linux/drivers/misc/eeprom/at25.c

549 lines
14 KiB
C
Raw Permalink Normal View History

// SPDX-License-Identifier: GPL-2.0-or-later
/*
* Driver for most of the SPI EEPROMs, such as Atmel AT25 models
* and Cypress FRAMs FM25 models.
*
* Copyright (C) 2006 David Brownell
*/
#include <linux/bits.h>
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
#include <linux/cleanup.h>
#include <linux/delay.h>
#include <linux/device.h>
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
#include <linux/iopoll.h>
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/property.h>
#include <linux/sched.h>
#include <linux/slab.h>
#include <linux/spi/eeprom.h>
#include <linux/spi/spi.h>
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
#include <linux/spi/spi-mem.h>
#include <linux/nvmem-provider.h>
/*
* NOTE: this is an *EEPROM* driver. The vagaries of product naming
* mean that some AT25 products are EEPROMs, and others are FLASH.
* Handle FLASH chips with the drivers/mtd/devices/m25p80.c driver,
* not this one!
*
* EEPROMs that can be used with this driver include, for example:
* AT25M02, AT25128B
*/
#define FM25_SN_LEN 8 /* serial number length */
eeprom: at25: Use DMA safe buffers Reading EEPROM fails with following warning: [ 16.357496] ------------[ cut here ]------------ [ 16.357529] fsl_spi b01004c0.spi: rejecting DMA map of vmalloc memory [ 16.357698] WARNING: CPU: 0 PID: 371 at include/linux/dma-mapping.h:326 fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.357775] CPU: 0 PID: 371 Comm: od Not tainted 5.16.11-s3k-dev-01743-g19beecbfe9d6-dirty #109 [ 16.357806] NIP: c03fbc9c LR: c03fbc9c CTR: 00000000 [ 16.357825] REGS: e68d9b20 TRAP: 0700 Not tainted (5.16.11-s3k-dev-01743-g19beecbfe9d6-dirty) [ 16.357849] MSR: 00029032 <EE,ME,IR,DR,RI> CR: 24002282 XER: 00000000 [ 16.357931] [ 16.357931] GPR00: c03fbc9c e68d9be0 c26d06a0 00000039 00000001 c0d36364 c0e96428 00000027 [ 16.357931] GPR08: 00000001 00000000 00000023 3fffc000 24002282 100d3dd6 100a2ffc 00000000 [ 16.357931] GPR16: 100cd280 100b0000 00000000 aff54f7e 100d0000 100d0000 00000001 100cf328 [ 16.357931] GPR24: 100cf328 00000000 00000003 e68d9e30 c156b410 e67ab4c0 e68d9d38 c24ab278 [ 16.358253] NIP [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.358292] LR [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.358325] Call Trace: [ 16.358336] [e68d9be0] [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 (unreliable) [ 16.358388] [e68d9c00] [c03fcb44] fsl_spi_bufs.isra.0+0x94/0x1a0 [ 16.358436] [e68d9c20] [c03fd970] fsl_spi_do_one_msg+0x254/0x3dc [ 16.358483] [e68d9cb0] [c03f7e50] __spi_pump_messages+0x274/0x8a4 [ 16.358529] [e68d9ce0] [c03f9d30] __spi_sync+0x344/0x378 [ 16.358573] [e68d9d20] [c03fb52c] spi_sync+0x34/0x60 [ 16.358616] [e68d9d30] [c03b4dec] at25_ee_read+0x138/0x1a8 [ 16.358667] [e68d9e50] [c04a8fb8] bin_attr_nvmem_read+0x98/0x110 [ 16.358725] [e68d9e60] [c0204b14] kernfs_fop_read_iter+0xc0/0x1fc [ 16.358774] [e68d9e80] [c0168660] vfs_read+0x284/0x410 [ 16.358821] [e68d9f00] [c016925c] ksys_read+0x6c/0x11c [ 16.358863] [e68d9f30] [c00160e0] ret_from_syscall+0x0/0x28 ... [ 16.359608] ---[ end trace a4ce3e34afef0cb5 ]--- [ 16.359638] fsl_spi b01004c0.spi: unable to map tx dma This is due to the AT25 driver using buffers on stack, which is not possible with CONFIG_VMAP_STACK. As mentionned in kernel Documentation (Documentation/spi/spi-summary.rst): - Follow standard kernel rules, and provide DMA-safe buffers in your messages. That way controller drivers using DMA aren't forced to make extra copies unless the hardware requires it (e.g. working around hardware errata that force the use of bounce buffering). Modify the driver to use a buffer located in the at25 device structure which is allocated via kmalloc during probe. Protect writes in this new buffer with the driver's mutex. Fixes: b587b13a4f67 ("[PATCH] SPI eeprom driver") Cc: stable <stable@kernel.org> Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Link: https://lore.kernel.org/r/230a9486fc68ea0182df46255e42a51099403642.1648032613.git.christophe.leroy@csgroup.eu Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2022-03-23 11:51:55 +01:00
#define EE_MAXADDRLEN 3 /* 24 bit addresses, up to 2 MBytes */
struct at25_data {
struct spi_eeprom chip;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
struct spi_mem *spimem;
struct mutex lock;
unsigned addrlen;
struct nvmem_config nvmem_config;
struct nvmem_device *nvmem;
u8 sernum[FM25_SN_LEN];
};
#define AT25_WREN 0x06 /* latch the write enable */
#define AT25_WRDI 0x04 /* reset the write enable */
#define AT25_RDSR 0x05 /* read status register */
#define AT25_WRSR 0x01 /* write status register */
#define AT25_READ 0x03 /* read byte(s) */
#define AT25_WRITE 0x02 /* write byte(s)/sector */
#define FM25_SLEEP 0xb9 /* enter sleep mode */
#define FM25_RDID 0x9f /* read device ID */
#define FM25_RDSN 0xc3 /* read S/N */
#define AT25_SR_nRDY 0x01 /* nRDY = write-in-progress */
#define AT25_SR_WEN 0x02 /* write enable (latched) */
#define AT25_SR_BP0 0x04 /* BP for software writeprotect */
#define AT25_SR_BP1 0x08
#define AT25_SR_WPEN 0x80 /* writeprotect enable */
#define AT25_INSTR_BIT3 0x08 /* additional address bit in instr */
#define FM25_ID_LEN 9 /* ID length */
/*
* Specs often allow 5ms for a page write, sometimes 20ms;
* it's important to recover from write timeouts.
*/
#define EE_TIMEOUT 25
/*-------------------------------------------------------------------------*/
#define io_limit PAGE_SIZE /* bytes */
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
/* Handle the address MSB as part of instruction byte */
static u8 at25_instr(struct at25_data *at25, u8 instr, unsigned int off)
{
if (!(at25->chip.flags & EE_INSTR_BIT3_IS_ADDR))
return instr;
if (off < BIT(at25->addrlen * 8))
return instr;
return instr | AT25_INSTR_BIT3;
}
static int at25_ee_read(void *priv, unsigned int offset,
void *val, size_t count)
{
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
u8 *bounce __free(kfree) = kmalloc(min(count, io_limit), GFP_KERNEL);
struct at25_data *at25 = priv;
char *buf = val;
unsigned int msg_offset = offset;
size_t bytes_left = count;
size_t segment;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
int status;
if (!bounce)
return -ENOMEM;
if (unlikely(offset >= at25->chip.byte_len))
return -EINVAL;
if ((offset + count) > at25->chip.byte_len)
count = at25->chip.byte_len - offset;
if (unlikely(!count))
return -EINVAL;
do {
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
struct spi_mem_op op;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
segment = min(bytes_left, io_limit);
eeprom: at25: Use DMA safe buffers Reading EEPROM fails with following warning: [ 16.357496] ------------[ cut here ]------------ [ 16.357529] fsl_spi b01004c0.spi: rejecting DMA map of vmalloc memory [ 16.357698] WARNING: CPU: 0 PID: 371 at include/linux/dma-mapping.h:326 fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.357775] CPU: 0 PID: 371 Comm: od Not tainted 5.16.11-s3k-dev-01743-g19beecbfe9d6-dirty #109 [ 16.357806] NIP: c03fbc9c LR: c03fbc9c CTR: 00000000 [ 16.357825] REGS: e68d9b20 TRAP: 0700 Not tainted (5.16.11-s3k-dev-01743-g19beecbfe9d6-dirty) [ 16.357849] MSR: 00029032 <EE,ME,IR,DR,RI> CR: 24002282 XER: 00000000 [ 16.357931] [ 16.357931] GPR00: c03fbc9c e68d9be0 c26d06a0 00000039 00000001 c0d36364 c0e96428 00000027 [ 16.357931] GPR08: 00000001 00000000 00000023 3fffc000 24002282 100d3dd6 100a2ffc 00000000 [ 16.357931] GPR16: 100cd280 100b0000 00000000 aff54f7e 100d0000 100d0000 00000001 100cf328 [ 16.357931] GPR24: 100cf328 00000000 00000003 e68d9e30 c156b410 e67ab4c0 e68d9d38 c24ab278 [ 16.358253] NIP [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.358292] LR [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.358325] Call Trace: [ 16.358336] [e68d9be0] [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 (unreliable) [ 16.358388] [e68d9c00] [c03fcb44] fsl_spi_bufs.isra.0+0x94/0x1a0 [ 16.358436] [e68d9c20] [c03fd970] fsl_spi_do_one_msg+0x254/0x3dc [ 16.358483] [e68d9cb0] [c03f7e50] __spi_pump_messages+0x274/0x8a4 [ 16.358529] [e68d9ce0] [c03f9d30] __spi_sync+0x344/0x378 [ 16.358573] [e68d9d20] [c03fb52c] spi_sync+0x34/0x60 [ 16.358616] [e68d9d30] [c03b4dec] at25_ee_read+0x138/0x1a8 [ 16.358667] [e68d9e50] [c04a8fb8] bin_attr_nvmem_read+0x98/0x110 [ 16.358725] [e68d9e60] [c0204b14] kernfs_fop_read_iter+0xc0/0x1fc [ 16.358774] [e68d9e80] [c0168660] vfs_read+0x284/0x410 [ 16.358821] [e68d9f00] [c016925c] ksys_read+0x6c/0x11c [ 16.358863] [e68d9f30] [c00160e0] ret_from_syscall+0x0/0x28 ... [ 16.359608] ---[ end trace a4ce3e34afef0cb5 ]--- [ 16.359638] fsl_spi b01004c0.spi: unable to map tx dma This is due to the AT25 driver using buffers on stack, which is not possible with CONFIG_VMAP_STACK. As mentionned in kernel Documentation (Documentation/spi/spi-summary.rst): - Follow standard kernel rules, and provide DMA-safe buffers in your messages. That way controller drivers using DMA aren't forced to make extra copies unless the hardware requires it (e.g. working around hardware errata that force the use of bounce buffering). Modify the driver to use a buffer located in the at25 device structure which is allocated via kmalloc during probe. Protect writes in this new buffer with the driver's mutex. Fixes: b587b13a4f67 ("[PATCH] SPI eeprom driver") Cc: stable <stable@kernel.org> Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Link: https://lore.kernel.org/r/230a9486fc68ea0182df46255e42a51099403642.1648032613.git.christophe.leroy@csgroup.eu Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2022-03-23 11:51:55 +01:00
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
op = (struct spi_mem_op)SPI_MEM_OP(SPI_MEM_OP_CMD(at25_instr(at25, AT25_READ,
msg_offset), 1),
SPI_MEM_OP_ADDR(at25->addrlen, msg_offset, 1),
SPI_MEM_OP_NO_DUMMY,
SPI_MEM_OP_DATA_IN(segment, bounce, 1));
eeprom: at25: Use DMA safe buffers Reading EEPROM fails with following warning: [ 16.357496] ------------[ cut here ]------------ [ 16.357529] fsl_spi b01004c0.spi: rejecting DMA map of vmalloc memory [ 16.357698] WARNING: CPU: 0 PID: 371 at include/linux/dma-mapping.h:326 fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.357775] CPU: 0 PID: 371 Comm: od Not tainted 5.16.11-s3k-dev-01743-g19beecbfe9d6-dirty #109 [ 16.357806] NIP: c03fbc9c LR: c03fbc9c CTR: 00000000 [ 16.357825] REGS: e68d9b20 TRAP: 0700 Not tainted (5.16.11-s3k-dev-01743-g19beecbfe9d6-dirty) [ 16.357849] MSR: 00029032 <EE,ME,IR,DR,RI> CR: 24002282 XER: 00000000 [ 16.357931] [ 16.357931] GPR00: c03fbc9c e68d9be0 c26d06a0 00000039 00000001 c0d36364 c0e96428 00000027 [ 16.357931] GPR08: 00000001 00000000 00000023 3fffc000 24002282 100d3dd6 100a2ffc 00000000 [ 16.357931] GPR16: 100cd280 100b0000 00000000 aff54f7e 100d0000 100d0000 00000001 100cf328 [ 16.357931] GPR24: 100cf328 00000000 00000003 e68d9e30 c156b410 e67ab4c0 e68d9d38 c24ab278 [ 16.358253] NIP [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.358292] LR [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.358325] Call Trace: [ 16.358336] [e68d9be0] [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 (unreliable) [ 16.358388] [e68d9c00] [c03fcb44] fsl_spi_bufs.isra.0+0x94/0x1a0 [ 16.358436] [e68d9c20] [c03fd970] fsl_spi_do_one_msg+0x254/0x3dc [ 16.358483] [e68d9cb0] [c03f7e50] __spi_pump_messages+0x274/0x8a4 [ 16.358529] [e68d9ce0] [c03f9d30] __spi_sync+0x344/0x378 [ 16.358573] [e68d9d20] [c03fb52c] spi_sync+0x34/0x60 [ 16.358616] [e68d9d30] [c03b4dec] at25_ee_read+0x138/0x1a8 [ 16.358667] [e68d9e50] [c04a8fb8] bin_attr_nvmem_read+0x98/0x110 [ 16.358725] [e68d9e60] [c0204b14] kernfs_fop_read_iter+0xc0/0x1fc [ 16.358774] [e68d9e80] [c0168660] vfs_read+0x284/0x410 [ 16.358821] [e68d9f00] [c016925c] ksys_read+0x6c/0x11c [ 16.358863] [e68d9f30] [c00160e0] ret_from_syscall+0x0/0x28 ... [ 16.359608] ---[ end trace a4ce3e34afef0cb5 ]--- [ 16.359638] fsl_spi b01004c0.spi: unable to map tx dma This is due to the AT25 driver using buffers on stack, which is not possible with CONFIG_VMAP_STACK. As mentionned in kernel Documentation (Documentation/spi/spi-summary.rst): - Follow standard kernel rules, and provide DMA-safe buffers in your messages. That way controller drivers using DMA aren't forced to make extra copies unless the hardware requires it (e.g. working around hardware errata that force the use of bounce buffering). Modify the driver to use a buffer located in the at25 device structure which is allocated via kmalloc during probe. Protect writes in this new buffer with the driver's mutex. Fixes: b587b13a4f67 ("[PATCH] SPI eeprom driver") Cc: stable <stable@kernel.org> Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Link: https://lore.kernel.org/r/230a9486fc68ea0182df46255e42a51099403642.1648032613.git.christophe.leroy@csgroup.eu Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2022-03-23 11:51:55 +01:00
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
status = spi_mem_adjust_op_size(at25->spimem, &op);
if (status)
return status;
segment = op.data.nbytes;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
mutex_lock(&at25->lock);
status = spi_mem_exec_op(at25->spimem, &op);
mutex_unlock(&at25->lock);
if (status)
return status;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
memcpy(buf, bounce, segment);
msg_offset += segment;
buf += segment;
bytes_left -= segment;
} while (bytes_left > 0);
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
dev_dbg(&at25->spimem->spi->dev, "read %zu bytes at %d\n",
count, offset);
return 0;
}
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
/*
* Read extra registers as ID or serial number
*
* Allow for the callers to provide @buf on stack (not necessary DMA-capable)
* by allocating a bounce buffer internally.
*/
static int fm25_aux_read(struct at25_data *at25, u8 *buf, uint8_t command,
int len)
{
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
u8 *bounce __free(kfree) = kmalloc(len, GFP_KERNEL);
struct spi_mem_op op;
int status;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
if (!bounce)
return -ENOMEM;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
op = (struct spi_mem_op)SPI_MEM_OP(SPI_MEM_OP_CMD(command, 1),
SPI_MEM_OP_NO_ADDR,
SPI_MEM_OP_NO_DUMMY,
SPI_MEM_OP_DATA_IN(len, bounce, 1));
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
status = spi_mem_exec_op(at25->spimem, &op);
dev_dbg(&at25->spimem->spi->dev, "read %d aux bytes --> %d\n", len, status);
if (status)
return status;
eeprom: at25: Use DMA safe buffers Reading EEPROM fails with following warning: [ 16.357496] ------------[ cut here ]------------ [ 16.357529] fsl_spi b01004c0.spi: rejecting DMA map of vmalloc memory [ 16.357698] WARNING: CPU: 0 PID: 371 at include/linux/dma-mapping.h:326 fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.357775] CPU: 0 PID: 371 Comm: od Not tainted 5.16.11-s3k-dev-01743-g19beecbfe9d6-dirty #109 [ 16.357806] NIP: c03fbc9c LR: c03fbc9c CTR: 00000000 [ 16.357825] REGS: e68d9b20 TRAP: 0700 Not tainted (5.16.11-s3k-dev-01743-g19beecbfe9d6-dirty) [ 16.357849] MSR: 00029032 <EE,ME,IR,DR,RI> CR: 24002282 XER: 00000000 [ 16.357931] [ 16.357931] GPR00: c03fbc9c e68d9be0 c26d06a0 00000039 00000001 c0d36364 c0e96428 00000027 [ 16.357931] GPR08: 00000001 00000000 00000023 3fffc000 24002282 100d3dd6 100a2ffc 00000000 [ 16.357931] GPR16: 100cd280 100b0000 00000000 aff54f7e 100d0000 100d0000 00000001 100cf328 [ 16.357931] GPR24: 100cf328 00000000 00000003 e68d9e30 c156b410 e67ab4c0 e68d9d38 c24ab278 [ 16.358253] NIP [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.358292] LR [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 [ 16.358325] Call Trace: [ 16.358336] [e68d9be0] [c03fbc9c] fsl_spi_cpm_bufs+0x2a0/0x2d8 (unreliable) [ 16.358388] [e68d9c00] [c03fcb44] fsl_spi_bufs.isra.0+0x94/0x1a0 [ 16.358436] [e68d9c20] [c03fd970] fsl_spi_do_one_msg+0x254/0x3dc [ 16.358483] [e68d9cb0] [c03f7e50] __spi_pump_messages+0x274/0x8a4 [ 16.358529] [e68d9ce0] [c03f9d30] __spi_sync+0x344/0x378 [ 16.358573] [e68d9d20] [c03fb52c] spi_sync+0x34/0x60 [ 16.358616] [e68d9d30] [c03b4dec] at25_ee_read+0x138/0x1a8 [ 16.358667] [e68d9e50] [c04a8fb8] bin_attr_nvmem_read+0x98/0x110 [ 16.358725] [e68d9e60] [c0204b14] kernfs_fop_read_iter+0xc0/0x1fc [ 16.358774] [e68d9e80] [c0168660] vfs_read+0x284/0x410 [ 16.358821] [e68d9f00] [c016925c] ksys_read+0x6c/0x11c [ 16.358863] [e68d9f30] [c00160e0] ret_from_syscall+0x0/0x28 ... [ 16.359608] ---[ end trace a4ce3e34afef0cb5 ]--- [ 16.359638] fsl_spi b01004c0.spi: unable to map tx dma This is due to the AT25 driver using buffers on stack, which is not possible with CONFIG_VMAP_STACK. As mentionned in kernel Documentation (Documentation/spi/spi-summary.rst): - Follow standard kernel rules, and provide DMA-safe buffers in your messages. That way controller drivers using DMA aren't forced to make extra copies unless the hardware requires it (e.g. working around hardware errata that force the use of bounce buffering). Modify the driver to use a buffer located in the at25 device structure which is allocated via kmalloc during probe. Protect writes in this new buffer with the driver's mutex. Fixes: b587b13a4f67 ("[PATCH] SPI eeprom driver") Cc: stable <stable@kernel.org> Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Link: https://lore.kernel.org/r/230a9486fc68ea0182df46255e42a51099403642.1648032613.git.christophe.leroy@csgroup.eu Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2022-03-23 11:51:55 +01:00
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
memcpy(buf, bounce, len);
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
return 0;
}
static ssize_t sernum_show(struct device *dev, struct device_attribute *attr, char *buf)
{
struct at25_data *at25;
at25 = dev_get_drvdata(dev);
return sysfs_emit(buf, "%*ph\n", (int)sizeof(at25->sernum), at25->sernum);
}
static DEVICE_ATTR_RO(sernum);
static struct attribute *sernum_attrs[] = {
&dev_attr_sernum.attr,
NULL,
};
ATTRIBUTE_GROUPS(sernum);
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
/*
* Poll Read Status Register with timeout
*
* Return:
* 0, if the chip is ready
* [positive] Status Register value as-is, if the chip is busy
* [negative] error code in case of read failure
*/
static int at25_wait_ready(struct at25_data *at25)
{
u8 *bounce __free(kfree) = kmalloc(1, GFP_KERNEL);
struct spi_mem_op op;
int status;
if (!bounce)
return -ENOMEM;
op = (struct spi_mem_op)SPI_MEM_OP(SPI_MEM_OP_CMD(AT25_RDSR, 1),
SPI_MEM_OP_NO_ADDR,
SPI_MEM_OP_NO_DUMMY,
SPI_MEM_OP_DATA_IN(1, bounce, 1));
read_poll_timeout(spi_mem_exec_op, status,
status || !(bounce[0] & AT25_SR_nRDY), false,
USEC_PER_MSEC, USEC_PER_MSEC * EE_TIMEOUT,
at25->spimem, &op);
if (status < 0)
return status;
if (!(bounce[0] & AT25_SR_nRDY))
return 0;
return bounce[0];
}
static int at25_ee_write(void *priv, unsigned int off, void *val, size_t count)
{
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
u8 *bounce __free(kfree) = kmalloc(min(count, io_limit), GFP_KERNEL);
struct at25_data *at25 = priv;
const char *buf = val;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
unsigned int buf_size;
int status;
if (unlikely(off >= at25->chip.byte_len))
return -EFBIG;
if ((off + count) > at25->chip.byte_len)
count = at25->chip.byte_len - off;
if (unlikely(!count))
return -EINVAL;
buf_size = at25->chip.page_size;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
if (!bounce)
return -ENOMEM;
/*
* For write, rollover is within the page ... so we write at
* most one page, then manually roll over to the next page.
*/
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
guard(mutex)(&at25->lock);
do {
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
struct spi_mem_op op = SPI_MEM_OP(SPI_MEM_OP_CMD(AT25_WREN, 1),
SPI_MEM_OP_NO_ADDR,
SPI_MEM_OP_NO_DUMMY,
SPI_MEM_OP_NO_DATA);
unsigned int segment;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
status = spi_mem_exec_op(at25->spimem, &op);
if (status < 0) {
dev_dbg(&at25->spimem->spi->dev, "WREN --> %d\n", status);
return status;
}
/* Write as much of a page as we can */
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
segment = buf_size - (off % buf_size);
if (segment > count)
segment = count;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
if (segment > io_limit)
segment = io_limit;
op = (struct spi_mem_op)SPI_MEM_OP(SPI_MEM_OP_CMD(at25_instr(at25, AT25_WRITE, off),
1),
SPI_MEM_OP_ADDR(at25->addrlen, off, 1),
SPI_MEM_OP_NO_DUMMY,
SPI_MEM_OP_DATA_OUT(segment, bounce, 1));
status = spi_mem_adjust_op_size(at25->spimem, &op);
if (status)
return status;
segment = op.data.nbytes;
memcpy(bounce, buf, segment);
status = spi_mem_exec_op(at25->spimem, &op);
dev_dbg(&at25->spimem->spi->dev, "write %u bytes at %u --> %d\n",
segment, off, status);
if (status)
return status;
/*
* REVISIT this should detect (or prevent) failed writes
* to read-only sections of the EEPROM...
*/
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
status = at25_wait_ready(at25);
if (status < 0) {
dev_err_probe(&at25->spimem->spi->dev, status,
"Read Status Redister command failed\n");
return status;
}
if (status) {
dev_dbg(&at25->spimem->spi->dev,
"Status %02x\n", status);
dev_err(&at25->spimem->spi->dev,
"write %u bytes offset %u, timeout after %u msecs\n",
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
segment, off, EE_TIMEOUT);
return -ETIMEDOUT;
}
off += segment;
buf += segment;
count -= segment;
} while (count > 0);
return status;
}
/*-------------------------------------------------------------------------*/
static int at25_fw_to_chip(struct device *dev, struct spi_eeprom *chip)
{
u32 val;
int err;
strscpy(chip->name, "at25", sizeof(chip->name));
err = device_property_read_u32(dev, "size", &val);
if (err)
err = device_property_read_u32(dev, "at25,byte-len", &val);
if (err) {
dev_err(dev, "Error: missing \"size\" property\n");
return err;
}
chip->byte_len = val;
err = device_property_read_u32(dev, "pagesize", &val);
if (err)
err = device_property_read_u32(dev, "at25,page-size", &val);
if (err) {
dev_err(dev, "Error: missing \"pagesize\" property\n");
return err;
}
chip->page_size = val;
err = device_property_read_u32(dev, "address-width", &val);
if (err) {
err = device_property_read_u32(dev, "at25,addr-mode", &val);
if (err) {
dev_err(dev, "Error: missing \"address-width\" property\n");
return err;
}
chip->flags = (u16)val;
} else {
switch (val) {
case 9:
chip->flags |= EE_INSTR_BIT3_IS_ADDR;
fallthrough;
case 8:
chip->flags |= EE_ADDR1;
break;
case 16:
chip->flags |= EE_ADDR2;
break;
case 24:
chip->flags |= EE_ADDR3;
break;
default:
dev_err(dev,
"Error: bad \"address-width\" property: %u\n",
val);
return -ENODEV;
}
if (device_property_present(dev, "read-only"))
chip->flags |= EE_READONLY;
}
return 0;
}
static int at25_fram_to_chip(struct device *dev, struct spi_eeprom *chip)
{
struct at25_data *at25 = container_of(chip, struct at25_data, chip);
u8 sernum[FM25_SN_LEN];
u8 id[FM25_ID_LEN];
int i;
strscpy(chip->name, "fm25", sizeof(chip->name));
/* Get ID of chip */
fm25_aux_read(at25, id, FM25_RDID, FM25_ID_LEN);
/* There are inside-out FRAM variations, detect them and reverse the ID bytes */
if (id[6] == 0x7f && id[2] == 0xc2)
for (i = 0; i < ARRAY_SIZE(id) / 2; i++) {
u8 tmp = id[i];
int j = ARRAY_SIZE(id) - i - 1;
id[i] = id[j];
id[j] = tmp;
}
if (id[6] != 0xc2) {
dev_err(dev, "Error: no Cypress FRAM (id %02x)\n", id[6]);
return -ENODEV;
}
switch (id[7]) {
case 0x21 ... 0x26:
chip->byte_len = BIT(id[7] - 0x21 + 4) * 1024;
break;
case 0x2a ... 0x30:
/* CY15B116QN ... CY15B116QN */
chip->byte_len = BIT(((id[7] >> 1) & 0xf) + 13);
break;
default:
dev_err(dev, "Error: unsupported size (id %02x)\n", id[7]);
return -ENODEV;
}
if (chip->byte_len > 64 * 1024)
chip->flags |= EE_ADDR3;
else
chip->flags |= EE_ADDR2;
if (id[8]) {
fm25_aux_read(at25, sernum, FM25_RDSN, FM25_SN_LEN);
/* Swap byte order */
for (i = 0; i < FM25_SN_LEN; i++)
at25->sernum[i] = sernum[FM25_SN_LEN - 1 - i];
}
chip->page_size = PAGE_SIZE;
return 0;
}
static const struct of_device_id at25_of_match[] = {
{ .compatible = "atmel,at25" },
{ .compatible = "cypress,fm25" },
{ }
};
MODULE_DEVICE_TABLE(of, at25_of_match);
static const struct spi_device_id at25_spi_ids[] = {
{ .name = "at25" },
{ .name = "fm25" },
{ }
};
MODULE_DEVICE_TABLE(spi, at25_spi_ids);
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
static int at25_probe(struct spi_mem *mem)
{
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
struct spi_device *spi = mem->spi;
struct spi_eeprom *pdata;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
struct at25_data *at25;
bool is_fram;
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
int err;
at25 = devm_kzalloc(&spi->dev, sizeof(*at25), GFP_KERNEL);
if (!at25)
return -ENOMEM;
at25->spimem = mem;
/*
* Ping the chip ... the status register is pretty portable,
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
* unlike probing manufacturer IDs.
*/
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
err = at25_wait_ready(at25);
if (err < 0)
return dev_err_probe(&spi->dev, err, "Read Status Register command failed\n");
if (err) {
dev_err(&spi->dev, "Not ready (%02x)\n", err);
return -ENXIO;
}
mutex_init(&at25->lock);
spi_set_drvdata(spi, at25);
is_fram = fwnode_device_is_compatible(dev_fwnode(&spi->dev), "cypress,fm25");
/* Chip description */
pdata = dev_get_platdata(&spi->dev);
if (pdata) {
at25->chip = *pdata;
} else {
if (is_fram)
err = at25_fram_to_chip(&spi->dev, &at25->chip);
else
err = at25_fw_to_chip(&spi->dev, &at25->chip);
if (err)
return err;
}
/* For now we only support 8/16/24 bit addressing */
if (at25->chip.flags & EE_ADDR1)
at25->addrlen = 1;
else if (at25->chip.flags & EE_ADDR2)
at25->addrlen = 2;
else if (at25->chip.flags & EE_ADDR3)
at25->addrlen = 3;
else {
dev_dbg(&spi->dev, "unsupported address type\n");
return -EINVAL;
}
at25->nvmem_config.type = is_fram ? NVMEM_TYPE_FRAM : NVMEM_TYPE_EEPROM;
at25->nvmem_config.name = dev_name(&spi->dev);
at25->nvmem_config.dev = &spi->dev;
at25->nvmem_config.read_only = at25->chip.flags & EE_READONLY;
at25->nvmem_config.root_only = true;
at25->nvmem_config.owner = THIS_MODULE;
at25->nvmem_config.compat = true;
at25->nvmem_config.base_dev = &spi->dev;
at25->nvmem_config.reg_read = at25_ee_read;
at25->nvmem_config.reg_write = at25_ee_write;
at25->nvmem_config.priv = at25;
at25->nvmem_config.stride = 1;
at25->nvmem_config.word_size = 1;
at25->nvmem_config.size = at25->chip.byte_len;
at25->nvmem = devm_nvmem_register(&spi->dev, &at25->nvmem_config);
if (IS_ERR(at25->nvmem))
return PTR_ERR(at25->nvmem);
dev_info(&spi->dev, "%d %s %s %s%s, pagesize %u\n",
nvmem: eeprom: at25: fix FRAM byte_len Commit fd307a4ad332 ("nvmem: prepare basics for FRAM support") added support for FRAM devices such as the Cypress FM25V. During testing, it was found that the FRAM detects properly, however reads and writes fail. Upon further investigation, two problem were found in at25_probe() routine. 1) In the case of an FRAM device without platform data, eg. fram == true && spi->dev.platform_data == NULL the stack local variable "struct spi_eeprom chip" is not initialized fully, prior to being copied into at25->chip. The chip.flags field in particular can cause problems. 2) The byte_len of FRAM is computed from its ID register, and is stored into the stack local "struct spi_eeprom chip" structure. This happens after the same structure has been copied into at25->chip. As a result, at25->chip.byte_len does not contain the correct length of the device. In turn this can cause checks at beginning of at25_ee_read() to fail (or equally, it could allow reads beyond the end of the device length). Fix both of these issues by eliminating the on-stack struct spi_eeprom. Instead use the one inside at25_data structure, which starts of zeroed. Fixes: fd307a4ad332 ("nvmem: prepare basics for FRAM support") Cc: stable <stable@vger.kernel.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Ralph Siemsen <ralph.siemsen@linaro.org> Link: https://lore.kernel.org/r/20211108181627.645638-1-ralph.siemsen@linaro.org Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2021-11-08 13:16:27 -05:00
(at25->chip.byte_len < 1024) ?
at25->chip.byte_len : (at25->chip.byte_len / 1024),
(at25->chip.byte_len < 1024) ? "Byte" : "KByte",
at25->chip.name, is_fram ? "fram" : "eeprom",
(at25->chip.flags & EE_READONLY) ? " (readonly)" : "",
at25->chip.page_size);
return 0;
}
/*-------------------------------------------------------------------------*/
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
static struct spi_mem_driver at25_driver = {
.spidrv = {
.driver = {
.name = "at25",
.of_match_table = at25_of_match,
.dev_groups = sernum_groups,
},
.id_table = at25_spi_ids,
},
.probe = at25_probe,
};
eeprom: at25: convert to spi-mem API Replace the RAW SPI accesses with spi-mem API. The latter will fall back to RAW SPI accesses if spi-mem callbacks are not implemented by a controller driver. Notable advantages: - read function now allocates a bounce buffer for SPI DMA compatibility, similar to write function; - the driver can now be used in conjunction with SPI controller drivers providing spi-mem API only, e.g. spi-nxp-fspi. - during the initial probe the driver polls busy/ready status bit for 25ms instead of giving up instantly and hoping that the FW didn't write the EEPROM Notes: - mutex_lock() has been dropped from fm25_aux_read() because the latter is only being called in probe phase and therefore cannot race with at25_ee_read() or at25_ee_write() Quick 4KB block size test with CY15B102Q 256KB F-RAM over spi_omap2_mcspi driver (no spi-mem ops provided, fallback to raw SPI inside spi-mem): OP | throughput, KB/s | change --------+-----------------------+------- write | 1717.847 -> 1656.684 | -3.6% read | 1115.868 -> 1059.367 | -5.1% The lower throughtput probably comes from the 3 messages per SPI transfer inside spi-mem instead of hand-crafted 2 messages per transfer in the former at25 code. However, if the raw SPI access is not preserved, then the driver doesn't grow from the lines-of-code perspective and subjectively could be considered even a bit simpler. Higher performance impact on the read operation could be explained by the newly introduced bounce buffer in read operation. I didn't find any explanation or guarantee, why would a bounce buffer be not needed on the read side, so I assume it's a pure luck that nobody read EEPROM into some variable on stack on an architecture where kernel stack would be not DMA-able. Cc: Michael Walle <mwalle@kernel.org> Cc: Hui Wang <hui.wang@canonical.com> Link: https://lore.kernel.org/all/28ab8b72afee1af59b628f7389f0d7f5@kernel.org/ Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com> Link: https://lore.kernel.org/r/20250702222823.864803-1-alexander.sverdlin@siemens.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2025-07-03 00:28:22 +02:00
module_spi_mem_driver(at25_driver);
MODULE_DESCRIPTION("Driver for most SPI EEPROMs");
MODULE_AUTHOR("David Brownell");
MODULE_LICENSE("GPL");