When the 'struct applespi_data' structure is allocated at the beginning of
applespi_probe(), 2504 bytes are allocated.
Because of the way memory is allocated, it ends to a 4096 bytes allocation.
So, about 1500 bytes are wasted.
Later in this function, when 'tx_buffer', 'tx_status', 'rx_buffer' and
'msg_buf' are allocated, 256, 4, 256 and 512 bytes are requested (~1 ko).
A devm_ memory allocation has a small overhead of 40 bytes. So, for the
same reason as above, it ends to allocate 512, 64, 512 and 1024 (~2 ko).
All that said, defining these 4 arrays as part of 'struct applespi_data'
saves 2 ko of runtime memory.
3504 bytes are now requested, and 4096 really allocated. All these 4
arrays fit in the 'wasted' memory of the first allocation.
Signed-off-by: Christophe JAILLET <[email protected]>
---
Compile tested only.
---
drivers/input/keyboard/applespi.c | 23 ++++-------------------
1 file changed, 4 insertions(+), 19 deletions(-)
diff --git a/drivers/input/keyboard/applespi.c b/drivers/input/keyboard/applespi.c
index fab5473ae5da..bee4ccfa2b05 100644
--- a/drivers/input/keyboard/applespi.c
+++ b/drivers/input/keyboard/applespi.c
@@ -373,11 +373,11 @@ struct applespi_data {
struct input_dev *keyboard_input_dev;
struct input_dev *touchpad_input_dev;
- u8 *tx_buffer;
- u8 *tx_status;
- u8 *rx_buffer;
+ u8 tx_buffer[APPLESPI_PACKET_SIZE];
+ u8 tx_status[APPLESPI_STATUS_SIZE];
+ u8 rx_buffer[APPLESPI_PACKET_SIZE];
- u8 *msg_buf;
+ u8 msg_buf[MAX_PKTS_PER_MSG * APPLESPI_PACKET_SIZE];
unsigned int saved_msg_len;
struct applespi_tp_info tp_info;
@@ -1659,21 +1659,6 @@ static int applespi_probe(struct spi_device *spi)
/* store the driver data */
spi_set_drvdata(spi, applespi);
- /* create our buffers */
- applespi->tx_buffer = devm_kmalloc(&spi->dev, APPLESPI_PACKET_SIZE,
- GFP_KERNEL);
- applespi->tx_status = devm_kmalloc(&spi->dev, APPLESPI_STATUS_SIZE,
- GFP_KERNEL);
- applespi->rx_buffer = devm_kmalloc(&spi->dev, APPLESPI_PACKET_SIZE,
- GFP_KERNEL);
- applespi->msg_buf = devm_kmalloc_array(&spi->dev, MAX_PKTS_PER_MSG,
- APPLESPI_PACKET_SIZE,
- GFP_KERNEL);
-
- if (!applespi->tx_buffer || !applespi->tx_status ||
- !applespi->rx_buffer || !applespi->msg_buf)
- return -ENOMEM;
-
/* set up our spi messages */
applespi_setup_read_txfrs(applespi);
applespi_setup_write_txfrs(applespi);
--
2.34.1
On Sun, Sep 18, 2022 at 03:08:17PM +0200, Christophe JAILLET wrote:
> When the 'struct applespi_data' structure is allocated at the beginning of
> applespi_probe(), 2504 bytes are allocated.
>
> Because of the way memory is allocated, it ends to a 4096 bytes allocation.
> So, about 1500 bytes are wasted.
>
> Later in this function, when 'tx_buffer', 'tx_status', 'rx_buffer' and
> 'msg_buf' are allocated, 256, 4, 256 and 512 bytes are requested (~1 ko).
> A devm_ memory allocation has a small overhead of 40 bytes. So, for the
> same reason as above, it ends to allocate 512, 64, 512 and 1024 (~2 ko).
>
> All that said, defining these 4 arrays as part of 'struct applespi_data'
> saves 2 ko of runtime memory.
>
> 3504 bytes are now requested, and 4096 really allocated. All these 4
> arrays fit in the 'wasted' memory of the first allocation.
>
> Signed-off-by: Christophe JAILLET <[email protected]>
> ---
> Compile tested only.
> ---
> drivers/input/keyboard/applespi.c | 23 ++++-------------------
> 1 file changed, 4 insertions(+), 19 deletions(-)
>
> diff --git a/drivers/input/keyboard/applespi.c b/drivers/input/keyboard/applespi.c
> index fab5473ae5da..bee4ccfa2b05 100644
> --- a/drivers/input/keyboard/applespi.c
> +++ b/drivers/input/keyboard/applespi.c
> @@ -373,11 +373,11 @@ struct applespi_data {
> struct input_dev *keyboard_input_dev;
> struct input_dev *touchpad_input_dev;
>
> - u8 *tx_buffer;
> - u8 *tx_status;
> - u8 *rx_buffer;
> + u8 tx_buffer[APPLESPI_PACKET_SIZE];
> + u8 tx_status[APPLESPI_STATUS_SIZE];
> + u8 rx_buffer[APPLESPI_PACKET_SIZE];
>
> - u8 *msg_buf;
> + u8 msg_buf[MAX_PKTS_PER_MSG * APPLESPI_PACKET_SIZE];
> unsigned int saved_msg_len;
>
> struct applespi_tp_info tp_info;
This kind of change is generally broken in case DMA can be involved.
Allocating the transfer buffers separately makes sure that alignment
requirements are met and avoids hard-to-debug memory corruption issues.
Johan
Le 19/09/2022 à 09:02, Johan Hovold a écrit :
> On Sun, Sep 18, 2022 at 03:08:17PM +0200, Christophe JAILLET wrote:
>> When the 'struct applespi_data' structure is allocated at the beginning of
>> applespi_probe(), 2504 bytes are allocated.
>>
>> Because of the way memory is allocated, it ends to a 4096 bytes allocation.
>> So, about 1500 bytes are wasted.
>>
>> Later in this function, when 'tx_buffer', 'tx_status', 'rx_buffer' and
>> 'msg_buf' are allocated, 256, 4, 256 and 512 bytes are requested (~1 ko).
>> A devm_ memory allocation has a small overhead of 40 bytes. So, for the
>> same reason as above, it ends to allocate 512, 64, 512 and 1024 (~2 ko).
>>
>> All that said, defining these 4 arrays as part of 'struct applespi_data'
>> saves 2 ko of runtime memory.
>>
>> 3504 bytes are now requested, and 4096 really allocated. All these 4
>> arrays fit in the 'wasted' memory of the first allocation.
>>
>> Signed-off-by: Christophe JAILLET <[email protected]>
>> ---
>> Compile tested only.
>> ---
>> drivers/input/keyboard/applespi.c | 23 ++++-------------------
>> 1 file changed, 4 insertions(+), 19 deletions(-)
>>
>> diff --git a/drivers/input/keyboard/applespi.c b/drivers/input/keyboard/applespi.c
>> index fab5473ae5da..bee4ccfa2b05 100644
>> --- a/drivers/input/keyboard/applespi.c
>> +++ b/drivers/input/keyboard/applespi.c
>> @@ -373,11 +373,11 @@ struct applespi_data {
>> struct input_dev *keyboard_input_dev;
>> struct input_dev *touchpad_input_dev;
>>
>> - u8 *tx_buffer;
>> - u8 *tx_status;
>> - u8 *rx_buffer;
>> + u8 tx_buffer[APPLESPI_PACKET_SIZE];
>> + u8 tx_status[APPLESPI_STATUS_SIZE];
>> + u8 rx_buffer[APPLESPI_PACKET_SIZE];
>>
>> - u8 *msg_buf;
>> + u8 msg_buf[MAX_PKTS_PER_MSG * APPLESPI_PACKET_SIZE];
>> unsigned int saved_msg_len;
>>
>> struct applespi_tp_info tp_info;
>
> This kind of change is generally broken in case DMA can be involved.
>
> Allocating the transfer buffers separately makes sure that alignment
> requirements are met and avoids hard-to-debug memory corruption issues.
>
> Johan
>
Got it. I'll keep away from it.
Thanks for the feed-back and explanation.
CJ