- 06 Mar, 2022 2 commits
-
-
Gwan-gyeong Mun authored
The current implementation of i915 prime mmap only works when initializing drm_i915_gem_object with shmem_region. When using LMEM, drm_i915_gem_object is initialized with ttm_system_region. In order to make prime mmap work even this case, when using LMEM (when using ttm in i915), dma_buf_ops.mmap callback function calls drm_gem_prime_mmap(). drm_gem_prime_mmap() of drm core calls internally i915_gem_mmap() so that prime mmap can perform normally. The fake offset is processed inside drm_gem_prime_mmap(). Testcase: igt/prime_mmap Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> Cc: Matthew Auld <matthew.auld@intel.com> Signed-off-by:
Gwan-gyeong Mun <gwan-gyeong.mun@intel.com> Reviewed-by:
Nirmoy Das <nirmoy.das@intel.com> Signed-off-by:
Ramalingam C <ramalingam.c@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20220225131316.1433515-3-gwan-gyeong.mun@intel.com
-
Gwan-gyeong Mun authored
The dma_buf_ops.unmap_dma_buf callback used in i915, i915_gem_unmap_dma_buf(), has the same code as drm_gem_unmap_dma_buf(). In order to eliminate defining and using duplicate function, it updates the dma_buf_ops.unmap_dma_buf callback to use drm_gem_unmap_dma_buf(). Signed-off-by:
-
- 04 Mar, 2022 2 commits
-
-
Stuart Summers authored
If RCS is not enumerated, GuC will return invalid parameters. Make sure we do not send RCS supported when we have not enumerated it. Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com> Signed-off-by:
Stuart Summers <stuart.summers@intel.com> Signed-off-by:
Matt Roper <matthew.d.roper@intel.com> Reviewed-by:
John Harrison <John.C.Harrison@Intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20220303223435.2793124-2-matthew.d.roper@intel.com
-
Matt Roper authored
In the past we've always assumed that an RCS engine is present on every platform. However now that we have compute engines there may be platforms that have CCS engines but no RCS, or platforms that are designed to have both, but have the RCS engine fused off. Various engine-centric initialization that only needs to be done a single time for the group of RCS+CCS engines can't rely on being setup with the RCS now; instead we add a I915_ENGINE_FIRST_RENDER_COMPUTE flag that will be assigned to a single engine in the group; whichever engine has this flag will be responsible for some of the general setup (RCU_MODE programming, initialization of certain workarounds, etc.). Signed-off-by:
Lucas De Marchi <lucas.demarchi@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20220303223435.2793124-1-matthew.d.roper@intel.com
-
- 03 Mar, 2022 8 commits
-
-
John Harrison authored
Some G2H handlers were reading the context id field from the payload before checking the payload met the minimum length required. Signed-off-by:
Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20220302003357.4188363-9-John.C.Harrison@Intel.com
-
John Harrison authored
The CTB registration process changed significantly a while back using a single KLV based H2G. So drop the original and now obsolete H2G definitions. Signed-off-by:
-
John Harrison authored
The LRC descriptor pool is going away. So, stop naming context ids as descriptor pool indecies. While at it, add a bunch of missing line feeds to some error messages. Signed-off-by:
-
John Harrison authored
The LRC descriptor was being initialised early on in the context registration sequence. It could then be determined that the actual registration needs to be delayed and the descriptor would be wiped out. This is inefficient, so move the setup to later in the process after the point of no return. v2: Move some split changes into the split patch (and do them correctly). Signed-off-by:
-
John Harrison authored
The LRC descriptor pool is going away. Further, the function that was populating it was also doing a bunch of logic about the context registration sequence. So, split that code apart into separate state setup and try to register functions. Note that some of those 'try to register' code paths actually undo the state setup and leave it to be redone again later (with potentially different values). This is inefficient. The next patch will correct this. Also, move a comment about ignoring return values to the place where the return values are actually ignored. v2: Move some more splitting from a later patch (and do it correctly). Signed-off-by:
-
John Harrison authored
The LRC descriptor pool is going away. So, stop using it as the limit for how many context ids are available. Instead, size the pool according to the number of contexts allowed. Note that this is just a naming change, the actual limit is identical in value. While at it, also update a kzalloc(sizeof()*count) to be a kcalloc(count,size). Signed-off-by:
-
John Harrison authored
The LRC descriptor pool is going away. So, stop using it as a check for whether submission has been initialised or not. Signed-off-by:
-
John Harrison authored
The LRC descriptor pool is going away. So, stop using it as a check for context registration, use the GuC id instead (being the thing that actually gets registered with the GuC). Also, rename the set/clear/query helper functions for context id mappings to better reflect their purpose and to differentiate from other registration related helper functions. Signed-off-by:
-
- 02 Mar, 2022 14 commits
-
-
Srinivasan Shanmugam authored
Registers that exist in the shared render/compute reset domain need to be placed on an engine workaround list to ensure that they are properly re-applied whenever an RCS or CCS engine is reset. We have a number of workarounds (updating registers MLTICTXCTL, L3SQCREG1_CCS0, GEN12_MERT_MOD_CTRL, and GEN12_GAMCNTRL_CTRL) that are incorrectly implemented on the 'gt' workaround list and need to be moved accordingly. Cc: Matt Roper <matthew.d.roper@intel.com> Signed-off-by:
Srinivasan Shanmugam <srinivasan.s@intel.com> Signed-off-by:
-
Matt Roper authored
Additional workarounds are required once we start exposing CCS engines. Note that we have a number of workarounds that update registers in the shared render/compute reset domain. Historically we've just added such registers to the RCS engine's workaround list. But going forward we should be more careful to place such workarounds on a wa_list for an engine that definitely exists and is not fused off (e.g., a platform with no RCS would never apply the RCS wa_list). We'll keep rcs_engine_wa_init() focused on RCS-specific workarounds that only need to be applied if the RCS engine is present. A separate general_render_compute_wa_init() function will be used to define workarounds that touch registers in the shared render/compute reset domain and that we need to apply regardless of what render and/or compute engines actually exist. Any workarounds defined in this new function will internally be added to the first present RCS or CCS engine's workaround list to ensure they get applied (and only get applied once rather than being needlessly re-applied several times). Co-author: Srinivasan Shanmugam Signed-off-by:
-
Daniele Ceraolo Spurio authored
HW resources are divided across the active CCS engines at the compute slice level, with each CCS having priority on one of the cslices. If a compute slice has no enabled DSS, its paired compute engine is not usable in full parallel execution because the other ones already fully saturate the HW, so consider it fused off. v2 (José): - moved it to its own function - fixed definition of ccs_mask v3 (Matt): - Replace fls() condition with a simple IP version test v4 (Matt): - Don't try to calculate a ccs_mask using intel_slicemask_from_dssmask() until we've determined that we're running on an Xe_HP platform where the logic makes sense (and won't overflow). Cc: Stuart Summers <stuart.summers@intel.com> Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com> Cc: Ashutosh Dixit <ashutosh.dixit@intel.com> Signed-off-by:
-
Matthew Brost authored
A different emit breadcrumbs ring programming is required for compute / render and we don't have UMD user so just reject parallel submission for these engine classes. Signed-off-by:
Matthew Brost <matthew.brost@intel.com> Signed-off-by:
-
Daniele Ceraolo Spurio authored
Tell GuC that CCS is enabled by setting the CCS mask in its ADS. Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com> Original-author: Michel Thierry Signed-off-by:
-
Matt Roper authored
We have to specify in the Render Control Unit Mode register when CCS is enabled. v2: - Move RCU_MODE programming to a helper function. (Tvrtko) - Clean up and clarify comments. (Tvrtko) - Add RCU_MODE to the GuC save/restore list. (Daniele) v3: - Move this patch before the GuC ADS update to enable compute engines; the definition of RCU_MODE and its insertion into the save/restore list moves to this patch. (Daniele) v4: - Call xehp_enable_ccs_engines() directly in guc_resume() and execlists_resume() rather than adding an extra layer of wrapping to the engine->resume() vfunc. (Umesh) Bspec: 46034 Original-author: Michel Thierry Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com> Cc: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> Signed-off-by:
Aravind Iddamsetty <aravind.iddamsetty@intel.com> Signed-off-by:
Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20220302001554.1836066-1-matthew.d.roper@intel.com
-
Matt Roper authored
In Dual Context mode the EUs are shared between render and compute command streamers. The hardware provides a field in the lrc descriptor to indicate the prioritization of the thread dispatch associated to the corresponding context. The context priority is set to 'low' at creation time and relies on the existing context priority to set it to low/normal/high. Bspec: 46145, 46260 Original-author: Michel Thierry Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Signed-off-by:
Prasad Nallani <prasad.nallani@intel.com> Signed-off-by:
Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20220301231549.1817978-8-matthew.d.roper@intel.com
-
Matt Roper authored
This is a more appropriate header for these definitions. v2: - Cleanup whitespace. (Lucas) Signed-off-by:
-
Matt Roper authored
The compute engine handles the same commands the render engine can (except 3D pipeline), so it makes sense that CCS is more similar to RCS than non-render engines. The CCS context state (lrc) is also similar to the render one, so reuse it. Note that the compute engine has its own CTX_R_PWR_CLK_STATE register. In order to avoid having multiple RCS && CCS checks, add the following engine flag: - I915_ENGINE_HAS_RCS_REG_STATE - use the render (larger) reg state ctx. BSpec: 46260 Original-author: Michel Thierry Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Signed-off-by:
-
Daniele Ceraolo Spurio authored
CCS will reuse the RCS functions for breadcrumb and flush emission. However, CCS pipe_control has additional programming restrictions: - Command Streamer Stall Enable must be always set - Post Sync Operations must not be set to Write PS Depth Count - 3D-related bits must not be set v2: - Drop unwanted blank line. (Lucas) Bspec: 47112 Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com> Signed-off-by:
-
Matt Roper authored
Add execlists and GuC interrupts for compute CS into existing IRQ handlers. All compute command streamers belong to the same compute class, so the only change needed to enable their interrupts is to program their GT engine interrupt mask registers. CCS0 shares the register with CCS1, while CCS2 and CCS3 are in a new one. BSpec: 50844, 54029, 54030, 53223, 53224. Original-author: Michel Thierry Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com> Signed-off-by:
-
Matt Roper authored
The reset domain is shared between render and all compute engines, so resetting one will affect the others. Note: Before performing a reset on an RCS or CCS engine, the GuC will attempt to preempt-to-idle the other non-hung RCS/CCS engines to avoid impacting other clients (since some shared modules will be reset). If other engines are executing non-preemptable workloads, the impact is unavoidable and some work may be lost. Bspec: 52549 Original-author: Michel Thierry Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com> Signed-off-by:
-
Matt Roper authored
Introduce a Compute Command Streamer (CCS), which has access to the media and GPGPU pipelines (but not the 3D pipeline). To begin with, define the compute class/engine common functions, based on the existing render ones. v2: - Add kerneldoc for drm_i915_gem_engine_class since we're adding a new element to it. (Daniel) - Make engine class <-> guc class converters use lookup tables to make it more clear/explicit how the IDs map. (Tvrtko) v3: - Don't update uapi for now; we'll just include the driver-internal changes for the time being. Bspec: 46167, 45544 Original-author: Michel Thierry Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com> Signed-off-by:
Rodrigo Vivi <rodrigo.vivi@intel.com> Signed-off-by:
-
Sebastian Andrzej Siewior authored
There are a few sections in the driver which are not compatible with PREEMPT_RT. They trigger warnings and can lead to deadlocks at runtime. Disable the i915 driver on a PREEMPT_RT enabled kernel. This way PREEMPT_RT itself can be enabled without needing to address the i915 issues first. The RT related patches are still in RT queue and will be handled later. Signed-off-by:
Sebastian Andrzej Siewior <bigeasy@linutronix.de> Acked-by:
-
- 01 Mar, 2022 7 commits
-
-
John Harrison authored
It is possible for reset notifications to arrive for a context that is in the process of being banned. So don't flag these as an error, just report it as informational (because it is still useful to know that resets are happening even if they are being ignored). v2: Better wording for the message (review feedback from Tvrtko). v3: Fix rebase issue (review feedback from Daniele). Signed-off-by:
-
Daniele Ceraolo Spurio authored
Move initialization of submission-related spinlock, lists and workers to init_early. This fixes an issue where if the GuC init fails we might still try to get the lock in the context cleanup code. Note that it is safe to call the GuC context cleanup code even if the init failed because all contexts are initialized with an invalid GuC ID, which will cause the GuC side of the cleanup to be skipped, so it is easier to just make sure the variables are initialized than to special case the cleanup to handle the case when they're not. References: https://gitlab.freedesktop.org/drm/intel/-/issues/4932Signed-off-by:
-
John Harrison authored
A flag query helper was actually writing to the flags word rather than just reading. Fix that. Also update the function's comment as it was out of date. NB: No need for a 'Fixes' tag. The test was only ever used inside a BUG_ON during context registration. Rather than asserting that the condition was true, it was making the condition true. So, in theory, there was no consequence because we should never have hit a BUG_ON anyway. Which means the write should always have been a no-op. Signed-off-by:
-
Matthew Auld authored
Exercise each of the migration scenarios, verifying that the final placement and buffer contents match our expectations. v2(Thomas): Replace for_i915_gem_ww() block with simpler object_lock() v3: - For testing purposes allow forcing the io_size such that we can exercise the allocation + migration path on devices that don't have the small BAR limit. Signed-off-by:
Matthew Auld <matthew.auld@intel.com> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> Reviewed-by:
Thomas Hellström <thomas.hellstrom@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20220228123607.580432-4-matthew.auld@intel.com
-
Matthew Auld authored
If we have to contend with non-mappable LMEM, then we need to ensure the object fits within the mappable portion, like in the selftests, where we later try to CPU access the pages. However if it can't then we need to gracefully handle this, without throwing an error. Also it looks like TTM will return -ENOMEM, in ttm_bo_mem_space() after exhausting all possible placements. Signed-off-by:
-
Matthew Auld authored
The end goal is to have userspace tell the kernel what buffers will require CPU access, however if we ever reach the CPU fault handler, and the current resource is not mappable, then we should attempt to migrate the buffer to the mappable portion of LMEM, or even system memory, if the allowable placements permit it. Signed-off-by:
-
Matthew Auld authored
If we need to make room for some mappable object, then we should only victimize objects that have one or pages that occupy the visible portion of LMEM. Let's also create a new priority hint for objects that are placed in mappable memory, where we know that CPU access was requested, that way we hopefully victimize these last. v2(Thomas): s/TTM_PL_PRIV/I915_PL_LMEM0/ Signed-off-by:
-
- 28 Feb, 2022 7 commits
-
-
Thomas Hellström authored
It's unclear what reference the initial vma kref reference refers to. A vma can have multiple weak references, the object vma list, the vm's bound list and the GT's closed_list, and the initial vma reference can be put from lookups of all these lists. With the current implementation this means that any holder of yet another vma refcount (currently only i915_gem_object_unbind()) needs to be holding two of either *) An object refcount, *) A vm open count *) A vma open count in order for us to not risk leaking a reference by having the initial vma reference being put twice. Address this by re-introducing i915_vma_destroy() which removes all weak references of the vma and *then* puts the initial vma refcount. This makes a strong vma reference hold on to the vma unconditionally. Perhaps a better name would be i915_vma_revoke() or i915_vma_zombify(), since other callers may still hold a refcount, but with the prospect of being able to replace the vma refcount with the object lock in the near future, let's stick with i915_vma_destroy(). Finally this commit fixes a race in that previously i915_vma_release() and now i915_vma_destroy() could destroy a vma without taking the vm->mutex after an advisory check that the vma mm_node was not allocated. This would race with the ungrab_vma() function creating a trace similar to the below one. This was fixed in one of the __i915_vma_put() callsites in commit bc1922e5 ("drm/i915: Fix a race between vma / object destruction and unbinding") but although not seemingly triggered by CI, that is not sufficient. This patch is needed to fix that properly. [823.012188] Console: switching to colour dummy device 80x25 [823.012422] [IGT] gem_ppgtt: executing [823.016667] [IGT] gem_ppgtt: starting subtest blt-vs-render-ctx0 [852.436465] stack segment: 0000 [#1] PREEMPT SMP NOPTI [852.436480] CPU: 0 PID: 3200 Comm: gem_ppgtt Not tainted 5.16.0-CI-CI_DRM_11115+ #1 [852.436489] Hardware name: Intel Corporation Alder Lake Client Platform/AlderLake-P DDR5 RVP, BIOS ADLPFWI1.R00.2422.A00.2110131104 10/13/2021 [852.436499] RIP: 0010:ungrab_vma+0x9/0x80 [i915] [852.436711] Code: ef e8 4b 85 cf e0 e8 36 a3 d6 e0 8b 83 f8 9c 00 00 85 c0 75 e1 5b 5d 41 5c 41 5d c3 e9 d6 fd 14 00 55 53 48 8b af c0 00 00 00 <8b> 45 00 85 c0 75 03 5b 5d c3 48 8b 85 a0 02 00 00 48 89 fb 48 8b [852.436727] RSP: 0018:ffffc90006db7880 EFLAGS: 00010246 [852.436734] RAX: 0000000000000000 RBX: ffffc90006db7598 RCX: 0000000000000000 [852.436742] RDX: ffff88815349e898 RSI: ffff88815349e858 RDI: ffff88810a284140 [852.436748] RBP: 6b6b6b6b6b6b6b6b R08: ffff88815349e898 R09: ffff88815349e8e8 [852.436754] R10: 0000000000000001 R11: 0000000051ef1141 R12: ffff88810a284140 [852.436762] R13: 0000000000000000 R14: ffff88815349e868 R15: ffff88810a284458 [852.436770] FS: 00007f5c04b04e40(0000) GS:ffff88849f000000(0000) knlGS:0000000000000000 [852.436781] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [852.436788] CR2: 00007f5c04b38fe0 CR3: 000000010a6e8001 CR4: 0000000000770ef0 [852.436797] PKRU: 55555554 [852.436801] Call Trace: [852.436806] <TASK> [852.436811] i915_gem_evict_for_node+0x33c/0x3c0 [i915] [852.437014] i915_gem_gtt_reserve+0x106/0x130 [i915] [852.437211] i915_vma_pin_ww+0x8f4/0xb60 [i915] [852.437412] eb_validate_vmas+0x688/0x860 [i915] [852.437596] i915_gem_do_execbuffer+0xc0e/0x25b0 [i915] [852.437770] ? deactivate_slab+0x5f2/0x7d0 [852.437778] ? _raw_spin_unlock_irqrestore+0x50/0x60 [852.437789] ? i915_gem_execbuffer2_ioctl+0xc6/0x2c0 [i915] [852.437944] ? init_object+0x49/0x80 [852.437950] ? __lock_acquire+0x5e6/0x2580 [852.437963] i915_gem_execbuffer2_ioctl+0x116/0x2c0 [i915] [852.438129] ? i915_gem_do_execbuffer+0x25b0/0x25b0 [i915] [852.438300] drm_ioctl_kernel+0xac/0x140 [852.438310] drm_ioctl+0x201/0x3d0 [852.438316] ? i915_gem_do_execbuffer+0x25b0/0x25b0 [i915] [852.438490] __x64_sys_ioctl+0x6a/0xa0 [852.438498] do_syscall_64+0x37/0xb0 [852.438507] entry_SYSCALL_64_after_hwframe+0x44/0xae [852.438515] RIP: 0033:0x7f5c0415b317 [852.438523] Code: b3 66 90 48 8b 05 71 4b 2d 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 2e 0f 1f 84 00 00 00 00 00 b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 41 4b 2d 00 f7 d8 64 89 01 48 [852.438542] RSP: 002b:00007ffd765039a8 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 [852.438553] RAX: ffffffffffffffda RBX: 000055e4d7829dd0 RCX: 00007f5c0415b317 [852.438562] RDX: 00007ffd76503a00 RSI: 00000000c0406469 RDI: 0000000000000017 [852.438571] RBP: 00007ffd76503a00 R08: 0000000000000000 R09: 0000000000000081 [852.438579] R10: 00000000ffffff7f R11: 0000000000000246 R12: 00000000c0406469 [852.438587] R13: 0000000000000017 R14: 00007ffd76503a00 R15: 0000000000000000 [852.438598] </TASK> [852.438602] Modules linked in: snd_hda_codec_hdmi i915 mei_hdcp x86_pkg_temp_thermal snd_hda_intel snd_intel_dspcfg drm_buddy coretemp crct10dif_pclmul crc32_pclmul snd_hda_codec ttm ghash_clmulni_intel snd_hwdep snd_hda_core e1000e drm_dp_helper ptp snd_pcm mei_me drm_kms_helper pps_core mei syscopyarea sysfillrect sysimgblt fb_sys_fops prime_numbers intel_lpss_pci smsc75xx usbnet mii [852.440310] ---[ end trace e52cdd2fe4fd911c ]--- v2: Fix typos in the commit message. Fixes: 7e00897b ("drm/i915: Add object locking to i915_gem_evict_for_node and i915_gem_evict_something, v2.") Fixes: bc1922e5 ("drm/i915: Fix a race between vma / object destruction and unbinding") Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by:
-
Matthew Auld authored
Check that mappable vs non-mappable matches our expectations. Signed-off-by:
-
Matthew Auld authored
Otherwise we get -EINVAL, instead of the more useful -E2BIG if the allocation doesn't fit within the pfn range, like with mappable lmem. The hugepages selftest, for example, needs this to know if a smaller size is needed. Signed-off-by:
-
Matthew Auld authored
Differentiate between mappable vs non-mappable resources, also if this is an actual range allocation ensure we set res->start as the starting pfn. Later when we need to do non-mappable -> mappable moves then we want TTM to see that the current placement is not compatible, which should result in an actual move, instead of being turned into a noop. Signed-off-by:
-
Matthew Auld authored
Track the total amount of available visible memory, and also track per-resource the amount of used visible memory. For now this is useful for our debug output, and deciding if it is even worth calling into the buddy allocator. In the future tracking the per-resource visible usage will be useful for when deciding if we should attempt to evict certain buffers. v2: - s/place->lpfn/lpfn/, that way we can avoid scanning the list if the entire range is already mappable. - Move the end declaration inside the if block(Thomas). - Make sure to also account for reserved memory. Signed-off-by:
-
Matthew Auld authored
If the user doesn't require CPU access for the buffer, then ALLOC_GPU_ONLY should be used, in order to prioritise allocating in the non-mappable portion of LMEM, on devices with small BAR. v2(Thomas): - The BO_ALLOC_TOPDOWN naming here is poor, since this is pure lies on systems that don't even have small BAR. A better name is GPU_ONLY, which is accurate regardless of the configuration. Signed-off-by: -
Matthew Auld authored
On devices with non-mappable LMEM ensure we always allocate the pages within the mappable portion. For now we assume that all LMEM buffers will require CPU access, which is also inline with pretty much all current kernel internal users. In the next patch we will introduce a new flag to override this behaviour. Signed-off-by:
-
