diff options
| author | Robin Murphy <robin.murphy@arm.com> | 2017-07-06 17:55:30 +0100 |
|---|---|---|
| committer | Will Deacon <will.deacon@arm.com> | 2017-07-20 10:30:27 +0100 |
| commit | 98a8f63e56a0bdcf1d0af8d840d011ab90386684 (patch) | |
| tree | ad2d756e8769b73d814d93cc818ce5f8598e0169 /drivers/iommu/mtk_iommu.c | |
| parent | 8e517e762a826d16451fb6ffb0a8722e4265582e (diff) | |
iommu/mtk: Avoid redundant TLB syncs locally
Under certain circumstances, the io-pgtable code may end up issuing two
TLB sync operations without any intervening invalidations. This goes
badly for the M4U hardware, since it means the second sync ends up
polling for a non-existent operation to finish, and as a result times
out and warns. The io_pgtable_tlb_* helpers implement a high-level
optimisation to avoid issuing the second sync at all in such cases, but
in order to work correctly that requires all pagetable operations to be
serialised under a lock, thus is no longer applicable to all io-pgtable
users.
Since we're the only user actually relying on this flag for correctness,
let's reimplement it locally to avoid the headache of trying to make the
high-level version concurrency-safe for other users.
CC: Yong Wu <yong.wu@mediatek.com>
CC: Matthias Brugger <matthias.bgg@gmail.com>
Tested-by: Yong Wu <yong.wu@mediatek.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Diffstat (limited to 'drivers/iommu/mtk_iommu.c')
| -rw-r--r-- | drivers/iommu/mtk_iommu.c | 6 |
1 files changed, 6 insertions, 0 deletions
diff --git a/drivers/iommu/mtk_iommu.c b/drivers/iommu/mtk_iommu.c index 5d14cd15198d..91c6d367ab35 100644 --- a/drivers/iommu/mtk_iommu.c +++ b/drivers/iommu/mtk_iommu.c @@ -129,6 +129,7 @@ static void mtk_iommu_tlb_add_flush_nosync(unsigned long iova, size_t size, writel_relaxed(iova, data->base + REG_MMU_INVLD_START_A); writel_relaxed(iova + size - 1, data->base + REG_MMU_INVLD_END_A); writel_relaxed(F_MMU_INV_RANGE, data->base + REG_MMU_INVALIDATE); + data->tlb_flush_active = true; } static void mtk_iommu_tlb_sync(void *cookie) @@ -137,6 +138,10 @@ static void mtk_iommu_tlb_sync(void *cookie) int ret; u32 tmp; + /* Avoid timing out if there's nothing to wait for */ + if (!data->tlb_flush_active) + return; + ret = readl_poll_timeout_atomic(data->base + REG_MMU_CPE_DONE, tmp, tmp != 0, 10, 100000); if (ret) { @@ -146,6 +151,7 @@ static void mtk_iommu_tlb_sync(void *cookie) } /* Clear the CPE status */ writel_relaxed(0, data->base + REG_MMU_CPE_DONE); + data->tlb_flush_active = false; } static const struct iommu_gather_ops mtk_iommu_gather_ops = { |