- 18 Dec, 2013 40 commits
-
-
Ben Widawsky authored
Originally this commit message said: Now that do_switch does the mm switch, and we always enable the aliasing PPGTT, and contexts at the same time, there is no need to continue doing this during PPGTT enabling. Since originally writing the patch however, I introduced the concept of synchronous mm switching (using MMIO). Since this is generally not recommended in the spec (for reasons unknown), I've isolated its usage as much as possible. As such the "extraneous" switch only ever will occur when we have full PPGTT. Signed-off-by:
Ben Widawsky <ben@bwidawsk.net> Signed-off-by:
Daniel Vetter <daniel.vetter@ffwll.ch>
-
Ben Widawsky authored
As with processes which run on the CPU, the goal of multiple VMs is to provide process isolation. Specific to GEN, there is also the ability to map more objects per process (2GB each instead of 2Gb-2k total). For the most part, all the pipes have been laid, and all we need to do is remove asserts and actually start changing address spaces with the context switch. Since prior to this we've converted the setting of the page tables to a streamed version, this is quite easy. One important thing to point out (since it'd been hotly contested) is that with this patch, every context created will have it's own address space (provided the HW can do it). v2: Disable BDW on rebase NOTE: I tried to make this commit as small as possible. I needed one place where I could "turn everything on" and that is here. It could be split into finer commits, but I didn't really see much point. Cc: Eric Anholt <eric@anholt.net> Cc: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by:
-
Daniel Vetter authored
I need the tricky do_switch fix before I can merge the final piece of the ppgtt enabling puzzle. Otherwise the conflict will be a real pain to resolve since the do_switch hunk from -fixes must be placed at the exact right place within a hunk in the next patch. Conflicts: drivers/gpu/drm/i915/i915_gem_context.c drivers/gpu/drm/i915/i915_gem_execbuffer.c drivers/gpu/drm/i915/intel_display.c Signed-off-by: -
Ben Widawsky authored
This is primarily a band aid for an unexplainable error in gem_reloc_vs_gpu/forked-faulting-reloc-thrashing. Essentially as soon as a relocated buffer (which had a non-zero presumed offset) moved to offset 0, something goes bad. Since I have been unable to solve this, and potentially this is a good thing to do anyway, since many things can accidentally write to offset 0, why not? Signed-off-by:
-
Ben Widawsky authored
It's quite common for an object to simply be on the inactive list (and not unbound) when we want to free the context. This of course happens with lazy unbinding. Simply, this is needed when an object isn't fully unbound but we want to free one VMA of the object, for whatever reason. NOTE: The aliasing PPGTT is not a proper VM, so it needs special casing. This addresses the fixup requirement mentioned in: drm/915: Better reset handling for contexts In the flink, and dmabuf case, we can't assert that the object isn't still active. To keep it more generic, just check the vma's link in the object vma list. If we wanted to do a better job, we could track last seqno (and active) per VMA. It was decided not to do this in the last iteration. Unfortunately this means the assertion can miss real bugs when using flink/dmabuf. v2: Use the newer introduced i915_gem_evict_vm(). Note that handling the aliasing PPGTT is special. Signed-off-by:
-
Ben Widawsky authored
With context destruction, we always want to be able to tear down the underlying address space. This is invoked on the last unreference to the context which could happen before we've moved all objects to the inactive list. To enable a clean tear down the address space, make sure to process the request free lastly. Without this change, we cannot guarantee to we don't still have active objects in the VM. As an example of a failing case: CTX-A is created, count=1 CTX-A is used during execbuf does a context switch count = 2 and add_request count = 3 CTX B runs, switches, CTX-A count = 2 CTX-A is destroyed, count = 1 retire requests is called free_request from CTX-A, count = 0 <--- free context with active object As mentioned above, by doing the free request after processing the active list, we can avoid this case. Signed-off-by:
-
Ben Widawsky authored
We need to have the address space when reserving space for the objects. Since the address space and context are tied together, and reserve occurs before context switch (for good reason), we must lookup our context earlier in the process. This leaves some room for optimizations where we no longer need to use ctx_id in certain places. This will be addressed in a subsequent patch. Important tricky bit: Because slow relocations during execbuffer drop struct_mutex Perhaps it would be best to acquire the reference when we get the context, but I'll save that for another day (note I have written the patch before, and I found the changes required to be uglier than this). Note that since we currently access everything via context id, and not the data structure this is fine, though not desirable. The next change attempts to get the context only once via the context ID idr lookup, and as such, the following can happen: CTX-A is created, refcount = 1 CTX-A execbuf, mutex dropped close IOCTL called on CTX-A, refcount = 0 CTX-A resumes in execbuf. v2: Rebased on top of commit b6359918 Author: Mika Kuoppala <mika.kuoppala@linux.intel.com> Date: Wed Oct 30 15:44:16 2013 +0200 drm/i915: add i915_get_reset_stats_ioctl v3: Rebased on top of commit 25b3dfc8 Author: Mika Westerberg <mika.westerberg@linux.intel.com> Date: Tue Nov 12 11:57:30 2013 +0200 Author: Mika Kuoppala <mika.kuoppala@linux.intel.com> Date: Tue Nov 26 16:14:33 2013 +0200 drm/i915: check context reset stats before relocations Signed-off-by:
-
Ben Widawsky authored
To simplify the codepaths somewhat, we can simply always create a context. Contexts already keep hangstat information. This prevents us from having to differentiate at other parts in the code. There is allocation overhead, but it should not be measurable. Signed-off-by:
-
Ben Widawsky authored
Every file will get it's own context, and we use this context instead of the default context. The default context still exists for future shrinker usage as well as reset handling. v2: Updated to address Mika's recent context guilty changes Some more changes around this come up in later patches as well. v3: Use a fake context to avoid allocation for the !HAS_HW_CONTEXT case. I've tried the alternatives. This looks the best to me. Removed hangstat stuff from v2 - for a separate patch Demote failed PPGTT set to DRM_DEBUG_DRIVER since it can now be invoked easily from userspace. Signed-off-by:
-
Ben Widawsky authored
We have a default context which suits the aliasing PPGTT well. Tie them together so it looks like any other context/PPGTT pair. This makes the code cleaner as it won't have to special case aliasing as often. The patch has one slightly tricky part in the default context creation function. In the future (and on aliased setup) we create a new VM for a context (potentially). However, if we have aliasing PPGTT, which occurs at this point in time for all platforms GEN6+, we can simply manage the refcounting to allow things to behave as normal. Now is a good time to recall that the aliasing_ppgtt doesn't have a real VM, it uses the GGTT drm_mm. Signed-off-by:
-
Ben Widawsky authored
In following with the old restore code, we must now restore ever PPGTT's PDEs, since they aren't proper GEM ojbects. v2: Rebased on BDW. Only do restore pdes for gen6 & 7 Signed-off-by:
-
Ben Widawsky authored
We won't be calling enable() for all PPGTTs. We do need to write PDEs for all PPGTTs however. By moving the writing to init (which is called for all PPGTTs) we should accomplish this. ADD NOTE ABOUT PDE restore TODO: Eventually, we should allocate the page tables on demand. v2: Rebased on BDW. Only do PDEs for pre-gen8 Signed-off-by:
-
Ben Widawsky authored
Pretty straightforward so far except for the bit about the refcounting. The PPGTT will potentially be shared amongst multiple contexts. Because contexts themselves have a refcounted lifecycle, the easiest way to manage this will be to refcount the PPGTT. To acheive this, we piggy back off of the existing context refcount, and will increment and decrement the PPGTT refcount with context creation, and destruction. To put it more clearly, if context A, and context B both use PPGTT 0, we can't free the PPGTT until both A, and B are destroyed. Note that because the PPGTT is permanently pinned (for now), it really just matters for the PPGTT destruction, as opposed to making space under memory pressure. Signed-off-by:
-
Ben Widawsky authored
This patch consolidates the way in which we handle the various supported PPGTT by module parameter in addition to what the hardware supports. It strives to make doing the right thing in the code as simple as possible, with the USES_ macros. I've opted to add the full PPGTT argument simply so one can see how I intend to use this function. It will not/cannot be used until later. Signed-off-by:
-
Ben Widawsky authored
Rearrange the initialization code to try to special case the aliasing PPGTT less, and provide usable interfaces for the general case later. Signed-off-by:
-
Ben Widawsky authored
I've found this by accident. The docs don't really come out and say you need to do this. What the docs do tell you is you need to flush the TLBs before you set the PP_DIR_BASE, and that the RCS will invalidate its TLBs upon setting the new PP_DIR_BASE. It makes no such comment about any of the other rings. Empirically, this indeed fixes a really obvious bug whereby the batches being sent to the blitter were not executing (we were executing the HSWP somehow instead). NOTE: This should make no difference with the current code. It only applies when we start using multiple VMs. NOTE2: HSW appears to be immune to this. Signed-off-by:
-
Ben Widawsky authored
The docs seem to suggest this is the appropriate method (though it doesn't say so outright). In other words, we probably should have done this before. We certainly must do this for switching VMs on the fly, since synchronizing the rings to MMIO updates isn't acceptable. v2: Make the reset code actually work for all rings. Note that this was fixed in subsequent commits, but was indeed broken for this commit. Add a posting read to the reset case. It probably should have existed before hand, but since we have no failures; there is no reason to make it a separate commit. Make IS_GEN6 not use the ring because I am seeing crashes when using it. It is a bit of a hack in this patch, it will get fixed up in a couple of patches. Signed-off-by:
-
Ben Widawsky authored
In order to do the full context switch with address space, it's convenient to have a way to switch the address space. We already have this in our code - just pull it out to be called by the context switch code later. v2: Rebased on BDW support. Required adding BDW. Signed-off-by:
-
Ben Widawsky authored
Signed-off-by:
-
Ben Widawsky authored
The patch before this changed the way in which we allocate space for the PPGTT PDEs. It began carving out the PPGTT PDEs (which live in the Global GTT) from the GGTT's drm_mm. Prior to that patch, the PDEs were hidden from the drm_mm, and therefore could never fail to be allocated. In unfortunate cases, the drm_mm may be full when we want to allocate the space. This can technically occur whenever we try to allocate, which happens in two places currently. Practically, it can only really ever happen at GPU reset. Later, when we allocate more PDEs for multiple PPGTTs this will potentially even more useful. Signed-off-by:
-
Ben Widawsky authored
When PPGTT support was originally enabled, it was only designed to support 1 PPGTT. It therefore made sense to simply hide the GGTT space required to enable this from the drm_mm allocator. Since we intend to support full PPGTT, which means more than 1, and they can be created and destroyed ad hoc it will be required to use the proper allocation techniques we already have. The first step here is to make the existing single PPGTT use the allocator. The astute observer will notice that we are reserving space in the GGTT for the PDEs for the lifetime of the address space, and would be right to question whether or not this is a good idea. It does not make a difference with this current patch only the aliasing PPGTT (indeed the PDEs should still be hidden from the shrinker). For the future, we are allocating from top to bottom to avoid using the precious "gtt space" The GGTT space at that point should only be used for scanout, HW contexts, ringbuffers, HWSP, PDEs, and a couple of other small buffers (potentially) used by the kernel. Everything else should be mapped into a PPGTT. To put the consumption in more tangible terms, it takes approximately 4 sets of PDEs to equal one 19x10 framebuffer (with no fancy stride or alignment constraints). 3/4 of the total [average] GGTT can be used for PDEs, and hopefully never touch the 1/4 that the framebuffer needs. The astute, and persistent observer might ask about the page tables which are also pinned for the address space. This waste is unfortunate. We use 2MB of memory per address space. We leave wrapping the PDEs as a real GEM object as a TODO. v2: Align PDEs to 64b in GTT Allocate the node dynamically so we can use drm_mm_put_block Now tested on IGT Allocate node at the top to avoid fragmentation (Chris) v3: Use Chris' top down allocator v4: Embed drm_mm_node into ppgtt struct (Jesse) Remove hunks which didn't belong (Jesse) v5: Don't subtract guard page since we now killed the guard page prior to this patch. (Ben) v6: Rebased and removed guard page stuff. Added a chunk to the commit message Allow adding a context to mappable region v7: Undo v3, so we can make the drm patch last in the series Cc: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org> (v4) Signed-off-by:
-
Ben Widawsky authored
Signed-off-by:
-
Ben Widawsky authored
The plan to to make every file descriptor have a default context. To accommodate this, generalize out default context setup function so it can be used at file open time. Signed-off-by:
-
Ben Widawsky authored
We **need** to do this for exactly 1 reason, because we want to embed a PPGTT into the context, but we don't want to special case the default context. To achieve that, we must be able to initialize contexts after the GTT is setup (so we can allocate and pin the default context's BO), but before the PPGTT and rings are initialized. This is because, currently, context initialization requires ring usage. We don't have rings until after the GTT is setup. If we split the enabling part of context initialization, the part requiring the ringbuffer, we can untangle this, and then later embed the PPGTT Incidentally this allows us to also adhere to the original design of context init/fini in future patches: they were only ever meant to be called at driver load and unload. v2: Move hw_contexts_disabled test in i915_gem_context_enable() (Chris) v3: BUG_ON after checking for disabled contexts. Or else it blows up pre gen6 (Ben) v4: Forward port Modified enable for each ring, since that patch is earlier in the series Dropped ring arg from create_default_context so it can be used by others Signed-off-by:
-
Ben Widawsky authored
This patch adds to changes for contexts on reset: Sets last context to default - this will prevent the context switch happening after a reset. That switch is not possible because the rings are hung during reset and context switch requires reset. This behavior will need to be reworked in the future, but this is what we want for now. In the future, we'll also want to reset the guilty context to uninitialized. We should wait for ARB_Robustness related code to land for that. This is somewhat for paranoia. Because we really don't know what the GPU was doing when it hung, or the state it was in (mid context write, for example), later restoring the context is a bad idea. By setting the flag to not initialized, the next load of that context will not restore the state, and thus on the subsequent switch away from the context will overwrite the old data. NOTE: This code needs a fixup when we actually have multiple VMs. The issue that can occur is inactive objects in a VM will need to be destroyed before the last context unref. This can now happen via the fake switch introduced in this patch (and it other ways in the future) Signed-off-by:
-
Ben Widawsky authored
Previously we dropped the association of a context to a ring. It is however very important to know which ring a context ran on (we could have reused the other member, but I was nitpicky). This is very important when we switch address spaces, which unlike context objects, do change per ring. As an example, if we have: RCS BCS ctx A ctx A ctx B ctx B Without tracking the last ring B ran on, we wouldn't know to switch the address space on BCS in the last row. As a result, we no longer need to track which ring a context "belongs" to, as it never really made much sense anyway. Signed-off-by: -
Ben Widawsky authored
If we want to use contexts in more abstract terms (specifically with PPGTT in mind), we need to allow them to be specified for any ring. Since the upcoming patches will bring about the use of multiple address spaces, and each ring needs to have an address space programmed (which we intend to do at context switch time), we can no longer only use RCS. With multiple rings having a last context, we must now unreference these contexts. NOTE: This commit requires an update to intel-gpu-tools to make it not fail. v2: Rebased with some logical conflicts. Squashed in the context fini refcount patch Signed-off-by:
-
Ben Widawsky authored
Signed-off-by:
-
Ben Widawsky authored
With the introduction of contexts per fd in the future, one can easily envision more contexts being used. We do not have an easy remedy to reduce the space requirements of the contexts, we can make things slightly better by using less stringent alignments on later hardware. Ville: Since I can almost predict you'll point this out. I can no longer find the docs which specify the 64k requirement on certain gen6 SKUs. If you'd like to change that too, be my guest. CC: Ville Syrjälä <ville.syrjala@linux.intel.com> Signed-off-by:
-
Ben Widawsky authored
We'll be doing a bit more stuff with each file, so having our own open function should make things clean. This also allows us to easily add conditionals for stuff we don't want to do when we don't have HW contexts. Signed-off-by:
-
Ben Widawsky authored
The only place we were using it was for GEN6, which won't have PPGTT support anyway (ie. the VM is always the same). To clear things up, (it only added confusion for me since it doesn't allow us to assert vma->vm is what we always want, when just looking at the code). Signed-off-by:
-
Ben Widawsky authored
To sum up what goes on here, we abstract the vma binding, similarly to the previous object binding. This helps for distinguishing legacy binding, versus modern binding. To keep the code churn as minimal as possible, I am leaving in insert_entries(). It serves as the per platform pte writing basically. bind_vma and insert_entries do share a lot of similarities, and I did have designs to combine the two, but as mentioned already... too much churn in an already massive patchset. What follows are the 3 commits which existed discretely in the original submissions. Upon rebasing on Broadwell support, it became clear that separation was not good, and only made for more error prone code. Below are the 3 commit messages with all their history. drm/i915: Add bind/unbind object functions to VMA drm/i915: Use the new vm [un]bind functions drm/i915: reduce vm->insert_entries() usage drm/i915: Add bind/unbind object functions to VMA As we plumb the code with more VM information, it has become more obvious that the easiest way to deal with bind and unbind is to simply put the function pointers in the vm, and let those choose the correct way to handle the page table updates. This change allows many places in the code to simply be vm->bind, and not have to worry about distinguishing PPGTT vs GGTT. Notice that this patch has no impact on functionality. I've decided to save the actual change until the next patch because I think it's easier to review that way. I'm happy to squash the two, or let Daniel do it on merge. v2: Make ggtt handle the quirky aliasing ppgtt Add flags to bind object to support above Don't ever call bind/unbind directly for PPGTT until we have real, full PPGTT (use NULLs to assert this) Make sure we rebind the ggtt if there already is a ggtt binding. This happens on set cache levels. Use VMA for bind/unbind (Daniel, Ben) v3: Reorganize ggtt_vma_bind to be more concise and easier to read (Ville). Change logic in unbind to only unbind ggtt when there is a global mapping, and to remove a redundant check if the aliasing ppgtt exists. v4: Make the bind function a bit smarter about the cache levels to avoid unnecessary multiple remaps. "I accept it is a wart, I think unifying the pin_vma / bind_vma could be unified later" (Chris) Removed the git notes, and put version info here. (Daniel) v5: Update the comment to not suck (Chris) v6: Move bind/unbind to the VMA. It makes more sense in the VMA structure (always has, but I was previously lazy). With this change, it will allow us to keep a distinct insert_entries. Reviewed-by:
Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by:
-
Ben Widawsky authored
Signed-off-by:
-
Ben Widawsky authored
Using the current state of the page directory registers, we can determine which of our address spaces was active when the hang occurred. This allows us to scan through all the address spaces to identify the "active" one during error capture. v2: Rebased for BDW error detection. BDW error detection is similar except instead of PP_DIR_BASE, we can use the PDP registers. Signed-off-by:
-
Ben Widawsky authored
The existing check was insufficient to determine whether we can use the GTT mapping to read out the object during error capture. The previous condition was, if the object has a GGTT mapping, and the reloc is in the GTT range... the can happen with opjects mapped into multiple vms (one of which being the GTT). There are two solutions to this problem: 1. This patch, which avoid reading the io mapping 2. Use the GGTT offset with the io mapping. Since error capture is about recording the most accurate possible error state, and the error was caused by the object not in the GGTT - I opted for the former. Signed-off-by:
-
Ben Widawsky authored
formerly: drm/i915: Create VMAs (part 6) - finish error plumbing Signed-off-by:
-
Ben Widawsky authored
This came from a patch called, "drm/i915: Move active to vma" When moving an object to the inactive list, we do it for all VMs for which the object is bound. The primary difference from that patch is this time around we don't not track 'active' per vma, but rather by object. Therefore, we only need one unref. Signed-off-by:
-
Ben Widawsky authored
This was found by code inspection. If the GTT setup fails then we are left without properly tearing down the drm_mm. Hopefully this never happens. Signed-off-by:
-
Ben Widawsky authored
To be able to effectively use the GGTT object lookup function, we don't want to warn when there is no GGTT mapping. Let the caller deal with it instead. Originally, I had intended to have this behavior, and has not introduced the WARN. It was introduced during review with the addition of the follow commit commit 5c2abbea Author: Ben Widawsky <benjamin.widawsky@intel.com> Date: Tue Sep 24 09:57:57 2013 -0700 drm/i915: Provide a cheap ggtt vma lookup Signed-off-by:
-
Ben Widawsky authored
Since the beginning, the functions which try to properly reference the aliasing PPGTT have deferences a potentially null aliasing_ppgtt member. Since the accessors are meant to be global, this will not do. Introduced originally in: commit a70a3148 Author: Ben Widawsky <ben@bwidawsk.net> Date: Wed Jul 31 16:59:56 2013 -0700 drm/i915: Make proper functions for VMs Signed-off-by:
-
