.c | 4 ++-- include/linux/hmm.h | 27 ++------------------------- mm/hmm.c | 13 ++++--------- 3 files changed, 8 insertions(+), 36 deletions(-)drivers/gpu/drm/nouveau/nouveau_svm.c +2
-2@@ -252,13 +252,13 @@ nouveau_svmm_invalidate(struct nouveau_svmm *svmm, u64 start, u64 limit)static int nouveau_svmm_sync_cpu_device_pagetables (struct hmm_mirror *mirror, -const struct hmm_update *update) +const struct mmu_notifier_range *update) {struct nouveau_svmm *svmm =container_of (mirror,typeof (*svmm), mirror);unsigned long start = update->start;unsigned long limit = update->end; -if (!update->blockable) +if (!mmu_notifier_range_blockable (update))return -EAGAIN ;SVMM_DBG (svmm,"invalidate %016lx-%016lx" , start, limit);include/linux/hmm.h +2
-25@@ -377,29 +377,6 @@ static inline uint64_t hmm_pfn_from_pfn(const struct hmm_range *range,struct hmm_mirror ; -/* -* enum hmm_update_event - type of update -* @HMM_UPDATE_INVALIDATE: invalidate range (no indication as to why) -*/ -enum hmm_update_event { -HMM_UPDATE_INVALIDATE , -}; - -/* -* struct hmm_update - HMM update information for callback -* -* @start: virtual start address of the range to update -* @end: virtual end address of the range to update -* @event: event triggering the update (what is happening) -* @blockable: can the callback block/sleep ? -*/ -struct hmm_update { -unsigned long start; -unsigned long end; -enum hmm_update_event event; -bool blockable; -}; -/* * struct hmm_mirror_ops - HMM mirror device operations callback * @@ -420,7 +397,7 @@ struct hmm_mirror_ops {/* sync_cpu_device_pagetables() - synchronize page tables * * @mirror: pointer to struct hmm_mirror -* @update: update information (see struct hmm_update) +* @update: update information (see struct mmu_notifier_range) * Return: -EAGAIN if update.blockable false and callback need to * block, 0 otherwise. * @@ -434,7 +411,7 @@ struct hmm_mirror_ops {* synchronous call. */ int (*sync_cpu_device_pagetables)(struct hmm_mirror *mirror, -const struct hmm_update *update); +const struct mmu_notifier_range *update); };/* mm/hmm.c +4
-9@@ -164,7 +164,6 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn, { struct hmm *hmm = container_of(mn, struct hmm, mmu_notifier);struct hmm_mirror *mirror; -struct hmm_update update;struct hmm_range *range;unsigned long flags;int ret =0 ;@@ -173,15 +172,10 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn,if (!kref_get_unless_zero (&hmm->kref))return 0 ; - update.start = nrange->start; - update.end = nrange->end; - update.event =HMM_UPDATE_INVALIDATE ; - update.blockable =mmu_notifier_range_blockable (nrange); -spin_lock_irqsave (&hmm->ranges_lock, flags); hmm->notifiers++;list_for_each_entry (range, &hmm->ranges, list) { -if (update.end < range->start || update.start >= range->end) +if (nrange->end < range->start || nrange->start >= range->end)continue ; range->valid = false;@@ -198,9 +192,10 @@ static int hmm_invalidate_range_start(struct mmu_notifier *mn,list_for_each_entry (mirror, &hmm->mirrors, list) {int rc; - rc = mirror->ops->sync_cpu_device_pagetables (mirror, &update); + rc = mirror->ops->sync_cpu_device_pagetables (mirror, nrange);if (rc) { -if (WARN_ON (update.blockable || rc != -EAGAIN )) +if (WARN_ON (mmu_notifier_range_blockable (nrange) || + rc != -EAGAIN ))continue ; ret = -EAGAIN ;break ;Signature · 1 line
2.20.1
Quoted 11 lines · Ralph Campbell
HMM defines its own struct
hmm_updatewhich is passed to thesync_cpu_device_pagetables()callback function. This is sufficient when the only action is to invalidate. However, a device may want to know the reason for the invalidation and be able to see the new permissions on a range, update device access rights or range statistics. Sincesync_cpu_device_pagetables()can be called fromtry_to_unmap(), themmap_semmay not be held andfind_vma()is not safe to be called.
Pass the structmmu_notifier_rangetosync_cpu_device_pagetables()to allow the full invalidation information to be used.
Quoted 5 lines
Signed-off-by Ralph Campbell <[email protected]>---
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
This make so much sense, I'd like to apply this in hmm.git, is there any objection?
Jason
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 15 lines · Jason Gunthorpe
Quoted 14 lines · Ralph Campbell
HMM defines its own struct
hmm_updatewhich is passed to thesync_cpu_device_pagetables()callback function. This is sufficient when the only action is to invalidate. However, a device may want to know the reason for the invalidation and be able to see the new permissions on a range, update device access rights or range statistics. Sincesync_cpu_device_pagetables()can be called fromtry_to_unmap(), themmap_semmay not be held andfind_vma()is not safe to be called.
Pass the structmmu_notifier_rangetosync_cpu_device_pagetables()to allow the full invalidation information to be used.Signed-off-by Ralph Campbell <[email protected]>---
Quoted 8 lines
Quoted 2 lines
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
This make so much sense, I'd like to apply this in hmm.git, is there any objection?
Jason
Not from me. :-)
Thanks!
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 5 lines · Jason Gunthorpe
Quoted 2 lines
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
This make so much sense, I'd like to apply this in hmm.git, is there any objection?
As this creates a somewhat hairy conflict for amdgpu, wouldn't it be a better idea to wait a bit and apply it first thing for next merge window?
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 10 lines · Christoph Hellwig
Quoted 5 lines · Jason Gunthorpe
Quoted 2 lines
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
This make so much sense, I'd like to apply this in hmm.git, is there any objection?
As this creates a somewhat hairy conflict for amdgpu, wouldn't it be a better idea to wait a bit and apply it first thing for next merge window?
My thinking is that AMD GPU already has a monster conflict from this:
int hmm_range_register(struct hmm_range *range,
- struct mm_struct *mm,
+ struct hmm_mirror *mirror,
unsigned long start,
unsigned long end,
unsigned page_shift);
So, depending on how that is resolved we might want to do both API changes at once.
Or we may have to revert the above change at this late date.
Waiting for AMDGPU team to discuss what process they want to use.
Jason
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Review
Quoted 12 lines · Jason Gunthorpe
Quoted 3 lines
As this creates a somewhat hairy conflict for amdgpu, wouldn't it be a better idea to wait a bit and apply it first thing for next merge window?
My thinking is that AMD GPU already has a monster conflict from this:
int hmm_range_register(struct
hmm_range*range,(quoted diff) +1
-1- struct mm_struct *mm, + struct hmm_mirror *mirror, unsigned long start, unsigned long end, unsigned page_shift);
Well, that seems like a relatively easy to fix conflict, at least as long as you have the mirror easily available. The notifier change on the other hand basically requires rewriting about two dozen lines of code entirely.
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 18 lines · Jason Gunthorpe
Quoted 8 lines · Christoph Hellwig
Quoted 4 lines · Jason Gunthorpe
Quoted 2 lines
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
This make so much sense, I'd like to apply this in hmm.git, is there any objection?
As this creates a somewhat hairy conflict for amdgpu, wouldn't it be a better idea to wait a bit and apply it first thing for next merge window?
My thinking is that AMD GPU already has a monster conflict from this:
int hmm_range_register(struct
hmm_range*range,(quoted diff) +1
-1- struct mm_struct *mm, + struct hmm_mirror *mirror, unsigned long start, unsigned long end, unsigned page_shift);
Quoted 2 lines
So, depending on how that is resolved we might want to do both API changes at once.
I just sent out a fix for the hmm_mirror API change.
Quoted 4 lines
Or we may have to revert the above change at this late date.
Waiting for AMDGPU team to discuss what process they want to use.
Yeah, I'm wondering what the process is myself. With HMM and driver development happening on different branches these kinds of API changes are painful. There seems to be a built-in assumption in the current process, that code flows mostly in one direction amd-staging-drm-next -> drm-next -> linux-next -> linux. That assumption is broken with HMM code evolving rapidly in both amdgpu and mm.
If we want to continue developing HMM driver changes in amd-staging-drm-next, we'll need to synchronize with hmm.git more frequently, both ways. I believe part of the problem is, that there is a fairly long lead-time from getting changes from amd-staging-drm-next into linux-next, as they are held for one release cycle in drm-next. Pushing HMM-related changes through drm-fixes may offer a kind of shortcut. Philip and my latest fixup is just bypassing drm-next completely and going straight into linux-next, though.
Regards,
Felix
Quoted 2 lines
Jason
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 19 lines · Felix
Quoted 18 lines · Jason Gunthorpe
Quoted 8 lines · Christoph Hellwig
Quoted 4 lines · Jason Gunthorpe
Quoted 2 lines
I'm sending this out now since we are updating many of the HMM APIs and I think it will be useful.
This make so much sense, I'd like to apply this in hmm.git, is there any objection?
As this creates a somewhat hairy conflict for amdgpu, wouldn't it be a better idea to wait a bit and apply it first thing for next merge window?
My thinking is that AMD GPU already has a monster conflict from this:
int hmm_range_register(struct
hmm_range*range,(quoted diff) +1
-1- struct mm_struct *mm, + struct hmm_mirror *mirror, unsigned long start, unsigned long end, unsigned page_shift);
Quoted 4 lines
Quoted 2 lines
So, depending on how that is resolved we might want to do both API changes at once.
I just sent out a fix for the
hmm_mirrorAPI change.
I think if you follow my suggestion to apply a prep patch to AMD GPU to make the conflict resolution simple, we should defer this patch until next kernel for the reasons CH gave.
Quoted 10 lines
Quoted 3 lines
Or we may have to revert the above change at this late date.
Waiting for AMDGPU team to discuss what process they want to use.
Yeah, I'm wondering what the process is myself. With HMM and driver development happening on different branches these kinds of API changes are painful. There seems to be a built-in assumption in the current process, that code flows mostly in one direction amd-staging-drm-next -> drm-next -> linux-next -> linux. That assumption is broken with HMM code evolving rapidly in both amdgpu and mm.
It looks to me like AMD GPU uses a pull request model. So a goal as a tree runner should be to work with the other trees (ie hmm.git, etc) to minimize conflicts between the PR you will send and the PR other trees will send.
Do not focus on linux-next, that is just an 'early warning system' that conflicts are on the horizon, we knew about this one :) (well, mostly, I was surprised how big it was, my bad)
So we must stay in co-ordination with patches in-flight on the list and make the right decision, depending on the situation. Communication here is key :)
We have lots of strategies available to deal with these situations.
Quoted 3 lines
If we want to continue developing HMM driver changes in amd-staging-drm-next, we'll need to synchronize with hmm.git more frequently, both ways.
It can't really go both ways. hmm.git has to be only the hmm topic, otherwise it doesn't really work.
Quoted 6 lines
I believe part of the problem is, that there is a fairly long lead-time from getting changes from amd-staging-drm-next into linux-next, as they are held for one release cycle in drm-next. Pushing HMM-related changes through drm-fixes may offer a kind of shortcut. Philip and my latest fixup is just bypassing drm-next completely and going straight into linux-next, though.
I'm not so familiar with the DRM work flow to give you advice on this.
Jason
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx