Re: [RESEND PATCH v3 5/8] mtd: spi-nor: cadence-quadspi: Handle probe deferral while requesting DMA channel

From: Vignesh Raghavendra
Date: Wed Aug 26 2020 - 08:18:51 EST




On 8/26/20 3:42 PM, Jan Kiszka wrote:
> On 24.08.20 19:20, Jan Kiszka wrote:
>> On 24.08.20 14:49, Jan Kiszka wrote:
>>> On 24.08.20 13:45, Vignesh Raghavendra wrote:
>>>>
>>>>
>>>> On 8/22/20 11:35 PM, Jan Kiszka wrote:
>>>>> On 01.06.20 09:04, Vignesh Raghavendra wrote:
>>>>>> dma_request_chan_by_mask() can throw EPROBE_DEFER if DMA provider
>>>>>> is not yet probed. Currently driver just falls back to using PIO mode
>>>>>> (which is less efficient) in this case. Instead return probe deferral
>>>>>> error as is so that driver will be re probed once DMA provider is
>>>>>> available.
>>>>>>
>>>>>> Signed-off-by: Vignesh Raghavendra <vigneshr@xxxxxx>
>>>>>> Reviewed-by: Tudor Ambarus <tudor.ambarus@xxxxxxxxxxxxx>
>>>>>> ---
>>>> [...]
>>>>>>
>>>>>> static const struct spi_nor_controller_ops cqspi_controller_ops = {
>>>>>> @@ -1269,8 +1274,11 @@ static int cqspi_setup_flash(struct cqspi_st *cqspi, struct device_node *np)
>>>>>> dev_dbg(nor->dev, "using direct mode for %s\n",
>>>>>> mtd->name);
>>>>>>
>>>>>> - if (!cqspi->rx_chan)
>>>>>> - cqspi_request_mmap_dma(cqspi);
>>>>>> + if (!cqspi->rx_chan) {
>>>>>> + ret = cqspi_request_mmap_dma(cqspi);
>>>>>> + if (ret == -EPROBE_DEFER)
>>>>>> + goto err;
>>>>>> + }
>>>>>> }
>>>>>> }
>>>>>>
>>>>>>
>>>>>
>>>>> This seem to break reading the SPI flash on our IOT2050 [1] (didn't test
>>>>> the eval board yet).
>>>>>
>>>>> Without that commit, read happens via PIO, and that works. With the
>>>>> commit, the pattern
>>>>>
>>>>> with open("out.bin", "wb") as out:
>>>>> pos = 0
>>>>> while pos < 2:
>>>>> with open("/dev/mtd0", "rb") as mtd:
>>>>> mtd.seek(pos * 0x10000)
>>>>> out.write(mtd.read(0x10000))
>>>>> pos += 1
>>>>>
>>>>> gives the wrong result for the second block while
>>>>
>>>> Interesting... Could you please explain wrong result? Is the data move
>>>> around or completely garbage?
>>>
>>> It looks like some stripes contain data from other parts of the flash or
>>> kernel RAM. It's not just garbage, there are readable strings included.
>>>
>>>>
>>>> Does this fail even on AM654 EVM? Could you share full script for me to
>>>> test locally?
>>>
>>> The scripts are complete (python). Just binary-diff the outputs.
>>>
>>> I'll try on the EVM later.
>>
>> Done so now, could reproduce.
>
> ..."could *not* reproduce" there. Sorry if that caused confusion.
>

Oh, thanks! I was wondering why I cannot see this issue on AM654 EVM at my end...

>>
>> But the OSPIs are definitely different. EVM reports
>>
>> spi-nor spi0.0: mt35xu512aba (65536 Kbytes)
>>
>> with 4K erase size. Our our board, we have
>>
>> spi-nor spi7.0: w25q128 (16384 Kbytes)
>>
>> with 64K erase size.
>>
>> Here is some extract of the hex-diffs between out.bin and out2.bin (the
>> latter being the good one):
>>
>> --- /dev/fd/63 2020-08-24 17:16:58.776409282 +0000
>> +++ /dev/fd/62 2020-08-24 17:16:58.776409282 +0000
>> @@ -6,18 +6,18 @@
>> 00000050 0f 30 0d 06 03 55 04 07 0c 06 44 61 6c 6c 61 73 |.0...U....Dallas|
>> 00000060 31 27 30 25 06 03 55 04 0a 0c 1e 54 65 78 61 73 |1'0%..U....Texas|
>> 00000070 20 49 6e 73 74 72 75 6d 65 6e 74 73 20 49 6e 63 | Instruments Inc|
>> -00000080 84 8b 96 2c 0c 12 18 03 01 05 05 04 01 02 00 00 |...,............|
>> -00000090 07 06 44 45 20 01 0d 14 2a 01 00 32 05 24 30 48 |..DE ...*..2.$0H|
>> -000000a0 60 6c 30 14 01 00 00 0f ac 04 01 00 00 0f ac 04 |`l0.............|
>> -000000b0 01 00 00 0f ac 02 0c 00 2d 1a 6f 18 17 ff ff ff |........-.o.....|
>> +00000080 6f 72 70 6f 72 61 74 65 64 31 13 30 11 06 03 55 |orporated1.0...U|
>> +00000090 04 0b 0c 0a 50 72 6f 63 65 73 73 6f 72 73 31 13 |....Processors1.|
>> +000000a0 30 11 06 03 55 04 03 0c 0a 54 49 20 73 75 70 70 |0...U....TI supp|
>> +000000b0 6f 72 74 31 1d 30 1b 06 09 2a 86 48 86 f7 0d 01 |ort1.0...*.H....|
>> 000000c0 09 01 16 0e 73 75 70 70 6f 72 74 40 74 69 2e 63 |....support@ti.c|
>> 000000d0 6f 6d 30 1e 17 0d 32 30 30 37 32 32 31 31 30 30 |om0...2007221100|
>> 000000e0 30 30 5a 17 0d 32 30 30 38 32 31 31 31 30 30 30 |00Z..20082111000|
>> 000000f0 30 5a 30 81 9d 31 0b 30 09 06 03 55 04 06 13 02 |0Z0..1.0...U....|
>> -00000100 00 00 27 a4 00 00 42 43 5e 00 62 32 2f 00 b4 96 |..'...BC^.b2/...|
>> -00000110 24 44 0c 00 c6 00 43 0a 00 00 0b f0 43 a5 2a 01 |$D....C.....C.*.|
>> -00000120 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
>> -*
>> +00000100 55 53 31 0b 30 09 06 03 55 04 08 0c 02 54 58 31 |US1.0...U....TX1|
>> +00000110 0f 30 0d 06 03 55 04 07 0c 06 44 61 6c 6c 61 73 |.0...U....Dallas|
>> +00000120 31 27 30 25 06 03 55 04 0a 0c 1e 54 65 78 61 73 |1'0%..U....Texas|
>> +00000130 20 49 6e 73 74 72 75 6d 65 6e 74 73 20 49 6e 63 | Instruments Inc|
>> 00000140 6f 72 70 6f 72 61 74 65 64 31 13 30 11 06 03 55 |orporated1.0...U|
>> 00000150 04 0b 0c 0a 50 72 6f 63 65 73 73 6f 72 73 31 13 |....Processors1.|
>> 00000160 30 11 06 03 55 04 03 0c 0a 54 49 20 73 75 70 70 |0...U....TI supp|
>>
>> [...]
>>
>> 000017a0 02 8a e5 06 c8 8c e2 14 c2 8a e5 7c 01 00 ea ed |...........|....|
>> 000017b0 1f 8f e2 66 02 00 ea 5b 45 72 72 6f 72 5d 20 52 |...f...[Error] R|
>> -000017c0 69 64 20 55 54 43 20 49 44 21 21 21 0a 00 00 5b |id UTC ID!!!...[|
>> -000017d0 45 72 72 6f 72 5d 20 49 6e 76 61 6c 69 64 20 50 |Error] Invalid P|
>> -000017e0 65 65 72 20 43 68 61 6e 6e 65 6c 20 4e 75 6d 62 |eer Channel Numb|
>> -000017f0 65 72 21 21 21 0a 00 41 73 73 65 72 74 69 6f 6e |er!!!..Assertion|
>> +000017c0 4d 20 41 6c 6c 6f 63 20 54 58 20 43 68 20 66 61 |M Alloc TX Ch fa|
>> +000017d0 69 6c 65 64 21 21 21 0a 00 00 00 73 72 63 2f 75 |iled!!!....src/u|
>> +000017e0 64 6d 61 5f 63 68 2e 63 00 00 00 75 74 63 49 6e |dma_ch.c...utcIn|
>> +000017f0 66 6f 20 21 3d 20 4e 55 4c 4c 5f 50 54 52 00 75 |fo != NULL_PTR.u|
>> 00001800 74 63 49 64 20 3c 3d 20 55 44 4d 41 5f 4e 55 4d |tcId <= UDMA_NUM|
>> 00001810 5f 55 54 43 5f 49 4e 53 54 41 4e 43 45 00 00 72 |_UTC_INSTANCE..r|
>>
>
> I've done [1] for now in order to make the OSPI usable again here. It
> looks like reading an mtd device in one chunk (single read syscall) is
> fine, ie. "dd if=/dev/mtd3 of=content2 bs=<sizeof-mtd3>", while reading
> it in multiple chunks is problematic, e.g. "dd if=/dev/mtd3 of=content2
> bs=4096". Interestingly, the deviation is already on the first block,
> which may speak against a setup issue for a second transfer.

I cannot think of a reasonable explanation for this. I wonder how the first chunk
gets affected when reading in multiple chunks.

>
> The content I've seen in the corrupted output may come from other parts
> of the memory. I've found my WIFI SSID there, which is definitely not
> part of our OSPI image.
>

Hmm, one guess it that QSPI on IoT board is 16MB and hence
does not support 4 byte addressing vs the OSPI on AM654 EVM.
There could be bug around that case.

So, could you apply diff [1] on linux-next and then execute
falling testcase and post the register dump printed?

[1]:


---><8---


diff --git a/drivers/spi/spi-cadence-quadspi.c b/drivers/spi/spi-cadence-quadspi.c
index 508b219eabf80..b9739ae919340 100644
--- a/drivers/spi/spi-cadence-quadspi.c
+++ b/drivers/spi/spi-cadence-quadspi.c
@@ -907,6 +907,10 @@ static int cqspi_direct_read_execute(struct cqspi_flash_pdata *f_pdata,
struct dma_async_tx_descriptor *tx;
dma_cookie_t cookie;
dma_addr_t dma_dst;
+ int i;
+
+ for (i = 0; i < 10; i++)
+ dev_err(dev, "REG off %x: val %x\n", i, readl(cqspi->iobase + (i << 2)));

if (!cqspi->rx_chan || !virt_addr_valid(buf)) {
memcpy_fromio(buf, cqspi->ahb_base + from, len);


Also, there seems to be DMA mapping related issue, that was always present in
older driver as well. Could you see if diff [2] fixes the issue?

[2] Use DMA device for mapping:

---><8---


diff --git a/drivers/spi/spi-cadence-quadspi.c b/drivers/spi/spi-cadence-quadspi.c
index b9739ae919340..a546aa4598758 100644
--- a/drivers/spi/spi-cadence-quadspi.c
+++ b/drivers/spi/spi-cadence-quadspi.c
@@ -901,6 +901,7 @@ static int cqspi_direct_read_execute(struct cqspi_flash_pdata *f_pdata,
{
struct cqspi_st *cqspi = f_pdata->cqspi;
struct device *dev = &cqspi->pdev->dev;
+ struct device *ddev = cqspi->rx_chan->device->dev;
enum dma_ctrl_flags flags = DMA_CTRL_ACK | DMA_PREP_INTERRUPT;
dma_addr_t dma_src = (dma_addr_t)cqspi->mmap_phys_base + from;
int ret = 0;
@@ -917,8 +918,8 @@ static int cqspi_direct_read_execute(struct cqspi_flash_pdata *f_pdata,
return 0;
}

- dma_dst = dma_map_single(dev, buf, len, DMA_FROM_DEVICE);
- if (dma_mapping_error(dev, dma_dst)) {
+ dma_dst = dma_map_single(ddev, buf, len, DMA_FROM_DEVICE);
+ if (dma_mapping_error(ddev, dma_dst)) {
dev_err(dev, "dma mapping failed\n");
return -ENOMEM;
}
@@ -952,7 +953,7 @@ static int cqspi_direct_read_execute(struct cqspi_flash_pdata *f_pdata,
}

err_unmap:
- dma_unmap_single(dev, dma_dst, len, DMA_FROM_DEVICE);
+ dma_unmap_single(ddev, dma_dst, len, DMA_FROM_DEVICE);

return ret;
}