This is the first chunk of the new VMM code that provides the structures
needed to describe a GPU virtual address-space layout, as well as common
interfaces to handle VMM creation, and connecting instances to a VMM.
The constructor now allocates the PD itself, rather than having the user
handle that manually. This won't/can't be used until after all backends
have been ported to these interfaces, so a little bit of memory will be
wasted on Fermi and newer for a couple of commits in the series.
Compatibility has been hacked into the old code to allow each GPU backend
to be ported individually.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
GP100 "big" (which is a funny name, when it supports "even bigger") page
tables are small enough that we want to be able to suballocate them from
a larger block of memory.
This builds on the previous page table cache interfaces so that the VMM
code doesn't need to know the difference.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Builds up and maintains a small cache of each page table size in order
to reduce the frequency of expensive allocations, particularly in the
pathological case where an address range ping-pongs between allocated
and free.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Removes the need to expose internals outside of MMU, and GP100 is both
different, and a lot harder to deal with.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Another transition step to allow finer-grained patches transitioning to
new MMU backends.
Old backends will continue operate as before (accessing nvkm_mem::tag),
and new backends will get a reference to the tags allocated here.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Upcoming MMU changes use nvkm_memory as its basic representation of memory,
so we need to be able to allocate VRAM like this.
The code is basically identical to the current chipset-specific allocators,
minus support for compression tags (which will be handled elsewhere anyway).
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
We need to be able to prevent memory from being freed while it's still
mapped in a GPU's address-space.
Will be used by upcoming MMU changes.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Map flags (access, kind, etc) are currently defined in either the VMA,
or the memory object, which turns out to not be ideal for things like
suballocated buffers, etc.
These will become per-map flags instead, so we need to support passing
these arguments in nvkm_memory_map().
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
We're moving towards having a central place to handle comptag allocation,
and as some GPUs don't have a ram submodule (ie. Tegra), we need to move
the mm somewhere else.
It probably never belonged in ram anyways.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Different sections of VRAM may have different properties (ie. can't be used
for compression/display, can't be mapped, etc).
We currently already support this, but it's a bit magic. This change makes
it more obvious where we're allocating from.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Before: "imem: init completed in 299277us"
After: "imem: init completed in 11574us"
Suspend from Fedora 26 gnome desktop on GP102.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Before: "imem: suspend completed in 5540487us"
After: "imem: suspend completed in 1871526us"
Suspend from Fedora 26 gnome desktop on GP102.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
A good deal of the structures we map into here aren't accessed very often
at all, and Fedora 26 has exposed an issue where after creating a heap of
channels, BAR2 space would run out, and we'd need to make use of the slow
path while accessing important structures like page tables.
This implements an LRU on BAR2 space, which allows eviction of mappings
that aren't currently needed, to make space for other objects.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Another piece of solving the "GP100 BAR2 VMM bootstrap" puzzle.
Without doing this, we'd attempt to write PDEs for the lower page table
levels through BAR2 before BAR2 access has been fully initialised.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This is not as simple as it was for earlier GPUs, due to the need to swap
accessor functions depending on whether BAR2 is usable or not.
We were previously protected by nvkm_instobj's accessor functions keeping
an object mapped permanently, with some unclear magic that managed to hit
the slow-path where needed even if an object was marked as mapped.
That's been replaced here by reference counting maps (some objects, like
page tables can be accessed concurrently), and swapping the functions as
necessary.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
This is to simplify upcoming changes. The slow-path is something that
currently occurs during bootstrap of the BAR2 VMM, while backing up an
object during suspend/resume, or when BAR2 address space runs out.
The latter is a real problem that can happen at runtime, and occurs in
Fedora 26 already (due to some change that causes a lot of channels to
be created at login), so ideally we'd prefer not to make it any slower.
We'd also like suspend/resume speed to not suffer.
Upcoming commits will solve those problems in a better way, making the
extra overhead of moving the locking here a non-issue.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
The accessor functions can change as a result of acquire()/release() calls,
and are protected by any refcounting done there.
Other functions must remain constant, as they can be called any time.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
