PAT and cache coherency support
v9 cover PAT and cache coherency support
[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Reply-To:
In-Reply-To: <[email protected]>
Quoted 39 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_type=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_indexwhen mapping memory via the ppGTT, in addtion to the CPU caching mode. This is very much needed on newer igpu platforms which allow incoherent GT access, where the choice over the cache level and expected coherency is best left to userspace depending on their usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU cache mode to use for system memory, including the expected GPU coherency mode. There are various restrictions here for the selected coherency mode and compatible CPU cache modes. With that in place the actualpat_indexcan now be provided as part ofvm_bind. The only restriction is that the coherency mode of thepat_indexmust be at least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.
Quoted 12 lines
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Signature · 1 line
2.41.0
I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Just to confirm my understanding. We will have 2 separate APIs?
1. For gem_create to explicitly set coherency/caching according to API enums
2. For vm_bind to set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?
Quoted 59 lines · Tomasz Mistat
[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Reply-To:
In-Reply-To: <[email protected]>Quoted 39 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_type=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_indexwhen mapping memory via the ppGTT, in addtion to the CPU caching mode. This is very much needed on newer igpu platforms which allow incoherent GT access, where the choice over the cache level and expected coherency is best left to userspace depending on their usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU cache mode to use for system memory, including the expected GPU coherency mode. There are various restrictions here for the selected coherency mode and compatible CPU cache modes. With that in place the actualpat_indexcan now be provided as part ofvm_bind. The only restriction is that the coherency mode of thepat_indexmust be at least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.Quoted 12 lines
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Signature · 1 line
2.41.0
Quoted 8 lines
I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportJust to confirm my understanding. We will have 2 separate APIs?
1. Forgem_createto explicitly set coherency/caching according to API enums
Yeah, you just need to provide the cpu_caching mode (wc or wb). If you later mmap() the object this same caching mode is used. Also any KMD internal things like swap-in will also respect this caching mode. The coh_mode is now removed.
Quoted 1 line
2. For
vm_bindto set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?
Right, as part of vm_bind you provide whatever PAT index control (as per Bspec) you want for each mapping of the object (you could map it multiple times).
When choosing a PAT index the cpu_caching only matters if you choose an incoherent PAT index (coh_none in Bspec), since we don't want to allow cpu_caching=wb and coh_none, since that leads to some complications in the KMD.
Quoted 2 lines
Hi Matthew,
In the compute driver we call GEM_MMAP_OFFSET with flags to specify cpu cache mode (wb/wc):GEM_CREATE_EXT() and mmap() + GEM_MMAP_OFFSET(flags=wb/wc)
Starting from Xe KMD we should be able to achieve the same CPU caching policies with GEM_CREATE_EXT:GEM_CREATE_EXT(flags=wb/wc) and mmap() + GEM_MMAP_OFFSET(flags=0)
Am I correct?
Thanks,
Slawek
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Wednesday, November 8, 2023 1:04 PM
To: Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Milczarek, Slawomir <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>; Dunajski, Bartosz <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Quoted 64 lines · Tomasz Mistat
[email protected],[email protected],saur [email protected],[email protected],[email protected] ,[email protected],[email protected],ayaz.siddiqui@intel .com,[email protected],[email protected],filip.hazubski@int el.com,[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support Reply-To:
In-Reply-To: <[email protected]>Quoted 49 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_typ
e=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_index
when mapping memory via the ppGTT, in addtion to the CPU caching
mode. This is very much needed on newer igpu platforms which allow
incoherent GT access, where the choice over the cache level and
expected coherency is best left to userspace depending on their
usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU
cache mode to use for system memory, including the expected GPU
coherency mode. There are various restrictions here for the selected
coherency mode and compatible CPU cache modes. With that in place
the actualpat_indexcan now be provided as part ofvm_bind. The only
restriction is that the coherency mode of thepat_indexmust be at
least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Quoted 3 lines
--
2.41.0
Quoted 8 lines
I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportJust to confirm my understanding. We will have 2 separate APIs?
1. Forgem_createto explicitly set coherency/caching according to API enums
Yeah, you just need to provide the cpu_caching mode (wc or wb). If you later mmap() the object this same caching mode is used. Also any KMD internal things like swap-in will also respect this caching mode. The coh_mode is now removed.
Quoted 1 line
2. For
vm_bindto set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?
Right, as part of vm_bind you provide whatever PAT index control (as per Bspec) you want for each mapping of the object (you could map it multiple times).
When choosing a PAT index the cpu_caching only matters if you choose an incoherent PAT index (coh_none in Bspec), since we don't want to allow cpu_caching=wb and coh_none, since that leads to some complications in the KMD.
Quoted 2 lines
---------------------------------------------------------------------
Intel Technology Poland sp. z o.o.
ul. Slowackiego 173 | 80-298 Gdansk | Sad Rejonowy Gdansk Polnoc | VII Wydzial Gospodarczy Krajowego Rejestru Sadowego - KRS 101882 | NIP 957-07-52-316 | Kapital zakladowy 200.000 PLN. Spolka oswiadcza, ze posiada status duzego przedsiebiorcy w rozumieniu ustawy z dnia 8 marca 2013 r. o przeciwdzialaniu nadmiernym opoznieniom w transakcjach handlowych.
Ta wiadomosc wraz z zalacznikami jest przeznaczona dla okreslonego adresata i moze zawierac informacje poufne. W razie przypadkowego otrzymania tej wiadomosci, prosimy o powiadomienie nadawcy oraz trwale jej usuniecie; jakiekolwiek przegladanie lub rozpowszechnianie jest zabronione. This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). If you are not the intended recipient, please contact the sender and delete all copies; any review or distribution by others is strictly prohibited.
Quoted 8 lines · Slawomir
Hi Matthew,
In the compute driver we call
GEM_MMAP_OFFSETwith flags to specify cpu cache mode (wb/wc):GEM_CREATE_EXT()andmmap()+GEM_MMAP_OFFSET(flags=wb/wc)Starting from Xe KMD we should be able to achieve the same CPU caching policies with
GEM_CREATE_EXT:GEM_CREATE_EXT(flags=wb/wc)andmmap()+GEM_MMAP_OFFSET(flags=0)
Am I correct?
Yes, that's correct.
Quoted 95 lines
Thanks,
Slawek-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Wednesday, November 8, 2023 1:04 PM
To: Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Milczarek, Slawomir <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>; Dunajski, Bartosz <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 64 lines · Tomasz Mistat
[email protected],[email protected],saur [email protected],[email protected],[email protected] ,[email protected],[email protected],ayaz.siddiqui@intel .com,[email protected],[email protected],filip.hazubski@int el.com,[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support Reply-To:
In-Reply-To: <[email protected]>Quoted 49 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_typ
e=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_index
when mapping memory via the ppGTT, in addtion to the CPU caching
mode. This is very much needed on newer igpu platforms which allow
incoherent GT access, where the choice over the cache level and
expected coherency is best left to userspace depending on their
usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU
cache mode to use for system memory, including the expected GPU
coherency mode. There are various restrictions here for the selected
coherency mode and compatible CPU cache modes. With that in place
the actualpat_indexcan now be provided as part ofvm_bind. The only
restriction is that the coherency mode of thepat_indexmust be at
least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Quoted 3 lines
--
2.41.0Quoted 8 lines
I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportJust to confirm my understanding. We will have 2 separate APIs?
1. Forgem_createto explicitly set coherency/caching according to API enumsYeah, you just need to provide the
cpu_cachingmode (wc or wb). If you latermmap()the object this same caching mode is used. Also any KMD internal things like swap-in will also respect this caching mode. Thecoh_modeis now removed.Quoted 1 line
2. For
vm_bindto set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?Right, as part of
vm_bindyou provide whatever PAT index control (as per Bspec) you want for each mapping of the object (you could map it multiple times).When choosing a PAT index the
cpu_cachingonly matters if you choose an incoherent PAT index (coh_nonein Bspec), since we don't want to allow cpu_caching=wb andcoh_none, since that leads to some complications in the KMD.Quoted 2 lines
Will KMD fail the mmap call if we provide incorrect mapping ? I.e. WB mmap on non-coherent GPU resource?
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 4:05 PM
To: Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Quoted 8 lines · Slawomir
Hi Matthew,
In the compute driver we call
GEM_MMAP_OFFSETwith flags to specify cpu cache mode (wb/wc):GEM_CREATE_EXT()andmmap()+GEM_MMAP_OFFSET(flags=wb/wc)Starting from Xe KMD we should be able to achieve the same CPU caching policies with
GEM_CREATE_EXT:GEM_CREATE_EXT(flags=wb/wc)andmmap()+GEM_MMAP_OFFSET(flags=0)Am I
correct?
Yes, that's correct.
Quoted 120 lines
Thanks,
Slawek-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Wednesday, November 8, 2023 1:04 PM
To: Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Milczarek, Slawomir <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>; Dunajski, Bartosz <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 72 lines · Tomasz Mistat
[email protected],[email protected],sau r [email protected],[email protected],[email protected] m ,[email protected],[email protected],ayaz.siddiqui@inte l .com,[email protected],[email protected],filip.hazubski@in t el.com,[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Reply-To:
In-Reply-To: <[email protected]>Quoted 50 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_ty
p
e=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_index
when mapping memory via the ppGTT, in addtion to the CPU caching
mode. This is very much needed on newer igpu platforms which allow
incoherent GT access, where the choice over the cache level and
expected coherency is best left to userspace depending on their
usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU
cache mode to use for system memory, including the expected GPU
coherency mode. There are various restrictions here for the selected
coherency mode and compatible CPU cache modes. With that in place
the actualpat_indexcan now be provided as part ofvm_bind. The
only restriction is that the coherency mode of thepat_indexmust be
at least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Quoted 3 lines
--
2.41.0I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Quoted 6 lines
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportJust to confirm my understanding. We will have 2 separate APIs?
1. Forgem_createto explicitly set coherency/caching according to API enumsYeah, you just need to provide the
cpu_cachingmode (wc or wb). If you latermmap()the object this same caching mode is used. Also any KMD internal things like swap-in will also respect this caching mode. Thecoh_modeis now removed.Quoted 1 line
2. For
vm_bindto set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?Right, as part of
vm_bindyou provide whatever PAT index control (as per
Bspec) you want for each mapping of the object (you could map it multiple times).When choosing a PAT index the
cpu_cachingonly matters if you choose an incoherent PAT index (coh_nonein Bspec), since we don't want to allow cpu_caching=wb andcoh_none, since that leads to some complications in the KMD.Quoted 2 lines
Quoted 1 line · Michal
Will KMD fail the mmap call if we provide incorrect mapping ?
With Xe there are no mmap flags or anything like that, so you can't really provide an incorrect mapping. Ignoring this series, the previous behaviour was for the KMD to implicitly select the cpu caching mode when allocating the memory, where roughly you would get WC for VRAM and then WB for system memory, unless it's used for scanout in which case you get WC. With this series userspace must now explicitly request the cpu caching mode at allocation via gem_create, which will be used for all CPU access, including KMD internal stuff as well as userspace mmap().
Quoted 1 line
I.e. WB mmap on non-coherent GPU resource?
Right, if you allocate WB memory (cpu_caching=wb), then you can't currently use an incoherent PAT index if accessing it from the GPU. KMD will reject it at vm_bind.
Quoted 141 lines
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 4:05 PM
To: Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 8 lines · Slawomir
Hi Matthew,
In the compute driver we call
GEM_MMAP_OFFSETwith flags to specify cpu cache mode (wb/wc):GEM_CREATE_EXT()andmmap()+GEM_MMAP_OFFSET(flags=wb/wc)Starting from Xe KMD we should be able to achieve the same CPU caching policies with
GEM_CREATE_EXT:GEM_CREATE_EXT(flags=wb/wc)andmmap()+GEM_MMAP_OFFSET(flags=0)Am I
correct?Yes, that's correct.
Quoted 120 lines
Thanks,
Slawek-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Wednesday, November 8, 2023 1:04 PM
To: Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Milczarek, Slawomir <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>; Dunajski, Bartosz <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 72 lines · Tomasz Mistat
[email protected],[email protected],sau r [email protected],[email protected],[email protected] m ,[email protected],[email protected],ayaz.siddiqui@inte l .com,[email protected],[email protected],filip.hazubski@in t el.com,[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Reply-To:
In-Reply-To: <[email protected]>Quoted 50 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_ty
p
e=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_index
when mapping memory via the ppGTT, in addtion to the CPU caching
mode. This is very much needed on newer igpu platforms which allow
incoherent GT access, where the choice over the cache level and
expected coherency is best left to userspace depending on their
usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU
cache mode to use for system memory, including the expected GPU
coherency mode. There are various restrictions here for the selected
coherency mode and compatible CPU cache modes. With that in place
the actualpat_indexcan now be provided as part ofvm_bind. The
only restriction is that the coherency mode of thepat_indexmust be
at least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Quoted 3 lines
--
2.41.0I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Quoted 6 lines
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportJust to confirm my understanding. We will have 2 separate APIs?
1. Forgem_createto explicitly set coherency/caching according to API enumsYeah, you just need to provide the
cpu_cachingmode (wc or wb). If you latermmap()the object this same caching mode is used. Also any KMD internal things like swap-in will also respect this caching mode. Thecoh_modeis now removed.Quoted 1 line
2. For
vm_bindto set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?Right, as part of
vm_bindyou provide whatever PAT index control (as per
Bspec) you want for each mapping of the object (you could map it multiple times).When choosing a PAT index the
cpu_cachingonly matters if you choose an incoherent PAT index (coh_nonein Bspec), since we don't want to allow cpu_caching=wb andcoh_none, since that leads to some complications in the KMD.Quoted 2 lines
"
Right, if you allocate WB memory (cpu_caching=wb), then you can't currently use an incoherent PAT index if accessing it from the GPU. KMD will reject it at vm_bind." "
What about scenarios where I want to have:
- CPU WB caching
- GPU non coherent access
And manually flush cache lines on synchronization points (i.e. clFlush on mapped OpenCL buffers) ? How to do it if KMD would reject VM_BIND operations?
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 5:08 PM
To: Mrozek, Michal <[email protected]>; Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Quoted 1 line · Michal
Will KMD fail the mmap call if we provide incorrect mapping ?
With Xe there are no mmap flags or anything like that, so you can't really provide an incorrect mapping. Ignoring this series, the previous behaviour was for the KMD to implicitly select the cpu caching mode when allocating the memory, where roughly you would get WC for VRAM and then WB for system memory, unless it's used for scanout in which case you get WC. With this series userspace must now explicitly request the cpu caching mode at allocation via gem_create, which will be used for all CPU access, including KMD internal stuff as well as userspace mmap().
Quoted 1 line
I.e. WB mmap on non-coherent GPU resource?
Right, if you allocate WB memory (cpu_caching=wb), then you can't currently use an incoherent PAT index if accessing it from the GPU. KMD will reject it at vm_bind.
Quoted 161 lines
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 4:05 PM
To: Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 8 lines · Slawomir
Hi Matthew,
In the compute driver we call
GEM_MMAP_OFFSETwith flags to specify cpu cache mode (wb/wc):GEM_CREATE_EXT()andmmap()+GEM_MMAP_OFFSET(flags=wb/wc)Starting from Xe KMD we should be able to achieve the same CPU caching policies with
GEM_CREATE_EXT:GEM_CREATE_EXT(flags=wb/wc)andmmap()+GEM_MMAP_OFFSET(flags=0)Am
I correct?Yes, that's correct.
Quoted 124 lines
Thanks,
Slawek-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Wednesday, November 8, 2023 1:04 PM
To: Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Milczarek, Slawomir <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>; Dunajski, Bartosz <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 76 lines · Tomasz Mistat
[email protected],[email protected],sa u
r [email protected],[email protected],[email protected] o
m ,[email protected],[email protected],ayaz.siddiqui@int e
l .com,[email protected],[email protected],filip.hazubski@i n t el.com,[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Reply-To:
In-Reply-To: <[email protected]>Quoted 51 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_t
y
p
e=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_index
when mapping memory via the ppGTT, in addtion to the CPU caching
mode. This is very much needed on newer igpu platforms which allow
incoherent GT access, where the choice over the cache level and
expected coherency is best left to userspace depending on their
usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU
cache mode to use for system memory, including the expected GPU
coherency mode. There are various restrictions here for the
selected coherency mode and compatible CPU cache modes. With that
in place the actualpat_indexcan now be provided as part ofvm_bind. The only restriction is that the coherency mode of thepat_indexmust be at least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Quoted 3 lines
--
2.41.0I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Quoted 6 lines
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportJust to confirm my understanding. We will have 2 separate APIs?
1. Forgem_createto explicitly set coherency/caching according to API enumsYeah, you just need to provide the
cpu_cachingmode (wc or wb). If you latermmap()the object this same caching mode is used. Also any KMD internal things like swap-in will also respect this caching mode. Thecoh_modeis now removed.Quoted 1 line
2. For
vm_bindto set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?Right, as part of
vm_bindyou provide whatever PAT index control (as per
Bspec) you want for each mapping of the object (you could map it multiple times).When choosing a PAT index the
cpu_cachingonly matters if you choose an incoherent PAT index (coh_nonein Bspec), since we don't want to allow cpu_caching=wb andcoh_none, since that leads to some complications in the KMD.Quoted 2 lines
Quoted 10 lines · Michal
"
Right, if you allocate WB memory (cpu_caching=wb), then you can't currently use an incoherent PAT index if accessing it from the GPU. KMD will reject it atvm_bind." "What about scenarios where I want to have:
- CPU WB caching
- GPU non coherent accessAnd manually flush cache lines on synchronization points (i.e. clFlush on mapped OpenCL buffers) ? How to do it if KMD would reject
VM_BINDoperations?
Currently it is not possible. The hope was to avoid the need for any kind of manual clflushing/wbinvd in the KMD, which would be needed to avoid potential security issues, if we start allowing incoherent GPU access to CPU cached memory.
Quoted 179 lines
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 5:08 PM
To: Mrozek, Michal <[email protected]>; Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 1 line · Michal
Will KMD fail the mmap call if we provide incorrect mapping ?
With Xe there are no mmap flags or anything like that, so you can't really provide an incorrect mapping. Ignoring this series, the previous behaviour was for the KMD to implicitly select the cpu caching mode when allocating the memory, where roughly you would get WC for VRAM and then WB for system memory, unless it's used for scanout in which case you get WC. With this series userspace must now explicitly request the cpu caching mode at allocation via
gem_create, which will be used for all CPU access, including KMD internal stuff as well as userspacemmap().Quoted 1 line
I.e. WB mmap on non-coherent GPU resource?
Right, if you allocate WB memory (cpu_caching=wb), then you can't currently use an incoherent PAT index if accessing it from the GPU. KMD will reject it at
vm_bind.Quoted 161 lines
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 4:05 PM
To: Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 8 lines · Slawomir
Hi Matthew,
In the compute driver we call
GEM_MMAP_OFFSETwith flags to specify cpu cache mode (wb/wc):GEM_CREATE_EXT()andmmap()+GEM_MMAP_OFFSET(flags=wb/wc)Starting from Xe KMD we should be able to achieve the same CPU caching policies with
GEM_CREATE_EXT:GEM_CREATE_EXT(flags=wb/wc)andmmap()+GEM_MMAP_OFFSET(flags=0)Am
I correct?Yes, that's correct.
Quoted 124 lines
Thanks,
Slawek-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Wednesday, November 8, 2023 1:04 PM
To: Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Milczarek, Slawomir <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>; Dunajski, Bartosz <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 76 lines · Tomasz Mistat
[email protected],[email protected],sa u
r [email protected],[email protected],[email protected] o
m ,[email protected],[email protected],ayaz.siddiqui@int e
l .com,[email protected],[email protected],filip.hazubski@i n t el.com,[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Reply-To:
In-Reply-To: <[email protected]>Quoted 51 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_t
y
p
e=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_index
when mapping memory via the ppGTT, in addtion to the CPU caching
mode. This is very much needed on newer igpu platforms which allow
incoherent GT access, where the choice over the cache level and
expected coherency is best left to userspace depending on their
usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU
cache mode to use for system memory, including the expected GPU
coherency mode. There are various restrictions here for the
selected coherency mode and compatible CPU cache modes. With that
in place the actualpat_indexcan now be provided as part ofvm_bind. The only restriction is that the coherency mode of thepat_indexmust be at least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Quoted 3 lines
--
2.41.0I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Quoted 6 lines
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportJust to confirm my understanding. We will have 2 separate APIs?
1. Forgem_createto explicitly set coherency/caching according to API enumsYeah, you just need to provide the
cpu_cachingmode (wc or wb). If you latermmap()the object this same caching mode is used. Also any KMD internal things like swap-in will also respect this caching mode. Thecoh_modeis now removed.Quoted 1 line
2. For
vm_bindto set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?Right, as part of
vm_bindyou provide whatever PAT index control (as per
Bspec) you want for each mapping of the object (you could map it multiple times).When choosing a PAT index the
cpu_cachingonly matters if you choose an incoherent PAT index (coh_nonein Bspec), since we don't want to allow cpu_caching=wb andcoh_none, since that leads to some complications in the KMD.Quoted 2 lines
But then we leave performance on the table:
- if we map on CPU as uncached, the allocation is slow on the CPU side
- if we map coherent on the GPU side, the allocation is slow on the GPU side
That's why OpenCL has clear synchronization points map/unmap buffer , that tells the driver where transition is coming from.
What kind of security is being compromised with this mode ?
The memory visibility is within a process , so there is no way to get anything from external processes.
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Monday, November 13, 2023 10:38 AM
To: Mrozek, Michal <[email protected]>; Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Quoted 10 lines · Michal
"
Right, if you allocate WB memory (cpu_caching=wb), then you can't currently use an incoherent PAT index if accessing it from the GPU. KMD will reject it atvm_bind." "What about scenarios where I want to have:
- CPU WB caching
- GPU non coherent accessAnd manually flush cache lines on synchronization points (i.e. clFlush on mapped OpenCL buffers) ? How to do it if KMD would reject
VM_BINDoperations?
Currently it is not possible. The hope was to avoid the need for any kind of manual clflushing/wbinvd in the KMD, which would be needed to avoid potential security issues, if we start allowing incoherent GPU access to CPU cached memory.
Quoted 179 lines
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 5:08 PM
To: Mrozek, Michal <[email protected]>; Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 1 line · Michal
Will KMD fail the mmap call if we provide incorrect mapping ?
With Xe there are no mmap flags or anything like that, so you can't really provide an incorrect mapping. Ignoring this series, the previous behaviour was for the KMD to implicitly select the cpu caching mode when allocating the memory, where roughly you would get WC for VRAM and then WB for system memory, unless it's used for scanout in which case you get WC. With this series userspace must now explicitly request the cpu caching mode at allocation via
gem_create, which will be used for all CPU access, including KMD internal stuff as well as userspacemmap().Quoted 1 line
I.e. WB mmap on non-coherent GPU resource?
Right, if you allocate WB memory (cpu_caching=wb), then you can't currently use an incoherent PAT index if accessing it from the GPU. KMD will reject it at
vm_bind.Quoted 161 lines
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 4:05 PM
To: Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 8 lines · Slawomir
Hi Matthew,
In the compute driver we call
GEM_MMAP_OFFSETwith flags to specify cpu cache mode (wb/wc):GEM_CREATE_EXT()andmmap()+GEM_MMAP_OFFSET(flags=wb/wc)Starting from Xe KMD we should be able to achieve the same CPU caching policies with
GEM_CREATE_EXT:GEM_CREATE_EXT(flags=wb/wc)andmmap()+GEM_MMAP_OFFSET(flags=0)Am
I correct?Yes, that's correct.
Quoted 124 lines
Thanks,
Slawek-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Wednesday, November 8, 2023 1:04 PM
To: Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Milczarek, Slawomir <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>; Dunajski, Bartosz <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 76 lines · Tomasz Mistat
[email protected],[email protected],sa u
r [email protected],[email protected],[email protected] o
m ,[email protected],[email protected],ayaz.siddiqui@int e
l .com,[email protected],[email protected],filip.hazubski@i n t el.com,[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Reply-To:
In-Reply-To: <[email protected]>Quoted 51 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_t
y
p
e=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_index
when mapping memory via the ppGTT, in addtion to the CPU caching
mode. This is very much needed on newer igpu platforms which allow
incoherent GT access, where the choice over the cache level and
expected coherency is best left to userspace depending on their
usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU
cache mode to use for system memory, including the expected GPU
coherency mode. There are various restrictions here for the
selected coherency mode and compatible CPU cache modes. With that
in place the actualpat_indexcan now be provided as part ofvm_bind. The only restriction is that the coherency mode of thepat_indexmust be at least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Quoted 3 lines
--
2.41.0I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Quoted 6 lines
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportJust to confirm my understanding. We will have 2 separate APIs?
1. Forgem_createto explicitly set coherency/caching according to API enumsYeah, you just need to provide the
cpu_cachingmode (wc or wb). If you latermmap()the object this same caching mode is used. Also any KMD internal things like swap-in will also respect this caching mode. Thecoh_modeis now removed.Quoted 1 line
2. For
vm_bindto set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?Right, as part of
vm_bindyou provide whatever PAT index control (as per
Bspec) you want for each mapping of the object (you could map it multiple times).When choosing a PAT index the
cpu_cachingonly matters if you choose an incoherent PAT index (coh_nonein Bspec), since we don't want to allow cpu_caching=wb andcoh_none, since that leads to some complications in the KMD.Quoted 2 lines
Quoted 7 lines · Michal
But then we leave performance on the table:
- if we map on CPU as uncached, the allocation is slow on the CPU side
- if we map coherent on the GPU side, the allocation is slow on the GPU sideThat's why OpenCL has clear synchronization points map/unmap buffer , that tells the driver where transition is coming from.
What kind of security is being compromised with this mode ?
When the KMD allocates some system memory and gives it to userspace it must always be zeroed first for security reasons. The core-kernel ensures the memory is zeroed, but this can leave the CPU caches dirty (like when allocated as WB). You also have things like like swap-in, where you also allocate system memory, but in this case we don't need to zero it and rather just use the CPU to copy the contents from swap back into usable system memory. But again CPU caches can be dirty here. Malicious user can be devious here and bind the memory as incoherent on the GPU, bypassing the CPU caches and potentially peek at the previous page contents.
In upstream i915, this is worked around by doing a clflush() of the pages for every newly allocated system memory object (unless it's dgpu or platform with shared-llc, since such platforms should always be coherent). There is also an additional clflush() during swap-in. Also every time you initially use some imported dma-buf there is a big nasty WBINVD, since it's slightly iffy on how you are meant to get the CPU address of the imported memory to be able to pass into clflush(). Also this only really works on x86, so you have to hide all this stuff when building on non-x86, which is now quite possible with dgpu.
The hope with Xe was to avoid all of that.
Quoted 204 lines
The memory visibility is within a process , so there is no way to get anything from external processes.
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Monday, November 13, 2023 10:38 AM
To: Mrozek, Michal <[email protected]>; Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 10 lines · Michal
"
Right, if you allocate WB memory (cpu_caching=wb), then you can't currently use an incoherent PAT index if accessing it from the GPU. KMD will reject it atvm_bind." "What about scenarios where I want to have:
- CPU WB caching
- GPU non coherent accessAnd manually flush cache lines on synchronization points (i.e. clFlush on mapped OpenCL buffers) ? How to do it if KMD would reject
VM_BINDoperations?Currently it is not possible. The hope was to avoid the need for any kind of manual clflushing/wbinvd in the KMD, which would be needed to avoid potential security issues, if we start allowing incoherent GPU access to CPU cached memory.
Quoted 179 lines
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 5:08 PM
To: Mrozek, Michal <[email protected]>; Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 1 line · Michal
Will KMD fail the mmap call if we provide incorrect mapping ?
With Xe there are no mmap flags or anything like that, so you can't really provide an incorrect mapping. Ignoring this series, the previous behaviour was for the KMD to implicitly select the cpu caching mode when allocating the memory, where roughly you would get WC for VRAM and then WB for system memory, unless it's used for scanout in which case you get WC. With this series userspace must now explicitly request the cpu caching mode at allocation via
gem_create, which will be used for all CPU access, including KMD internal stuff as well as userspacemmap().Quoted 1 line
I.e. WB mmap on non-coherent GPU resource?
Right, if you allocate WB memory (cpu_caching=wb), then you can't currently use an incoherent PAT index if accessing it from the GPU. KMD will reject it at
vm_bind.Quoted 161 lines
Michal
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Thursday, November 9, 2023 4:05 PM
To: Milczarek, Slawomir <[email protected]>; Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 8 lines · Slawomir
Hi Matthew,
In the compute driver we call
GEM_MMAP_OFFSETwith flags to specify cpu cache mode (wb/wc):GEM_CREATE_EXT()andmmap()+GEM_MMAP_OFFSET(flags=wb/wc)Starting from Xe KMD we should be able to achieve the same CPU caching policies with
GEM_CREATE_EXT:GEM_CREATE_EXT(flags=wb/wc)andmmap()+GEM_MMAP_OFFSET(flags=0)Am
I correct?Yes, that's correct.
Quoted 124 lines
Thanks,
Slawek-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Wednesday, November 8, 2023 1:04 PM
To: Tomasz Mistat <[email protected]>; [email protected]; Arteaga Molina, Jaime A <[email protected]>; Cetnerowski, Adam <[email protected]>; Chodor, Jaroslaw <[email protected]>; Dunajski, Bartosz <[email protected]>; Harasimiuk, Artur <[email protected]>; Hoppe, Mateusz <[email protected]>; Jablonski, Mateusz <[email protected]>; Jobczyk, Lukasz <[email protected]>; Kasper, Kacper K <[email protected]>; Milczarek, Slawomir <[email protected]>; Mrozek, Michal <[email protected]>; Rozenfeld, Piotr <[email protected]>; Wilma, Pawel <[email protected]>; Wolny, Bartlomiej <[email protected]>; Yates, Brandon <[email protected]>; Zdanowicz, Zbigniew <[email protected]>; Dunajski, Bartosz <[email protected]>
Subject: Re: FW: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportQuoted 76 lines · Tomasz Mistat
[email protected],[email protected],sa u
r [email protected],[email protected],[email protected] o
m ,[email protected],[email protected],ayaz.siddiqui@int e
l .com,[email protected],[email protected],filip.hazubski@i n t el.com,[email protected],[email protected] Bcc:
Subject: Re: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency support
Reply-To:
In-Reply-To: <[email protected]>Quoted 51 lines · Matthew Auld
Branch available here:
https://gitlab.freedesktop.org/mwa/kernel/-/tree/xe-pat-index?ref_t
y
p
e=headsIGT changes:
https://patchwork.freedesktop.org/series/124667/Mesa:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25462Goal here is to allow userspace to directly control the
pat_index
when mapping memory via the ppGTT, in addtion to the CPU caching
mode. This is very much needed on newer igpu platforms which allow
incoherent GT access, where the choice over the cache level and
expected coherency is best left to userspace depending on their
usecase. In the future there may also be other stuff encoded in thepat_index, so giving userspace direct control will also be needed there.To support this we added new
gem_createuAPI for selecting the CPU
cache mode to use for system memory, including the expected GPU
coherency mode. There are various restrictions here for the
selected coherency mode and compatible CPU cache modes. With that
in place the actualpat_indexcan now be provided as part ofvm_bind. The only restriction is that the coherency mode of thepat_indexmust be at least as coherent as thegem_createcoherency mode. There are also some special cases like with userptr and dma-buf.v2:
- Loads of improvements/tweaks. Main changes are to now allow gem_create.coh_mode <=coh_mode(pat_index), rather than it needing to match exactly. This simplifies the dma-buf policy from userspace pov. Also we now only considerCOH_NONEandCOH_AT_LEAST_1WAY.
v3:
- Rebase. Split thepte_encode()refactoring, plus various smaller tweaks and fixes.
v4:
- Rebase on Lucas' new series.
- Drop UC cache mode.
- s/smem_cpu_caching/cpu_caching/. Idea is to make VRAM WC explicit in the uapi, plus make it more future proof.
v5:
- Rebase, plus some small tweaks and fixes.
v6:
- CI hooks fixes + checkpatch.
v7:
- Some small tweaks
v8:
- Rebase on Xe2 PAT table additions.
v9:
- Dropcoh_mode.Quoted 3 lines
--
2.41.0I'm sharing down here questions passed by [email protected] lets continue review here at respective ML
Quoted 6 lines
Sent: Wednesday, November 8, 2023 09:05
Subject: RE: [Intel-xe] [PATCH v9 0/3] PAT and cache coherency supportJust to confirm my understanding. We will have 2 separate APIs?
1. Forgem_createto explicitly set coherency/caching according to API enumsYeah, you just need to provide the
cpu_cachingmode (wc or wb). If you latermmap()the object this same caching mode is used. Also any KMD internal things like swap-in will also respect this caching mode. Thecoh_modeis now removed.Quoted 1 line
2. For
vm_bindto set PAT index that is aligned with Bspec and controls more than #1. It can also override coherency/caching that was previously set?Right, as part of
vm_bindyou provide whatever PAT index control (as per
Bspec) you want for each mapping of the object (you could map it multiple times).When choosing a PAT index the
cpu_cachingonly matters if you choose an incoherent PAT index (coh_nonein Bspec), since we don't want to allow cpu_caching=wb andcoh_none, since that leads to some complications in the KMD.Quoted 2 lines