linux-wireless.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [PATCH 0/4] wcn36xx: Fix DMA buffer allocation and free logic
@ 2021-10-15 13:17 Bryan O'Donoghue
  2021-10-15 13:17 ` [PATCH 1/4] wcn36xx: Fix DXE lock layering violation Bryan O'Donoghue
                   ` (3 more replies)
  0 siblings, 4 replies; 6+ messages in thread
From: Bryan O'Donoghue @ 2021-10-15 13:17 UTC (permalink / raw)
  To: kvalo, linux-wireless, wcn36xx
  Cc: loic.poulain, benl, daniel.thompson, bryan.odonoghue

Digging around through some bugs reported from an extensive testing cycle
we've found that wcn36xx has a number of unexplained RX related oopses.

In at least one case we appear to have DMA'd data to an unmapped region.
The written data appears to be a correctly formed DMA buffer descriptor - a
DXE BD in WCNSS parlance, with an AP beacon inside of it.

Reasoning about how such a situation might come about and reviewing the
run-time code, there was no obvious path where we might free a BD or an
skbuff pointed to by a BD, which DXE might not be aware of.

However looking at the ieee80211_ops.start and ieee80211_ops.stop callbacks
in wcn36xx we can see a number of bugs associated with BD allocation, error
handling and leaving the DMA engine active, despite freeing SKBs on the MSM
side.

This last mention - failure to quiesce potential DMA from the downstream
agent - WCNSS DXE despite freeing the memory @ the skbuffs is a decent
candidate for our unexplained upstream DMA transaction to unmapped memory.

Since wcn36xx_stop and wcn36xx_start can be called a number of times by the
controlling upper layers it means there is a potential gap between
wcn36xx_stop and wcn36xx_start which could leave WCNSS in a state where it
will try to DMA to memory we have freed.

This series addresses the obvious bugs that jump out on the start()/stop()
path.

Patch #1
  In order to make it easier to read the DXE code, I've moved all of the
  lock taking and freeing for DXE into dxe.c

Patch #2
  Fixes a very obviously broken channel enable/disable cycle

Patch #3
  Fixes a very obvious memory leak with dma_alloc_coherent()

Patch #4
  Makes sure before we release skbuffs which we assigned to the RX channels
  that we ensure the DXE block is put into reset

Bryan O'Donoghue (4):
  wcn36xx: Fix DXE lock layering violation
  wcn36xx: Fix DXE/DMA channel enable/disable cycle
  wcn36xx: Release DMA channel descriptor allocations
  wcn36xx: Put DXE block into reset before freeing memory

 drivers/net/wireless/ath/wcn36xx/dxe.c  | 75 +++++++++++++++++++++----
 drivers/net/wireless/ath/wcn36xx/dxe.h  |  2 +
 drivers/net/wireless/ath/wcn36xx/txrx.c | 15 +----
 3 files changed, 68 insertions(+), 24 deletions(-)

-- 
2.33.0


^ permalink raw reply	[flat|nested] 6+ messages in thread
end of thread, other threads:[~2021-10-17  1:26 UTC | newest]

Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-10-15 13:17 [PATCH 0/4] wcn36xx: Fix DMA buffer allocation and free logic Bryan O'Donoghue
2021-10-15 13:17 ` [PATCH 1/4] wcn36xx: Fix DXE lock layering violation Bryan O'Donoghue
2021-10-15 13:17 ` [PATCH 2/4] wcn36xx: Fix DXE/DMA channel enable/disable cycle Bryan O'Donoghue
2021-10-15 13:17 ` [PATCH 3/4] wcn36xx: Release DMA channel descriptor allocations Bryan O'Donoghue
2021-10-15 13:17 ` [PATCH 4/4] wcn36xx: Put DXE block into reset before freeing memory Bryan O'Donoghue
2021-10-17  1:28   ` Bryan O'Donoghue

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).