Message ID | 1448908321-3042-1-git-send-email-nbd@openwrt.org (mailing list archive) |
---|---|
State | Not Applicable |
Delegated to: | Kalle Valo |
Headers | show |
Felix Fietkau <nbd@openwrt.org> writes: > Coherent memory is more expensive to allocate (and constrained on some > architectures where it has to be pre-allocated). It is also completely > unnecessary, since the host has no reason to even access these allocated > memory spaces > > Signed-off-by: Felix Fietkau <nbd@openwrt.org> Applied, thanks.
On 11/30/2015 10:32 AM, Felix Fietkau wrote: > Coherent memory is more expensive to allocate (and constrained on some > architectures where it has to be pre-allocated). It is also completely > unnecessary, since the host has no reason to even access these allocated > memory spaces > > Signed-off-by: Felix Fietkau <nbd@openwrt.org> > - memset(ar->wmi.mem_chunks[idx].vaddr, 0, pool_size); > + if (!num_units) > + return -ENOMEM; > + > + paddr = dma_map_single(ar->dev, vaddr, pool_size, DMA_TO_DEVICE); > + if (dma_mapping_error(ar->dev, paddr)) { > + kfree(vaddr); > + return -ENOMEM; > + } Are you sure you have the direction correct for the 'dma_map_single' call? I thought this memory was for the NIC to scribble in, and probably host should never even bother reading or writing it? The reason I started looking at this is that I see these errors when trying to use a 4x4 wave-2 ath10k NIC in a 4.4.0-rc7 ath kernel: [ 202.489625] wlan0: authenticate with 00:0e:8e:f8:73:96 [ 202.784533] DMAR: DRHD: handling fault status reg 3 [ 202.786246] wlan0: send auth to 00:0e:8e:f8:73:96 (try 1/3) [ 202.788133] DMAR: DMAR:[DMA Write] Request device [05:00.0] fault addr ff5de000 DMAR:[fault reason 05] PTE Write access is not set [ 202.887410] wlan0: send auth to 00:0e:8e:f8:73:96 (try 2/3) [ 202.988423] wlan0: send auth to 00:0e:8e:f8:73:96 (try 3/3) [ 203.089437] wlan0: authentication with 00:0e:8e:f8:73:96 timed out This is on an Intel x86-64 system with IOMMU (VT-d) enabled. I'm likely having more than one problem since a 4.3.0-rc6+ kernel is not working with this NIC either... Thanks, Ben
diff --git a/drivers/net/wireless/ath/ath10k/wmi.c b/drivers/net/wireless/ath/ath10k/wmi.c index 9021079..1386dd8 100644 --- a/drivers/net/wireless/ath/ath10k/wmi.c +++ b/drivers/net/wireless/ath/ath10k/wmi.c @@ -4300,34 +4300,58 @@ void ath10k_wmi_event_vdev_resume_req(struct ath10k *ar, struct sk_buff *skb) ath10k_dbg(ar, ATH10K_DBG_WMI, "WMI_VDEV_RESUME_REQ_EVENTID\n"); } -static int ath10k_wmi_alloc_host_mem(struct ath10k *ar, u32 req_id, - u32 num_units, u32 unit_len) +static int ath10k_wmi_alloc_chunk(struct ath10k *ar, u32 req_id, + u32 num_units, u32 unit_len) { dma_addr_t paddr; - u32 pool_size; + u32 pool_size = 0; int idx = ar->wmi.num_mem_chunks; + void *vaddr = NULL; - pool_size = num_units * round_up(unit_len, 4); + if (ar->wmi.num_mem_chunks == ARRAY_SIZE(ar->wmi.mem_chunks)) + return -ENOMEM; - if (!pool_size) - return -EINVAL; + while (!vaddr && num_units) { + pool_size = num_units * round_up(unit_len, 4); + if (!pool_size) + return -EINVAL; - ar->wmi.mem_chunks[idx].vaddr = dma_alloc_coherent(ar->dev, - pool_size, - &paddr, - GFP_KERNEL); - if (!ar->wmi.mem_chunks[idx].vaddr) { - ath10k_warn(ar, "failed to allocate memory chunk\n"); - return -ENOMEM; + vaddr = kzalloc(pool_size, GFP_KERNEL | __GFP_NOWARN); + if (!vaddr) + num_units /= 2; } - memset(ar->wmi.mem_chunks[idx].vaddr, 0, pool_size); + if (!num_units) + return -ENOMEM; + + paddr = dma_map_single(ar->dev, vaddr, pool_size, DMA_TO_DEVICE); + if (dma_mapping_error(ar->dev, paddr)) { + kfree(vaddr); + return -ENOMEM; + } + ar->wmi.mem_chunks[idx].vaddr = vaddr; ar->wmi.mem_chunks[idx].paddr = paddr; ar->wmi.mem_chunks[idx].len = pool_size; ar->wmi.mem_chunks[idx].req_id = req_id; ar->wmi.num_mem_chunks++; + return num_units; +} + +static int ath10k_wmi_alloc_host_mem(struct ath10k *ar, u32 req_id, + u32 num_units, u32 unit_len) +{ + int ret; + + while (num_units) { + ret = ath10k_wmi_alloc_chunk(ar, req_id, num_units, unit_len); + if (ret < 0) + return ret; + + num_units -= ret; + } + return 0; } @@ -7705,10 +7729,11 @@ void ath10k_wmi_free_host_mem(struct ath10k *ar) /* free the host memory chunks requested by firmware */ for (i = 0; i < ar->wmi.num_mem_chunks; i++) { - dma_free_coherent(ar->dev, - ar->wmi.mem_chunks[i].len, - ar->wmi.mem_chunks[i].vaddr, - ar->wmi.mem_chunks[i].paddr); + dma_unmap_single(ar->dev, + ar->wmi.mem_chunks[i].paddr, + ar->wmi.mem_chunks[i].len, + DMA_TO_DEVICE); + kfree(ar->wmi.mem_chunks[i].vaddr); } ar->wmi.num_mem_chunks = 0;
Coherent memory is more expensive to allocate (and constrained on some architectures where it has to be pre-allocated). It is also completely unnecessary, since the host has no reason to even access these allocated memory spaces Signed-off-by: Felix Fietkau <nbd@openwrt.org> --- drivers/net/wireless/ath/ath10k/wmi.c | 61 ++++++++++++++++++++++++----------- 1 file changed, 43 insertions(+), 18 deletions(-)