[PATCH v6 RESED 1/2] dma: replace zone_dma_bits by zone_dma_limit

Tue Aug 27 16:14:03 AEST 2024

On 27.08.2024 06:52, Baruch Siach wrote:
> Hi Marek,
>
> Thanks for your report.
>
> On Mon, Aug 26 2024, Marek Szyprowski wrote:
>> On 11.08.2024 09:09, Baruch Siach wrote:
>>> From: Catalin Marinas <catalin.marinas at arm.com>
>>>
>>> Hardware DMA limit might not be power of 2. When RAM range starts above
>>> 0, say 4GB, DMA limit of 30 bits should end at 5GB. A single high bit
>>> can not encode this limit.
>>>
>>> Use plain address for DMA zone limit.
>>>
>>> Since DMA zone can now potentially span beyond 4GB physical limit of
>>> DMA32, make sure to use DMA zone for GFP_DMA32 allocations in that case.
>>>
>>> Signed-off-by: Catalin Marinas <catalin.marinas at arm.com>
>>> Co-developed-by: Baruch Siach <baruch at tkos.co.il>
>>> Signed-off-by: Baruch Siach <baruch at tkos.co.il>
>>> ---
>> This patch landed recently in linux-next as commit ba0fb44aed47
>> ("dma-mapping: replace zone_dma_bits by zone_dma_limit"). During my
>> tests I found that it introduces the following warning on ARM64/Rockchip
>> based Odroid M1 board (arch/arm64/boot/dts/rockchip/rk3568-odroid-m1.dts):
> Does this warning go away if you revert both 3be9b846896d and ba0fb44aed47?

Yes, linux-next with above mentioned commits reverted works fine.

> Upstream rockchip DTs have no dma-ranges property. Is that the case for
> your platform as well?
>
> Can you share kernel report of DMA zones and swiotlb? On my platform I get:
>
> [    0.000000] Zone ranges:
> [    0.000000]   DMA      [mem 0x0000000800000000-0x000000083fffffff]
> [    0.000000]   DMA32    empty
> [    0.000000]   Normal   [mem 0x0000000840000000-0x0000000fffffffff]
> ...
> [    0.000000] software IO TLB: area num 8.
> [    0.000000] software IO TLB: mapped [mem 0x000000083be38000-0x000000083fe38000] (64MB)
>
> What do you get at your end?

On ba0fb44aed47 I got:

[    0.000000] NUMA: No NUMA configuration found
[    0.000000] NUMA: Faking a node at [mem 
0x0000000000200000-0x00000001ffffffff]
[    0.000000] NUMA: NODE_DATA [mem 0x1ff7a0600-0x1ff7a2fff]
[    0.000000] Zone ranges:
[    0.000000]   DMA      [mem 0x0000000000200000-0x00000001ffffffff]
[    0.000000]   DMA32    empty
[    0.000000]   Normal   empty
[    0.000000] Movable zone start for each node
[    0.000000] Early memory node ranges
[    0.000000]   node   0: [mem 0x0000000000200000-0x00000000083fffff]
[    0.000000]   node   0: [mem 0x0000000009400000-0x00000000efffffff]
[    0.000000]   node   0: [mem 0x00000001f0000000-0x00000001ffffffff]
[    0.000000] Initmem setup node 0 [mem 
0x0000000000200000-0x00000001ffffffff]
[    0.000000] On node 0, zone DMA: 512 pages in unavailable ranges
[    0.000000] On node 0, zone DMA: 4096 pages in unavailable ranges
[    0.000000] cma: Reserved 96 MiB at 0x00000001f0000000 on node -1

...

[    0.000000] software IO TLB: SWIOTLB bounce buffer size adjusted to 3MB
[    0.000000] software IO TLB: area num 4.
[    0.000000] software IO TLB: mapped [mem 
0x00000001fac00000-0x00000001fb000000] (4MB)

On the fa3c109a6d30 (parent commit of the $subject) I got:

[    0.000000] NUMA: No NUMA configuration found
[    0.000000] NUMA: Faking a node at [mem 
0x0000000000200000-0x00000001ffffffff]
[    0.000000] NUMA: NODE_DATA [mem 0x1ff7a0600-0x1ff7a2fff]
[    0.000000] Zone ranges:
[    0.000000]   DMA      [mem 0x0000000000200000-0x00000000ffffffff]
[    0.000000]   DMA32    empty
[    0.000000]   Normal   [mem 0x0000000100000000-0x00000001ffffffff]
[    0.000000] Movable zone start for each node
[    0.000000] Early memory node ranges
[    0.000000]   node   0: [mem 0x0000000000200000-0x00000000083fffff]
[    0.000000]   node   0: [mem 0x0000000009400000-0x00000000efffffff]
[    0.000000]   node   0: [mem 0x00000001f0000000-0x00000001ffffffff]
[    0.000000] Initmem setup node 0 [mem 
0x0000000000200000-0x00000001ffffffff]
[    0.000000] On node 0, zone DMA: 512 pages in unavailable ranges
[    0.000000] On node 0, zone DMA: 4096 pages in unavailable ranges
[    0.000000] cma: Reserved 96 MiB at 0x00000000ea000000 on node -1

...

[    0.000000] software IO TLB: area num 4.
[    0.000000] software IO TLB: mapped [mem 
0x00000000e6000000-0x00000000ea000000] (64MB)

It looks that for some reasons $subject patch changes the default zone 
and swiotlb configuration.

>> ------------[ cut here ]------------
>> dwmmc_rockchip fe2b0000.mmc: swiotlb addr 0x00000001faf00000+4096
>> overflow (mask ffffffff, bus limit 0).
>> WARNING: CPU: 3 PID: 1 at kernel/dma/swiotlb.c:1594 swiotlb_map+0x2f0/0x308
>> Modules linked in:
>> CPU: 3 UID: 0 PID: 1 Comm: swapper/0 Not tainted 6.11.0-rc4+ #15278
>> Hardware name: Hardkernel ODROID-M1 (DT)
>> pstate: 60400009 (nZCv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>> pc : swiotlb_map+0x2f0/0x308
>> lr : swiotlb_map+0x2f0/0x308
>> ...
>> Call trace:
>>    swiotlb_map+0x2f0/0x308
>>    dma_direct_map_sg+0x9c/0x2e4
>>    __dma_map_sg_attrs+0x28/0x94
>>    dma_map_sg_attrs+0x10/0x24
>>    dw_mci_pre_dma_transfer+0xb8/0xf4
>>    dw_mci_pre_req+0x50/0x68
>>    mmc_blk_mq_issue_rq+0x3e0/0x964
>>    mmc_mq_queue_rq+0x118/0x2b4
>>    blk_mq_dispatch_rq_list+0x21c/0x714
>>    __blk_mq_sched_dispatch_requests+0x490/0x58c
>>    blk_mq_sched_dispatch_requests+0x30/0x6c
>>    blk_mq_run_hw_queue+0x284/0x40c
>>    blk_mq_flush_plug_list.part.0+0x190/0x974
>>    blk_mq_flush_plug_list+0x1c/0x2c
>>    __blk_flush_plug+0xe4/0x140
>>    blk_finish_plug+0x38/0x4c
>>    __ext4_get_inode_loc+0x22c/0x654
>>    __ext4_get_inode_loc_noinmem+0x40/0xa8
>>    __ext4_iget+0x154/0xcc0
>>    ext4_get_journal_inode+0x30/0x110
>>    ext4_load_and_init_journal+0x9c/0xaf0
>>    ext4_fill_super+0x1fec/0x2d90
>>    get_tree_bdev+0x140/0x1d8
>>    ext4_get_tree+0x18/0x24
>>    vfs_get_tree+0x28/0xe8
>>    path_mount+0x3e8/0xb7c
>>    init_mount+0x68/0xac
>>    do_mount_root+0x108/0x1dc
>>    mount_root_generic+0x100/0x330
>>    mount_root+0x160/0x2d0
>>    initrd_load+0x1f0/0x2a0
>>    prepare_namespace+0x4c/0x29c
>>    kernel_init_freeable+0x4b4/0x50c
>>    kernel_init+0x20/0x1d8
>>    ret_from_fork+0x10/0x20
>> irq event stamp: 1305682
>> hardirqs last  enabled at (1305681): [<ffff8000800e332c>]
>> console_unlock+0x124/0x130
>> hardirqs last disabled at (1305682): [<ffff80008124e684>] el1_dbg+0x24/0x8c
>> softirqs last  enabled at (1305678): [<ffff80008005be1c>]
>> handle_softirqs+0x4cc/0x4e4
>> softirqs last disabled at (1305665): [<ffff8000800105b0>]
>> __do_softirq+0x14/0x20
>> ---[ end trace 0000000000000000 ]---
>>
>> This "bus limit 0" seems to be a bit suspicious to me as well as the
>> fact that swiotlb is used for the MMC DMA. I will investigate this
>> further tomorrow. The board boots fine though.
> Looking at the code I guess that bus_dma_limit set to 0 means no bus
> limit. But dma_mask for your device indicates 32-bit device limit. This
> can't work with address above 4GB. For some reason DMA code tries to
> allocate from higher address. This is most likely the reason
> dma_capable() returns false.

Indeed this looks like a source of the problem:

[    3.123618] Synopsys Designware Multimedia Card Interface Driver
[    3.139653] dwmmc_rockchip fe2b0000.mmc: IDMAC supports 32-bit 
address mode.
[    3.147739] dwmmc_rockchip fe2b0000.mmc: Using internal DMA controller.
[    3.161659] dwmmc_rockchip fe2b0000.mmc: Version ID is 270a
[    3.168455] dwmmc_rockchip fe2b0000.mmc: DW MMC controller at irq 
56,32 bit host data width,256 deep fifo
[    3.182651] dwmmc_rockchip fe2b0000.mmc: Got CD GPIO

...

[   11.009258] ------------[ cut here ]------------
[   11.014762] dwmmc_rockchip fe2b0000.mmc: swiotlb addr 
0x00000001faf00000+4096 overflow (mask ffffffff, bus limit 0).

> ...

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland