| Message ID | 20180917133652.12442-1-boris.brezillon@bootlin.com |
|---|---|
| State | Superseded |
| Delegated to: | Boris Brezillon |
| Headers | show |
| Series | mtd: devices: m25p80: Make the buffer passed in op is DMA-able | expand |
Oops. You should read "Make *sure* ...". On Mon, 17 Sep 2018 15:36:52 +0200 Boris Brezillon <boris.brezillon@bootlin.com> wrote: > As documented in spi-mem.h, spi_mem_op->data.buf.{in,out} must be > DMA-able, and commit 4120f8d158ef ("mtd: spi-nor: Use the spi_mem_xx() > API") failed to follow this rule as buffers passed to > ->{read,write}_reg() are usually placed on the stack. > > Fix that by allocating a scratch buffer and copying the data in there > before passing it to the spi-mem layer. > > Fixes: 4120f8d158ef ("mtd: spi-nor: Use the spi_mem_xx() API") > Reported-by: Jarkko Nikula <jarkko.nikula@linux.intel.com> > Cc: <stable@vger.kernel.org> > Signed-off-by: Boris Brezillon <boris.brezillon@bootlin.com> > --- > Note that the ->{read,write}() path is still buggy since nothing > guarantees that buffers passed by the MTD layer to the SPI NOR layer > are DMA-able, but this is a long-standing issue which we'll have to > address at the spi-nor level (this layer can choose the bounce buffer > size based on nor->page_size). > --- > drivers/mtd/devices/m25p80.c | 24 +++++++++++++++++++++--- > 1 file changed, 21 insertions(+), 3 deletions(-) > > diff --git a/drivers/mtd/devices/m25p80.c b/drivers/mtd/devices/m25p80.c > index cbfafc453274..3b7fafa4bbd6 100644 > --- a/drivers/mtd/devices/m25p80.c > +++ b/drivers/mtd/devices/m25p80.c > @@ -39,14 +39,22 @@ static int m25p80_read_reg(struct spi_nor *nor, u8 code, u8 *val, int len) > struct spi_mem_op op = SPI_MEM_OP(SPI_MEM_OP_CMD(code, 1), > SPI_MEM_OP_NO_ADDR, > SPI_MEM_OP_NO_DUMMY, > - SPI_MEM_OP_DATA_IN(len, val, 1)); > + SPI_MEM_OP_DATA_IN(len, NULL, 1)); > + void *scratchbuf; > int ret; > > + scratchbuf = kmemdup(val, len, GFP_KERNEL); > + if (!scratchbuf) > + return -ENOMEM; > + > + op.data.buf.in = scratchbuf; > ret = spi_mem_exec_op(flash->spimem, &op); > if (ret < 0) > dev_err(&flash->spimem->spi->dev, "error %d reading %x\n", ret, > code); > > + kfree(scratchbuf); > + > return ret; > } > > @@ -56,9 +64,19 @@ static int m25p80_write_reg(struct spi_nor *nor, u8 opcode, u8 *buf, int len) > struct spi_mem_op op = SPI_MEM_OP(SPI_MEM_OP_CMD(opcode, 1), > SPI_MEM_OP_NO_ADDR, > SPI_MEM_OP_NO_DUMMY, > - SPI_MEM_OP_DATA_OUT(len, buf, 1)); > + SPI_MEM_OP_DATA_OUT(len, NULL, 1)); > + void *scratchbuf; > + int ret; > > - return spi_mem_exec_op(flash->spimem, &op); > + scratchbuf = kmemdup(buf, len, GFP_KERNEL); > + if (!scratchbuf) > + return -ENOMEM; > + > + op.data.buf.out = scratchbuf; > + ret = spi_mem_exec_op(flash->spimem, &op); > + kfree(scratchbuf); > + > + return ret; > } > > static ssize_t m25p80_write(struct spi_nor *nor, loff_t to, size_t len,
On 09/17/2018 04:36 PM, Boris Brezillon wrote: > As documented in spi-mem.h, spi_mem_op->data.buf.{in,out} must be > DMA-able, and commit 4120f8d158ef ("mtd: spi-nor: Use the spi_mem_xx() > API") failed to follow this rule as buffers passed to > ->{read,write}_reg() are usually placed on the stack. > > Fix that by allocating a scratch buffer and copying the data in there > before passing it to the spi-mem layer. > > Fixes: 4120f8d158ef ("mtd: spi-nor: Use the spi_mem_xx() API") > Reported-by: Jarkko Nikula <jarkko.nikula@linux.intel.com> > Cc: <stable@vger.kernel.org> > Signed-off-by: Boris Brezillon <boris.brezillon@bootlin.com> > --- > Note that the ->{read,write}() path is still buggy since nothing > guarantees that buffers passed by the MTD layer to the SPI NOR layer > are DMA-able, but this is a long-standing issue which we'll have to > address at the spi-nor level (this layer can choose the bounce buffer > size based on nor->page_size). > --- > drivers/mtd/devices/m25p80.c | 24 +++++++++++++++++++++--- > 1 file changed, 21 insertions(+), 3 deletions(-) > > diff --git a/drivers/mtd/devices/m25p80.c b/drivers/mtd/devices/m25p80.c > index cbfafc453274..3b7fafa4bbd6 100644 > --- a/drivers/mtd/devices/m25p80.c > +++ b/drivers/mtd/devices/m25p80.c > @@ -39,14 +39,22 @@ static int m25p80_read_reg(struct spi_nor *nor, u8 code, u8 *val, int len) > struct spi_mem_op op = SPI_MEM_OP(SPI_MEM_OP_CMD(code, 1), > SPI_MEM_OP_NO_ADDR, > SPI_MEM_OP_NO_DUMMY, > - SPI_MEM_OP_DATA_IN(len, val, 1)); > + SPI_MEM_OP_DATA_IN(len, NULL, 1)); > + void *scratchbuf; > int ret; > > + scratchbuf = kmemdup(val, len, GFP_KERNEL); > + if (!scratchbuf) > + return -ENOMEM; > + > + op.data.buf.in = scratchbuf; > ret = spi_mem_exec_op(flash->spimem, &op); > if (ret < 0) > dev_err(&flash->spimem->spi->dev, "error %d reading %x\n", ret, > code); > > + kfree(scratchbuf); > + Actually this makes things even worse: "unrecognized JEDEC id bytes: b1, 74, 01" since we don't copy scratchbuf to original buffer and kmemdup before bus read is needless :-)
On Mon, 17 Sep 2018 17:13:32 +0300 Jarkko Nikula <jarkko.nikula@linux.intel.com> wrote: > On 09/17/2018 04:36 PM, Boris Brezillon wrote: > > As documented in spi-mem.h, spi_mem_op->data.buf.{in,out} must be > > DMA-able, and commit 4120f8d158ef ("mtd: spi-nor: Use the spi_mem_xx() > > API") failed to follow this rule as buffers passed to > > ->{read,write}_reg() are usually placed on the stack. > > > > Fix that by allocating a scratch buffer and copying the data in there > > before passing it to the spi-mem layer. > > > > Fixes: 4120f8d158ef ("mtd: spi-nor: Use the spi_mem_xx() API") > > Reported-by: Jarkko Nikula <jarkko.nikula@linux.intel.com> > > Cc: <stable@vger.kernel.org> > > Signed-off-by: Boris Brezillon <boris.brezillon@bootlin.com> > > --- > > Note that the ->{read,write}() path is still buggy since nothing > > guarantees that buffers passed by the MTD layer to the SPI NOR layer > > are DMA-able, but this is a long-standing issue which we'll have to > > address at the spi-nor level (this layer can choose the bounce buffer > > size based on nor->page_size). > > --- > > drivers/mtd/devices/m25p80.c | 24 +++++++++++++++++++++--- > > 1 file changed, 21 insertions(+), 3 deletions(-) > > > > diff --git a/drivers/mtd/devices/m25p80.c b/drivers/mtd/devices/m25p80.c > > index cbfafc453274..3b7fafa4bbd6 100644 > > --- a/drivers/mtd/devices/m25p80.c > > +++ b/drivers/mtd/devices/m25p80.c > > @@ -39,14 +39,22 @@ static int m25p80_read_reg(struct spi_nor *nor, u8 code, u8 *val, int len) > > struct spi_mem_op op = SPI_MEM_OP(SPI_MEM_OP_CMD(code, 1), > > SPI_MEM_OP_NO_ADDR, > > SPI_MEM_OP_NO_DUMMY, > > - SPI_MEM_OP_DATA_IN(len, val, 1)); > > + SPI_MEM_OP_DATA_IN(len, NULL, 1)); > > + void *scratchbuf; > > int ret; > > > > + scratchbuf = kmemdup(val, len, GFP_KERNEL); > > + if (!scratchbuf) > > + return -ENOMEM; > > + > > + op.data.buf.in = scratchbuf; > > ret = spi_mem_exec_op(flash->spimem, &op); > > if (ret < 0) > > dev_err(&flash->spimem->spi->dev, "error %d reading %x\n", ret, > > code); > > > > + kfree(scratchbuf); > > + > > Actually this makes things even worse: "unrecognized JEDEC id bytes: b1, > 74, 01" since we don't copy scratchbuf to original buffer and kmemdup > before bus read is needless :-) True :-). I'll fix that and send a v2.
diff --git a/drivers/mtd/devices/m25p80.c b/drivers/mtd/devices/m25p80.c index cbfafc453274..3b7fafa4bbd6 100644 --- a/drivers/mtd/devices/m25p80.c +++ b/drivers/mtd/devices/m25p80.c @@ -39,14 +39,22 @@ static int m25p80_read_reg(struct spi_nor *nor, u8 code, u8 *val, int len) struct spi_mem_op op = SPI_MEM_OP(SPI_MEM_OP_CMD(code, 1), SPI_MEM_OP_NO_ADDR, SPI_MEM_OP_NO_DUMMY, - SPI_MEM_OP_DATA_IN(len, val, 1)); + SPI_MEM_OP_DATA_IN(len, NULL, 1)); + void *scratchbuf; int ret; + scratchbuf = kmemdup(val, len, GFP_KERNEL); + if (!scratchbuf) + return -ENOMEM; + + op.data.buf.in = scratchbuf; ret = spi_mem_exec_op(flash->spimem, &op); if (ret < 0) dev_err(&flash->spimem->spi->dev, "error %d reading %x\n", ret, code); + kfree(scratchbuf); + return ret; } @@ -56,9 +64,19 @@ static int m25p80_write_reg(struct spi_nor *nor, u8 opcode, u8 *buf, int len) struct spi_mem_op op = SPI_MEM_OP(SPI_MEM_OP_CMD(opcode, 1), SPI_MEM_OP_NO_ADDR, SPI_MEM_OP_NO_DUMMY, - SPI_MEM_OP_DATA_OUT(len, buf, 1)); + SPI_MEM_OP_DATA_OUT(len, NULL, 1)); + void *scratchbuf; + int ret; - return spi_mem_exec_op(flash->spimem, &op); + scratchbuf = kmemdup(buf, len, GFP_KERNEL); + if (!scratchbuf) + return -ENOMEM; + + op.data.buf.out = scratchbuf; + ret = spi_mem_exec_op(flash->spimem, &op); + kfree(scratchbuf); + + return ret; } static ssize_t m25p80_write(struct spi_nor *nor, loff_t to, size_t len,
As documented in spi-mem.h, spi_mem_op->data.buf.{in,out} must be DMA-able, and commit 4120f8d158ef ("mtd: spi-nor: Use the spi_mem_xx() API") failed to follow this rule as buffers passed to ->{read,write}_reg() are usually placed on the stack. Fix that by allocating a scratch buffer and copying the data in there before passing it to the spi-mem layer. Fixes: 4120f8d158ef ("mtd: spi-nor: Use the spi_mem_xx() API") Reported-by: Jarkko Nikula <jarkko.nikula@linux.intel.com> Cc: <stable@vger.kernel.org> Signed-off-by: Boris Brezillon <boris.brezillon@bootlin.com> --- Note that the ->{read,write}() path is still buggy since nothing guarantees that buffers passed by the MTD layer to the SPI NOR layer are DMA-able, but this is a long-standing issue which we'll have to address at the spi-nor level (this layer can choose the bounce buffer size based on nor->page_size). --- drivers/mtd/devices/m25p80.c | 24 +++++++++++++++++++++--- 1 file changed, 21 insertions(+), 3 deletions(-)