Message ID | 1888937035.82162.1441309955075.JavaMail.zimbra@xes-inc.com |
---|---|
State | Superseded |
Headers | show |
Hi Aaron, Sorry for the late reply. On Thu, 3 Sep 2015 14:52:35 -0500 (CDT), Aaron Sierra wrote: > Introduce at24_smbus_write_i2c_block_data() to allow very slow write > access to 16-bit EEPROM devices attached to SMBus controllers like > the Intel I801. This would be very bad hardware design. Are there actual systems out there which use large EEPROMs over SMBus? I would only consider adding this feature to the at24 driver if there is a real-world use case. Otherwise the same can be done from user-space with eeprog. > > With an AT24C512 device: > 248 B/s with 1-byte page (default) > 3.9 KB/s with 128-byte* page (via platform data) > > *limited to 16-bytes by I2C_SMBUS_BLOCK_MAX / 2. > > Signed-off-by: Aaron Sierra <asierra@xes-inc.com> > --- > drivers/misc/eeprom/at24.c | 35 ++++++++++++++++++++++++++++++++--- > 1 file changed, 32 insertions(+), 3 deletions(-) > > diff --git a/drivers/misc/eeprom/at24.c b/drivers/misc/eeprom/at24.c > index 2d3db81..4cf53a0 100644 > --- a/drivers/misc/eeprom/at24.c > +++ b/drivers/misc/eeprom/at24.c > @@ -134,6 +134,34 @@ MODULE_DEVICE_TABLE(i2c, at24_ids); > /*-------------------------------------------------------------------------*/ > > /* > + * Write block data to an AT24 device using SMBus cycles. > + */ > +static inline s32 at24_smbus_write_i2c_block_data(struct at24_data *at24, > + const struct i2c_client *client, u16 off, u8 len, const u8 *vals) > +{ > + u8 *addr16; > + s32 res; > + > + if (!(at24->chip.flags & AT24_FLAG_ADDR16)) > + return i2c_smbus_write_i2c_block_data(client, off, len, vals); > + > + addr16 = kzalloc(len + 1, GFP_KERNEL); > + if (!addr16) > + return -ENOMEM; Allocating and freeing memory with every transfer is a bad idea, as this slows the operation even more and favors memory fragmentation. Why don't you use at24_data.writebuf? > + > + /* Insert extra address byte into data stream */ > + addr16[0] = off & 0xff; > + memcpy(addr16 + 1, vals, len); > + > + res = i2c_smbus_write_i2c_block_data(client, > + (off >> 8) & 0xff, len + 1, addr16); Useless masking. > + > + kfree(addr16); > + > + return res; > +} > + > +/* > * This routine supports chips which consume multiple I2C addresses. It > * computes the addressing information to be used for a given r/w request. > * Assumes that sanity checks for offset happened at sysfs-layer. > @@ -369,8 +397,8 @@ static ssize_t at24_eeprom_write(struct at24_data *at24, const char *buf, > if (at24->use_smbus_write) { > switch (at24->use_smbus_write) { > case I2C_SMBUS_I2C_BLOCK_DATA: > - status = i2c_smbus_write_i2c_block_data(client, > - offset, count, buf); > + status = at24_smbus_write_i2c_block_data(at24, > + client, offset, count, buf); > break; > case I2C_SMBUS_BYTE_DATA: > status = i2c_smbus_write_byte_data(client, > @@ -612,7 +640,8 @@ static int at24_probe(struct i2c_client *client, const struct i2c_device_id *id) > if (write_max > io_limit) > write_max = io_limit; > if (use_smbus && write_max > I2C_SMBUS_BLOCK_MAX) > - write_max = I2C_SMBUS_BLOCK_MAX; > + write_max = I2C_SMBUS_BLOCK_MAX >> > + !!(chip.flags & AT24_FLAG_ADDR16); Looks bogus when AT24_FLAG_ADDR16 flag is set, in two ways: 1* I2C_SMBUS_BLOCK_MAX is 32. If write_max was 33, you'll set it to 33 >> 1 = 16. But if write_max was 31, you'll leave it to 31. Makes no sense to me. 2* Why I2C_SMBUS_BLOCK_MAX >> 1 in the first place? You only need to steal one byte from the data buffer for the extra address byte, so I'd expect write_max to be capped at I2C_SMBUS_BLOCK_MAX - 1. > at24->write_max = write_max; > > /* buffer (data + address at the beginning) */ The code you added will never be executed anyway, because of this test in at24_probe(): if (!i2c_check_functionality(client->adapter, I2C_FUNC_I2C)) { if (chip.flags & AT24_FLAG_ADDR16) return -EPFNOSUPPORT;
----- Original Message ----- > From: "Jean Delvare" <jdelvare@suse.de> > Sent: Monday, November 2, 2015 7:42:09 AM > > Hi Aaron, > > Sorry for the late reply. > > On Thu, 3 Sep 2015 14:52:35 -0500 (CDT), Aaron Sierra wrote: > > Introduce at24_smbus_write_i2c_block_data() to allow very slow write > > access to 16-bit EEPROM devices attached to SMBus controllers like > > the Intel I801. > > This would be very bad hardware design. Are there actual systems out > there which use large EEPROMs over SMBus? I would only consider adding > this feature to the at24 driver if there is a real-world use case. > Otherwise the same can be done from user-space with eeprog. Jean, No problem. Thanks for taking the time to consider these patches. We've used the AT24C128 and larger devices on all of our embedded x86 products since 2010. For their use in our products, sacrificing performance for higher storage capacity is an acceptable tradeoff. We've used variations of this patchset since 2.6.30, making changes along the way to adapt to upstream changes (which may explain some of the warts you've pointed out). Based on commit log messages and references within the driver to the challenges associated with supporting 16-bit devices on SMBus, I thought our patches might prove useful to the larger Linux community. > > > > With an AT24C512 device: > > 248 B/s with 1-byte page (default) > > 3.9 KB/s with 128-byte* page (via platform data) > > > > *limited to 16-bytes by I2C_SMBUS_BLOCK_MAX / 2. > > > > Signed-off-by: Aaron Sierra <asierra@xes-inc.com> > > --- > > drivers/misc/eeprom/at24.c | 35 ++++++++++++++++++++++++++++++++--- > > 1 file changed, 32 insertions(+), 3 deletions(-) > > > > diff --git a/drivers/misc/eeprom/at24.c b/drivers/misc/eeprom/at24.c > > index 2d3db81..4cf53a0 100644 > > --- a/drivers/misc/eeprom/at24.c > > +++ b/drivers/misc/eeprom/at24.c > > @@ -134,6 +134,34 @@ MODULE_DEVICE_TABLE(i2c, at24_ids); > > /*-------------------------------------------------------------------------*/ > > > > /* > > + * Write block data to an AT24 device using SMBus cycles. > > + */ > > +static inline s32 at24_smbus_write_i2c_block_data(struct at24_data *at24, > > + const struct i2c_client *client, u16 off, u8 len, const u8 *vals) > > +{ > > + u8 *addr16; > > + s32 res; > > + > > + if (!(at24->chip.flags & AT24_FLAG_ADDR16)) > > + return i2c_smbus_write_i2c_block_data(client, off, len, vals); > > + > > + addr16 = kzalloc(len + 1, GFP_KERNEL); > > + if (!addr16) > > + return -ENOMEM; > > Allocating and freeing memory with every transfer is a bad idea, as > this slows the operation even more and favors memory fragmentation. > Why don't you use at24_data.writebuf? I didn't like the repeated allocation, either, but performance was already expected to be poor. There doesn't seem to be any reason that we can't use .writebuf. > > + > > + /* Insert extra address byte into data stream */ > > + addr16[0] = off & 0xff; > > + memcpy(addr16 + 1, vals, len); > > + > > + res = i2c_smbus_write_i2c_block_data(client, > > + (off >> 8) & 0xff, len + 1, addr16); > > Useless masking. Agreed. > > + > > + kfree(addr16); > > + > > + return res; > > +} > > + > > +/* > > * This routine supports chips which consume multiple I2C addresses. It > > * computes the addressing information to be used for a given r/w request. > > * Assumes that sanity checks for offset happened at sysfs-layer. > > @@ -369,8 +397,8 @@ static ssize_t at24_eeprom_write(struct at24_data > > *at24, const char *buf, > > if (at24->use_smbus_write) { > > switch (at24->use_smbus_write) { > > case I2C_SMBUS_I2C_BLOCK_DATA: > > - status = i2c_smbus_write_i2c_block_data(client, > > - offset, count, buf); > > + status = at24_smbus_write_i2c_block_data(at24, > > + client, offset, count, buf); > > break; > > case I2C_SMBUS_BYTE_DATA: > > status = i2c_smbus_write_byte_data(client, > > @@ -612,7 +640,8 @@ static int at24_probe(struct i2c_client *client, const > > struct i2c_device_id *id) > > if (write_max > io_limit) > > write_max = io_limit; > > if (use_smbus && write_max > I2C_SMBUS_BLOCK_MAX) > > - write_max = I2C_SMBUS_BLOCK_MAX; > > + write_max = I2C_SMBUS_BLOCK_MAX >> > > + !!(chip.flags & AT24_FLAG_ADDR16); > > Looks bogus when AT24_FLAG_ADDR16 flag is set, in two ways: > > 1* I2C_SMBUS_BLOCK_MAX is 32. If write_max was 33, you'll set it to > 33 >> 1 = 16. But if write_max was 31, you'll leave it to 31. Makes > no sense to me. That's true, I hadn't considered that case. I assumed page_size would be a power-of-two based on its name, though that isn't enforced anywhere. > 2* Why I2C_SMBUS_BLOCK_MAX >> 1 in the first place? You only need to > steal one byte from the data buffer for the extra address byte, so > I'd expect write_max to be capped at I2C_SMBUS_BLOCK_MAX - 1. I think that I had a vague concern about sending a non-power-of-two amount of data per transfer. I don't see why that should be an issue. > > at24->write_max = write_max; > > > > /* buffer (data + address at the beginning) */ > > The code you added will never be executed anyway, because of this test > in at24_probe(): > > if (!i2c_check_functionality(client->adapter, I2C_FUNC_I2C)) { > if (chip.flags & AT24_FLAG_ADDR16) > return -EPFNOSUPPORT; That's true, if this patch is considered by itself. This test is updated by the 3rd patch in the series of three that I submitted. -Aaron S. P.S. I submitted a v2 of these patches, but only 3/3 is affected. -- To unsubscribe from this list: send the line "unsubscribe linux-i2c" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Aaron, On Mon, 2 Nov 2015 10:35:22 -0600 (CST), Aaron Sierra wrote: > ----- Original Message ----- > > From: "Jean Delvare" <jdelvare@suse.de> > > Sent: Monday, November 2, 2015 7:42:09 AM > > On Thu, 3 Sep 2015 14:52:35 -0500 (CDT), Aaron Sierra wrote: > > > @@ -612,7 +640,8 @@ static int at24_probe(struct i2c_client *client, const > > > struct i2c_device_id *id) > > > if (write_max > io_limit) > > > write_max = io_limit; > > > if (use_smbus && write_max > I2C_SMBUS_BLOCK_MAX) > > > - write_max = I2C_SMBUS_BLOCK_MAX; > > > + write_max = I2C_SMBUS_BLOCK_MAX >> > > > + !!(chip.flags & AT24_FLAG_ADDR16); > > > > Looks bogus when AT24_FLAG_ADDR16 flag is set, in two ways: > > > > 1* I2C_SMBUS_BLOCK_MAX is 32. If write_max was 33, you'll set it to > > 33 >> 1 = 16. But if write_max was 31, you'll leave it to 31. Makes > > no sense to me. > > That's true, I hadn't considered that case. I assumed page_size would > be a power-of-two based on its name, though that isn't enforced anywhere. It is checked but not enforced: if (!is_power_of_2(chip.page_size)) dev_warn(&client->dev, "page_size looks suspicious (no power of 2)!\n"); The thing is that I don't think a write operation can cross a page boundary. But as I understand it, this piece of code in at24_eeprom_write() takes care of that: /* Never roll over backwards, to the start of this page */ next_page = roundup(offset + 1, at24->chip.page_size); if (offset + count > next_page) count = next_page - offset; So in the above example, if page size is 32 and write_max is 31, you'll still need 2 write operations per page. If you write the whole eeprom, that's no different from having write_max at 16. But if page size is 64 or more, having write_max at 31 instead of 16 will improve the performance. Also think of partial writes, they can always benefit from greater write_max. > > 2* Why I2C_SMBUS_BLOCK_MAX >> 1 in the first place? You only need to > > steal one byte from the data buffer for the extra address byte, so > > I'd expect write_max to be capped at I2C_SMBUS_BLOCK_MAX - 1. > > I think that I had a vague concern about sending a non-power-of-two amount > of data per transfer. I don't see why that should be an issue. Should indeed not be an issue as long as you don't cross a page boundary. > > (...) > > The code you added will never be executed anyway, because of this test > > in at24_probe(): > > > > if (!i2c_check_functionality(client->adapter, I2C_FUNC_I2C)) { > > if (chip.flags & AT24_FLAG_ADDR16) > > return -EPFNOSUPPORT; > > That's true, if this patch is considered by itself. This test is > updated by the 3rd patch in the series of three that I submitted. The whole point of splitting the changes into multiple patches is that they can be applied, tested and accepted (or rejected) independently. Patches can depend on previous patch in the series but not on following patches. In general I am very much in favor of splitting the changes to make reviews easier, but for this patch series, if there is no easy way to split the changes, I believe it would make more sense to go with a single patch. > P.S. I submitted a v2 of these patches, but only 3/3 is affected. I don't think I ever received v2 :(
diff --git a/drivers/misc/eeprom/at24.c b/drivers/misc/eeprom/at24.c index 2d3db81..4cf53a0 100644 --- a/drivers/misc/eeprom/at24.c +++ b/drivers/misc/eeprom/at24.c @@ -134,6 +134,34 @@ MODULE_DEVICE_TABLE(i2c, at24_ids); /*-------------------------------------------------------------------------*/ /* + * Write block data to an AT24 device using SMBus cycles. + */ +static inline s32 at24_smbus_write_i2c_block_data(struct at24_data *at24, + const struct i2c_client *client, u16 off, u8 len, const u8 *vals) +{ + u8 *addr16; + s32 res; + + if (!(at24->chip.flags & AT24_FLAG_ADDR16)) + return i2c_smbus_write_i2c_block_data(client, off, len, vals); + + addr16 = kzalloc(len + 1, GFP_KERNEL); + if (!addr16) + return -ENOMEM; + + /* Insert extra address byte into data stream */ + addr16[0] = off & 0xff; + memcpy(addr16 + 1, vals, len); + + res = i2c_smbus_write_i2c_block_data(client, + (off >> 8) & 0xff, len + 1, addr16); + + kfree(addr16); + + return res; +} + +/* * This routine supports chips which consume multiple I2C addresses. It * computes the addressing information to be used for a given r/w request. * Assumes that sanity checks for offset happened at sysfs-layer. @@ -369,8 +397,8 @@ static ssize_t at24_eeprom_write(struct at24_data *at24, const char *buf, if (at24->use_smbus_write) { switch (at24->use_smbus_write) { case I2C_SMBUS_I2C_BLOCK_DATA: - status = i2c_smbus_write_i2c_block_data(client, - offset, count, buf); + status = at24_smbus_write_i2c_block_data(at24, + client, offset, count, buf); break; case I2C_SMBUS_BYTE_DATA: status = i2c_smbus_write_byte_data(client, @@ -612,7 +640,8 @@ static int at24_probe(struct i2c_client *client, const struct i2c_device_id *id) if (write_max > io_limit) write_max = io_limit; if (use_smbus && write_max > I2C_SMBUS_BLOCK_MAX) - write_max = I2C_SMBUS_BLOCK_MAX; + write_max = I2C_SMBUS_BLOCK_MAX >> + !!(chip.flags & AT24_FLAG_ADDR16); at24->write_max = write_max; /* buffer (data + address at the beginning) */
Introduce at24_smbus_write_i2c_block_data() to allow very slow write access to 16-bit EEPROM devices attached to SMBus controllers like the Intel I801. With an AT24C512 device: 248 B/s with 1-byte page (default) 3.9 KB/s with 128-byte* page (via platform data) *limited to 16-bytes by I2C_SMBUS_BLOCK_MAX / 2. Signed-off-by: Aaron Sierra <asierra@xes-inc.com> --- drivers/misc/eeprom/at24.c | 35 ++++++++++++++++++++++++++++++++--- 1 file changed, 32 insertions(+), 3 deletions(-)