Both changes in dc97b3409a cause serious
regressions in the nouveau driver.
move_notify() was originally able to presume that bo->mem is the old node,
and new_mem is the new node. The above commit moves the call to
move_notify() to after move() has been done, which means that now, sometimes,
new_mem isn't the new node at all, bo->mem is, and new_mem points at a
stale, possibly-just-been-killed-by-move node.
This is clearly not a good situation. This patch reverts this change, and
replaces it with a cleanup in the move() failure path instead.
The second issue is that the call to move_notify() from cleanup_memtype_use()
causes the TTM ghost objects to get passed into the driver. This is clearly
bad as the driver knows nothing about these "fake" TTM BOs, and ends up
accessing uninitialised memory.
I worked around this in nouveau's move_notify() hook by ensuring the BO
destructor was nouveau's. I don't particularly like this solution, and
would rather TTM never pass the driver these objects. However, I don't
clearly understand the reason why we're calling move_notify() here anyway
and am happy to work around the problem in nouveau instead of breaking the
behaviour expected by other drivers.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Jerome Glisse <j.glisse@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
. It was useful during development, but now on a production system
we can get this (if the user forgot to upload the firmware):
[drm] radeon: irq initialized.
[drm] GART: num cpu pages 131072, num gpu pages 131072
[drm] radeon: ib pool ready.
[drm] Loading SUMO Microcode
r600_cp: Failed to load firmware "radeon/SUMO_pfp.bin"
atl1c 0000:03:00.0: version 1.0.1.0-NAPI.213057] [drm:evergreen_startup] *ERROR* Failed to load firmware!
radeon 0000:00:01.0: disabling GPU acceleration
88] radeon 0000:00:01.0: ffff8801bb782400 unpin not necessary
------------[ cut here ]------------
WARNING: at /home/konrad/linux-linus/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c:956 ttm_dma_unpopulate+0x79/0x300 [ttm]()
Hardware name: System Product Name
Modules linked in: e1000e atl1c radeon(+) ahci libahci libata scsi_mod fbcon tileblit font ttm bitblit softcursor drm_kms_helper wmi xen_blkfront xen_netfront fb_sys_fops sysimgblt sysfillrect syscopyarea xenfs xen_privcmd
Pid: 1600, comm: modprobe Not tainted 3.2.0-06100-ge343a89 #1
Call Trace:
[<ffffffff8108973a>] warn_slowpath_common+0x7a/0xb0
[<ffffffff81089785>] warn_slowpath_null+0x15/0x20
[<ffffffffa0060309>] ttm_dma_unpopulate+0x79/0x300 [ttm]
[<ffffffffa01341c0>] radeon_ttm_tt_unpopulate+0x120/0x130 [radeon]
[<ffffffffa0056e0c>] ttm_tt_destroy+0x2c/0x70 [ttm]
[<ffffffffa0057a4e>] ttm_bo_cleanup_memtype_use+0x3e/0x80 [ttm]
[<ffffffffa00595a1>] ttm_bo_release+0x251/0x280 [ttm]
[<ffffffffa0059610>] ttm_bo_unref+0x40/0x60 [ttm]
[<ffffffffa0134d02>] radeon_bo_unref+0x42/0x80 [radeon]
[<ffffffffa0186dfb>] radeon_sa_bo_manager_fini+0x6b/0x80 [radeon]
[<ffffffffa0146b8f>] radeon_ib_pool_fini+0x6f/0x90 [radeon]
[<ffffffffa014be49>] r100_ib_fini+0x19/0x20 [radeon]
[<ffffffffa017b47e>] evergreen_init+0x1ee/0x2d0 [radeon]
The big WARN() has nothing to do with the culprit - which is that
the firmware was not loaded. So lets remove the WARN() from the TTM DMA code.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The "if (!p && !p->dev)" condition isn't right because || was intended
instead of &&. But actually, "p" is the list cursor and so it's always
non-NULL and we can just remove that bit. We can remove the another
similar check as well.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Acked-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
ttm tt rework modified the way we allocate and populate the
ttm_tt structure, the AGP side was missing some bit to properly
work. Fix those and fix radeon and nouveau AGP support.
Tested on radeon only so far.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The code to figure out how many pages to shrink the pool
ends up capping the 'count' at _manager->options.max_size - which is OK.
Except that the 'count' is also used when accounting for how many pages
are recycled - which we end up with the invalid values. This fixes
it by using a different value for the amount of pages to shrink.
On top of that we would free the cached page pool - which is nonsense
as they are deleted from the pool - so there are no free pages in that
pool..
Also we also missed the opportunity to batch the amount of pages
to free (similar to how ttm_page_alloc.c does it). This reintroduces
the code that was lost during rebasing.
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Previously we were calling back move_notify in error path when the
bo is returned to it's original position or when destroy the bo.
When destroying the bo set the new mem placement as NULL when calling
back in the driver.
Updating nouveau to deal with NULL placement properly.
v2: reserve the object before calling move_notify in bo destroy path
at that point ttm should be the only piece of code interacting
with the object so atomic_set is safe here.
v3: callback move notify only once the bo is in its new position
call move notify want swaping out the buffer
v4:- don't call move_notify when swapin out bo, assume driver should
do what is appropriate in swap notify
- move move_notify call back to ttm_bo_cleanup_memtype_use for
destroy path
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Provide helper function to compute the kernel memory size needed
for each buffer object. Move all the accounting inside ttm, simplifying
driver and avoiding code duplication accross them.
v2 fix accounting of ghost object, one would have thought that i
would have run into the issue since a longtime but it seems
ghost object are rare when you have plenty of vram ;)
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Move dma data to a superset ttm_dma_tt structure which herit
from ttm_tt. This allow driver that don't use dma functionalities
to not have to waste memory for it.
V2 Rebase on top of no memory account changes (where/when is my
delorean when i need it ?)
V3 Make sure page list is initialized empty
V4 typo/syntax fixes
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
In TTM world the pages for the graphic drivers are kept in three different
pools: write combined, uncached, and cached (write-back). When the pages
are used by the graphic driver the graphic adapter via its built in MMU
(or AGP) programs these pages in. The programming requires the virtual address
(from the graphic adapter perspective) and the physical address (either System RAM
or the memory on the card) which is obtained using the pci_map_* calls (which does the
virtual to physical - or bus address translation). During the graphic application's
"life" those pages can be shuffled around, swapped out to disk, moved from the
VRAM to System RAM or vice-versa. This all works with the existing TTM pool code
- except when we want to use the software IOTLB (SWIOTLB) code to "map" the physical
addresses to the graphic adapter MMU. We end up programming the bounce buffer's
physical address instead of the TTM pool memory's and get a non-worky driver.
There are two solutions:
1) using the DMA API to allocate pages that are screened by the DMA API, or
2) using the pci_sync_* calls to copy the pages from the bounce-buffer and back.
This patch fixes the issue by allocating pages using the DMA API. The second
is a viable option - but it has performance drawbacks and potential correctness
issues - think of the write cache page being bounced (SWIOTLB->TTM), the
WC is set on the TTM page and the copy from SWIOTLB not making it to the TTM
page until the page has been recycled in the pool (and used by another application).
The bounce buffer does not get activated often - only in cases where we have
a 32-bit capable card and we want to use a page that is allocated above the
4GB limit. The bounce buffer offers the solution of copying the contents
of that 4GB page to an location below 4GB and then back when the operation has been
completed (or vice-versa). This is done by using the 'pci_sync_*' calls.
Note: If you look carefully enough in the existing TTM page pool code you will
notice the GFP_DMA32 flag is used - which should guarantee that the provided page
is under 4GB. It certainly is the case, except this gets ignored in two cases:
- If user specifies 'swiotlb=force' which bounces _every_ page.
- If user is using a Xen's PV Linux guest (which uses the SWIOTLB and the
underlaying PFN's aren't necessarily under 4GB).
To not have this extra copying done the other option is to allocate the pages
using the DMA API so that there is not need to map the page and perform the
expensive 'pci_sync_*' calls.
This DMA API capable TTM pool requires for this the 'struct device' to
properly call the DMA API. It also has to track the virtual and bus address of
the page being handed out in case it ends up being swapped out or de-allocated -
to make sure it is de-allocated using the proper's 'struct device'.
Implementation wise the code keeps two lists: one that is attached to the
'struct device' (via the dev->dma_pools list) and a global one to be used when
the 'struct device' is unavailable (think shrinker code). The global list can
iterate over all of the 'struct device' and its associated dma_pool. The list
in dev->dma_pools can only iterate the device's dma_pool.
/[struct device_pool]\
/---------------------------------------------------| dev |
/ +-------| dma_pool |
/-----+------\ / \--------------------/
|struct device| /-->[struct dma_pool for WC]</ /[struct device_pool]\
| dma_pools +----+ /-| dev |
| ... | \--->[struct dma_pool for uncached]<-/--| dma_pool |
\-----+------/ / \--------------------/
\----------------------------------------------/
[Two pools associated with the device (WC and UC), and the parallel list
containing the 'struct dev' and 'struct dma_pool' entries]
The maximum amount of dma pools a device can have is six: write-combined,
uncached, and cached; then there are the DMA32 variants which are:
write-combined dma32, uncached dma32, and cached dma32.
Currently this code only gets activated when any variant of the SWIOTLB IOMMU
code is running (Intel without VT-d, AMD without GART, IBM Calgary and Xen PV
with PCI devices).
Tested-by: Michel Dänzer <michel@daenzer.net>
[v1: Using swiotlb_nr_tbl instead of swiotlb_enabled]
[v2: Major overhaul - added 'inuse_list' to seperate used from inuse and reorder
the order of lists to get better performance.]
[v3: Added comments/and some logic based on review, Added Jerome tag]
[v4: rebase on top of ttm_tt & ttm_backend merge]
[v5: rebase on top of ttm memory accounting overhaul]
[v6: New rebase on top of more memory accouting changes]
[v7: well rebase on top of no memory accounting changes]
[v8: make sure pages list is initialized empty]
[v9: calll ttm_mem_global_free_page in unpopulate for accurate accountg]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Acked-by: Thomas Hellstrom <thellstrom@vmware.com>
Move the page allocation and freeing to driver callback and
provide ttm code helper function for those.
Most intrusive change, is the fact that we now only fully
populate an object this simplify some of code designed around
the page fault design.
V2 Rebase on top of memory accounting overhaul
V3 New rebase on top of more memory accouting changes
V4 Rebase on top of no memory account changes (where/when is my
delorean when i need it ?)
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
ttm_backend will only exist with a ttm_tt, and ttm_tt
will only be of interest when bound to a backend. Merge them
to avoid code and data duplication.
V2 Rebase on top of memory accounting overhaul
V3 Rebase on top of more memory accounting changes
V4 Rebase on top of no memory account changes (where/when is my
delorean when i need it ?)
V5 make sure ttm is unbound before destroying, change commit
message on suggestion from Tormod Volden
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Use the ttm_tt pages array for pages allocations, move the list
unwinding into the page allocation functions.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
On failure we need to make sure the page we free has wb cache
attribute. Do this pas call the proper ttm page helper function.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Split btw highmem and lowmem page was rendered useless by the
pool code. Remove it. Note further cleanup would change the
ttm page allocation helper to actualy take an array instead
of relying on list this could drasticly reduce the number of
function call in the common case of allocation whole buffer.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
This was never use in none of the driver, properly using userspace
page for bo would need more code (vma interaction mostly). Removing
this dead code in preparation of ttm_tt & backend merge.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
An unlikely race could case a bo to be returned reserved on an error path.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This reverts commit dfadbbdb57.
Further upstream discussion between Marek and Thomas decided this wasn't
fully baked and needed further work, so revert it before it hits mainline.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Fixes an information leak to userspace, we were handing out un-zeroed pages
for any newly created TTM_PL_TT buffer.
Reported-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Tested-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Cc: stable@kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
Sometimes we want to know whether a buffer is busy and wait for it (bo_wait).
However, sometimes it would be more useful to be able to query whether
a buffer is busy and being either read or written, and wait until it's stopped
being either read or written. The point of this is to be able to avoid
unnecessary waiting, e.g. if a GPU has written something to a buffer and is now
reading that buffer, and a CPU wants to map that buffer for read, it needs to
only wait for the last write. If there were no write, there wouldn't be any
waiting needed.
This, or course, requires user space drivers to send read/write flags
with each relocation (like we have read/write domains in radeon, so we can
actually use those for something useful now).
Now how this patch works:
The read/write flags should passed to ttm_validate_buffer. TTM maintains
separate sync objects of the last read and write for each buffer, in addition
to the sync object of the last use of a buffer. ttm_bo_wait then operates
with one the sync objects.
Signed-off-by: Marek Olšák <maraeo@gmail.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This was true for new TTM_PL_SYSTEM and new TTM_PL_TT cases, but wasn't
the case on TTM_PL_SYSTEM<->TTM_PL_TT moves, which causes trouble on some
paths as nouveau's move_notify() hook requires that the dma addresses be
valid at this point.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Nouveau makes the assumption that if a TTM is bound there will be a mm_node
around for it and the backwards ordering here resulted in a use-after-free
on some eviction paths.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
ttm_tt_destroy kfrees passed object, so we need to nullify
a reference to it.
Signed-off-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Cc: stable@kernel.org
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This allows us to move duplicated code in <asm/atomic.h>
(atomic_inc_not_zero() for now) to <linux/atomic.h>
Signed-off-by: Arun Sharma <asharma@fb.com>
Reviewed-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: Mike Frysinger <vapier@gentoo.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Soon tmpfs will stop supporting ->readpage and read_mapping_page(): once
"tmpfs: add shmem_read_mapping_page_gfp" has been applied, this patch can
be applied to ease the transition.
ttm_tt_swapin() and ttm_tt_swapout() use shmem_read_mapping_page() in
place of read_mapping_page(), since their swap_space has been created with
shmem_file_setup().
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Dave Airlie <airlied@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
. and some comments to make it easier to understand.
Ackedby: Randy Dunlap <randy.dunlap@oracle.com>
[v2: Added some more updates from Randy Dunlap]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Change each shrinker's API by consolidating the existing parameters into
shrink_control struct. This will simplify any further features added w/o
touching each file of shrinker.
[akpm@linux-foundation.org: fix build]
[akpm@linux-foundation.org: fix warning]
[kosaki.motohiro@jp.fujitsu.com: fix up new shrinker API]
[akpm@linux-foundation.org: fix xfs warning]
[akpm@linux-foundation.org: update gfs2]
Signed-off-by: Ying Han <yinghan@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Acked-by: Pavel Emelyanov <xemul@openvz.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: Rik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This reverts commit 69a07f0b11.
We've tracked a number of problems back to this, and Thomas
thinks we should redesign this for .40/41 anyways so I'm
happy to revert it.
Signed-off-by: Dave Airlie <airlied@redhat.com>
There's no need to pass kref_put() the address of a function (just the
function will do just fine) nor to cast its unused return to void.
Signed-off-by: Paul Bolle <pebolle@tiscali.nl>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
This reverts commit 5a893fc28f.
This causes a use after free in the ttm free alloc pages path,
when it tries to get the be after the be has been destroyed.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Nouveau doesn't have enough information at ttm_backend_func.bind() time
to implement things like tiled GART, or to keep a buffer at a constant
address in the GPU virtual address space no matter where in physical
memory it's placed.
To resolve this, nouveau will handle binding of all buffers to the GPU
itself from the move_notify() hook. This commit ensures it's called
for all buffer moves.
Talked to Dave about the impact on radeon, which uses move_notify, it
doesn't look like anything should break there.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Reviewed-by: Thomas Hellstrom <thomas@shipmail.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This makes the accounting when using 'debug_dma_dump_mappings()'
and CONFIG_DMA_API_DEBUG=y be assigned to the correct device
instead of 'fallback'.
No functional change - just cosmetic.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
We pass in the array of ttm pages to be populated in the GART/MM
of the card (or AGP). Patch titled: "ttm: Utilize the DMA API for
pages that have TTM_PAGE_FLAG_DMA32 set." uses the DMA API to make
those pages have a proper DMA addresses (in the situation where
page_to_phys or virt_to_phys do not give use the DMA (bus) address).
Since we are using the DMA API on those pages, we should pass in the
DMA address to this function so it can save it in its proper fields
(later patches use it).
[v2: Added reviewed-by tag]
Reviewed-by: Thomas Hellstrom <thellstrom@shipmail.org>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Tested-by: Ian Campbell <ian.campbell@citrix.com>
For pages that have the TTM_PAGE_FLAG_DMA32 flag set we
use the DMA API. We save the bus address in our array which we
use to program the GART (see "radeon/ttm/PCIe: Use dma_addr if TTM
has set it." and "nouveau/ttm/PCIe: Use dma_addr if TTM has set it.").
The reason behind using the DMA API is that under Xen we would
end up programming the GART with the bounce buffer (SWIOTLB)
DMA address instead of the physical DMA address of the TTM page.
The reason being that alloc_page with GFP_DMA32 does not allocate
pages under the the 4GB mark when running under Xen hypervisor.
Under baremetal this means we do the DMA API call earlier instead
of when we program the GART.
For details please refer to:
https://lkml.org/lkml/2011/1/7/251
[v2: Fixed indentation, revised desc, added Reviewed-by]
Reviewed-by: Thomas Hellstrom <thomas@shipmail.org>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Tested-by: Ian Campbell <ian.campbell@citrix.com>
This is right now limited to only non-pool constructs.
[v2: Fixed indentation issues, add review-by tag]
Reviewed-by: Thomas Hellstrom <thomas@shipmail.org>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Tested-by: Ian Campbell <ian.campbell@citrix.com>
Make ttm_bo::ttm_bo_device_release call cancel_delayed_work_sync()
instead of calling cancel_delayed_work() followed by
flush_scheduled_work().
This is to prepare for the deprecation and removal of
flush_scheduled_work().
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc:: Thomas Hellstrom <thellstrom@vmware.com>
Cc:: Dave Airlie <airlied@redhat.com>
Drivers using their own implementation of io_mem_reserve/io_mem_free are
likely to store the tracking information for the map in mem.mm_node, so
it can't be freed while still mapped.
Signed-off-by: Ben Skeggs<bskeggs@redhat.com>
Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This patch attempts to fix up shortcomings with the current calling
sequences.
1) There's a fastpath where no locking occurs and only io_mem_reserved is
called to obtain needed info for mapping. The fastpath is set per
memory type manager.
2) If the fastpath is disabled, io_mem_reserve and io_mem_free will be exactly
balanced and not called recursively for the same struct ttm_mem_reg.
3) Optionally the driver can choose to enable a per memory type manager LRU
eviction mechanism that, when io_mem_reserve returns -EAGAIN will attempt
to kill user-space mappings of memory in that manager to free up needed
resources
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Ben Skeggs <bskeggs@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Rather than having the driver supply the validation sequence, leave that
responsibility to TTM. This saves some confusion and a function argument.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Drastically reduce the number of spin lock / unlock operations by performing
unreserving and fencing under global locks.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jerome Glisse <j.glisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The bo lock used only to protect the bo sync object members, and since it
is a per bo lock, fencing a buffer list will see a lot of locks and unlocks.
Replace it with a per-device lock that protects the sync object members on
*all* bos. Reading and setting these members will always be very quick, so
the risc of heavy lock contention is microscopic. Note that waiting for
sync objects will always take place outside of this lock.
The bo device fence lock will eventually be replaced with a seqlock /
rcu mechanism so we can determine that a bo is idle under a
rcu / read seqlock.
However this change will allow us to batch fencing and unreserving of
buffers with a minimal amount of locking.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jerome Glisse <j.glisse@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Add an aid for the driver to detect deadlocks on multi-bo reservations
Update documentation.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jerome Glisse <j.glisse@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Avoid the ttm_bo_unreserve() spinlocks by calling
ttm_eu_backoff_reservation_locked under the lru spinlock.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jerome Glisse <j.glisse@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Makes it possible to reserve a list of buffer objects with a single
spin lock / unlock if there is no contention.
Should improve cpu usage on SMP kernels.
v2: Initialize private list members on reserve and don't call
ttm_bo_list_ref_sub() with zero put_count.
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Ben Skeggs