sfp: sfp_read: split-up request when hw rx buffer is too small.
Subsystems (4)
PATCH 1/1 sfp: sfp_read: split-up request when hw rx buffer is too small.
Without this patch sfp code retries to read the full struct sfp_eeprom_id id out of the SFP eeprom. Sizeof(id) is 96 bytes.
My i2c hardware, Mediatek mt7621, has a rx buffer of 64 bytes. So sfp_read gets -NOSUPPORTED back on his turn return -EAGAIN. Same issue is with the SFP_EXT_STATUS data which is 92 bytes.
By split-up the request in multiple smaller requests with a max size of i2c max_read_len, we can readout the SFP module successfully.
Tested with MT7621 and two Fiberstore modules SFP-GB-GE-T and SFP-GE-BX.
--- drivers/net/phy/sfp.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)
drivers/net/phy/sfp.c +22 -1
@@ -367,7 +367,28 @@ static void sfp_set_state(struct sfp *sfp, unsigned int state)static int sfp_read (struct sfp *sfp,bool a2,u8 addr,void *buf,size_t len) { -return sfp->read (sfp, a2, addr, buf, len); +const struct i2c_adapter_quirks *q = sfp->i2c->quirks; +int ret; +size_t rx_bytes =0 ; + +/* Many i2c hw have limited rx buffers, split-up request when needed. */ +while ((q->max_read_len) && (len > q->max_read_len)) { + ret = sfp->read (sfp, a2, addr, buf, q->max_read_len); +if (ret <0 ) +return ret; + rx_bytes += ret; + addr += q->max_read_len; + buf += q->max_read_len; + len -= q->max_read_len; + } + + ret = sfp->read (sfp, a2, addr, buf, len); +if (ret <0 ) +return ret; + + rx_bytes += ret; + +return rx_bytes; }static int sfp_write (struct sfp *sfp,bool a2,u8 addr,void *buf,size_t len)
Signature · 1 line
2.19.1
Andrew Lunn
maintainerQuoted 22 lines · René van Dorst
Without this patch sfp code retries to read the full struct
sfp_eeprom_idid out of the SFP eeprom. Sizeof(id) is 96 bytes.
My i2c hardware, Mediatek mt7621, has a rx buffer of 64 bytes. Sosfp_readgets -NOSUPPORTED back on his turn return-EAGAIN. Same issue is with theSFP_EXT_STATUSdata which is 92 bytes.By split-up the request in multiple smaller requests with a max size of i2c
max_read_len, we can readout the SFP module successfully.Tested with MT7621 and two Fiberstore modules SFP-GB-GE-T and SFP-GE-BX.
Signed-off-by René van Dorst <[email protected]>--- drivers/net/phy/sfp.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)drivers/net/phy/sfp.c +0
-0@@ -367,7 +367,28 @@ static void sfp_set_state(struct sfp *sfp, unsigned int state)
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 376
Quoted 10 lines
drivers/net/phy/sfp.c +7
-1static int sfp_read (struct sfp *sfp,bool a2,u8 addr,void *buf,size_t len) { -return sfp->read (sfp, a2, addr, buf, len); +const struct i2c_adapter_quirks *q = sfp->i2c->quirks; +int ret; +size_t rx_bytes =0 ; + +/* Many i2c hw have limited rx buffers, split-up request when needed. */ +while ((q->max_read_len) && (len > q->max_read_len)) { + ret = sfp->read (sfp, a2, addr, buf, q->max_read_len);
Hi René
I think you want to pass MIN(len, q->max_read_len) to read().
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 382
Quoted 6 lines
drivers/net/phy/sfp.c +6
-0+if (ret <0 ) +return ret; + rx_bytes += ret; + addr += q->max_read_len; + buf += q->max_read_len; + len -= q->max_read_len;
I would prefer you add ret, not q->max_read_len. There is a danger it returned less bytes than you asked for.
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 394
Quoted 12 lines
drivers/net/phy/sfp.c +9
-0+ } + + ret = sfp->read (sfp, a2, addr, buf, len); +if (ret <0 ) +return ret; + + rx_bytes += ret; + +return rx_bytes; }static int sfp_write (struct sfp *sfp,bool a2,u8 addr,void *buf,size_t len)
Doesn't write need the same handling?
Thanks
Andrew
Quoting Andrew Lunn <[email protected]>:
Quoted 38 lines · René van Dorst
Quoted 24 lines
Without this patch sfp code retries to read the full struct
sfp_eeprom_idid out of the SFP eeprom. Sizeof(id) is 96 bytes.
My i2c hardware, Mediatek mt7621, has a rx buffer of 64 bytes. Sosfp_readgets -NOSUPPORTED back on his turn return-EAGAIN. Same issue is with theSFP_EXT_STATUSdata which is 92 bytes.By split-up the request in multiple smaller requests with a max size of i2c
max_read_len, we can readout the SFP module successfully.Tested with MT7621 and two Fiberstore modules SFP-GB-GE-T and SFP-GE-BX.
Signed-off-by René van Dorst <[email protected]>--- drivers/net/phy/sfp.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)drivers/net/phy/sfp.c +0
-0@@ -367,7 +367,28 @@ static void sfp_set_state(struct sfp *sfp, unsigned int state) static int sfp_read(struct sfp *sfp, bool a2, u8 addr, void *buf,drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, · line 378
Quoted 10 lines
drivers/net/phy/sfp.c +7
-1size_t len) { -return sfp->read (sfp, a2, addr, buf, len); +const struct i2c_adapter_quirks *q = sfp->i2c->quirks; +int ret; +size_t rx_bytes =0 ; + +/* Many i2c hw have limited rx buffers, split-up request when needed. */ +while ((q->max_read_len) && (len > q->max_read_len)) { + ret = sfp->read (sfp, a2, addr, buf, q->max_read_len);Hi René
Quoted 1 line
I think you want to pass
MIN(len, q->max_read_len)toread().
Hi Andrew,
max_read_len is 0 when there is no quirk.
I can write it a bit differently depending on the outcome of my other email.
Quoted 9 lines
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 382
Quoted 6 lines
drivers/net/phy/sfp.c +6
-0+if (ret <0 ) +return ret; + rx_bytes += ret; + addr += q->max_read_len; + buf += q->max_read_len; + len -= q->max_read_len;I would prefer you add ret, not
q->max_read_len. There is a danger it returned less bytes than you asked for.
Getting less bytes then asked is already an error I think.
I could check the return size and directly return the number of bytes that I have. The callers are checking for size and they can retry if wanted. So that should not be an issue.
Quoted 11 lines
Quoted 11 lines
drivers/net/phy/sfp.c +9
-0+ } + + ret = sfp->read (sfp, a2, addr, buf, len); +if (ret <0 ) +return ret; + + rx_bytes += ret; + +return rx_bytes; }
Quoted 4 lines
Quoted 2 lines
static int sfp_write(struct sfp *sfp, bool a2, u8 addr, void *buf,
size_tlen)Doesn't write need the same handling?
By reading the SSF spec we can write to a user writable EERPOM area of 120 bytes. But the current code has only has 1 sfp_write for a byte value.
So for now I should say no.
Greats,
René
Andrew Lunn
maintainerQuoted 6 lines
Quoted 6 lines
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 376
Quoted 3 lines
drivers/net/phy/sfp.c +3
-0+/* Many i2c hw have limited rx buffers, split-up request when needed. */ +while ((q->max_read_len) && (len > q->max_read_len)) { + ret = sfp->read (sfp, a2, addr, buf, q->max_read_len);Hi René
Quoted 6 lines
Quoted 1 line
I think you want to pass
MIN(len, q->max_read_len)toread().Hi Andrew,
max_read_lenis 0 when there is no quirk.
I can write it a bit differently depending on the outcome of my other email.
Hi René
No, you misunderstood me.
Quoted 1 line
Quoted 1 line
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 376
Quoted 1 line
drivers/net/phy/sfp.c +1
-0+ ret = sfp->read (sfp, a2, addr, buf, q->max_read_len);
Say max_read_len = 64
The SFP code asks to read 68 bytes. The first call to read() is going to ask for 64 bytes. The second call is going to also ask for 64 bytes, writing 60 bytes passed the end of buf. Bad things then happen.
Quoted 11 lines
Quoted 10 lines
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 382
Quoted 7 lines
drivers/net/phy/sfp.c +6
-0+if (ret <0 ) +return ret; + rx_bytes += ret; + addr += q->max_read_len; + buf += q->max_read_len; + len -= q->max_read_len;I would prefer you add ret, not
q->max_read_len. There is a danger it returned less bytes than you asked for.
Quoted 4 lines
Getting less bytes then asked is already an error I think.
I could check the return size and directly return the number of bytes that I have. The callers are checking for size and they can retry if wanted. So that should not be an issue.
If that is true, why is rx_bytes += ret, where as all the others are += q->max_read_len. Please be consistent. The general pattern of a read function in POSIX systems is that it returns how many bytes were actually returned. So i would always use += ret.
Quoted 5 lines
By reading the SSF spec we can write to a user writable EERPOM area of 120 bytes.
But the current code has only has 1sfp_writefor a byte value.So for now I should say no.
So how about adding a WARN_ON. If the request is bigger than what the quirk allows, make it very obvious we have an issue.
Andrew
Quoting Andrew Lunn <[email protected]>:
Quoted 17 lines
Quoted 6 lines
Quoted 6 lines
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 376
Quoted 3 lines
drivers/net/phy/sfp.c +3
-0+/* Many i2c hw have limited rx buffers, split-up request when needed. */ +while ((q->max_read_len) && (len > q->max_read_len)) { + ret = sfp->read (sfp, a2, addr, buf, q->max_read_len);Hi René
Quoted 6 lines
Quoted 1 line
I think you want to pass
MIN(len, q->max_read_len)toread().Hi Andrew,
max_read_lenis 0 when there is no quirk.
I can write it a bit differently depending on the outcome of my other email.Hi René
No, you misunderstood me.
Quoted 7 lines
Quoted 1 line
Quoted 1 line
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 376
Quoted 1 line
drivers/net/phy/sfp.c +1
-0+ ret = sfp->read (sfp, a2, addr, buf, q->max_read_len);Say
max_read_len= 64The SFP code asks to read 68 bytes. The first call to
read()is going to ask for 64 bytes. The second call is going to also ask for 64 bytes, writing 60 bytes passed the end of buf. Bad things then happen.
Hi Andrew,
This can't happen, while loop is only executed when len > max_read_len. len is reduced after a successful read in the while loop.
So sizes <= max_read_len is handled by sfp->read below the while loop.
Quoted 25 lines
Quoted 11 lines
Quoted 10 lines
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 382
Quoted 7 lines
drivers/net/phy/sfp.c +6
-0+if (ret <0 ) +return ret; + rx_bytes += ret; + addr += q->max_read_len; + buf += q->max_read_len; + len -= q->max_read_len;I would prefer you add ret, not
q->max_read_len. There is a danger it returned less bytes than you asked for.Quoted 4 lines
Getting less bytes then asked is already an error I think.
I could check the return size and directly return the number of bytes that I have. The callers are checking for size and they can retry if wanted. So that should not be an issue.If that is true, why is
rx_bytes+= ret, where as all the others are +=q->max_read_len. Please be consistent. The general pattern of a read function in POSIX systems is that it returns how many bytes were actually returned. So i would always use += ret.Quoted 4 lines
By reading the SSF spec we can write to a user writable EERPOM area of 120 bytes.
But the current code has only has 1sfp_writefor a byte value.
Quoted 6 lines
Quoted 1 line
So for now I should say no.
So how about adding a
WARN_ON. If the request is bigger than what the quirk allows, make it very obvious we have an issue.Andrew
i2c-core already reports an error in i2c_check_for_quirks [0].
Is that sufficient?
This is how I know that something was wrong.
But the driver should not retry it for infinitely like what is happening with read. And keeps spamming the error logs.
Greats,
René
Quoted 29 lines · René van Dorst
Without this patch sfp code retries to read the full struct
sfp_eeprom_idid out of the SFP eeprom. Sizeof(id) is 96 bytes.
My i2c hardware, Mediatek mt7621, has a rx buffer of 64 bytes. Sosfp_readgets -NOSUPPORTED back on his turn return-EAGAIN. Same issue is with theSFP_EXT_STATUSdata which is 92 bytes.By split-up the request in multiple smaller requests with a max size of i2c
max_read_len, we can readout the SFP module successfully.Tested with MT7621 and two Fiberstore modules SFP-GB-GE-T and SFP-GE-BX.
Signed-off-by René van Dorst <[email protected]>--- drivers/net/phy/sfp.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)drivers/net/phy/sfp.c +4
-1@@ -367,7 +367,28 @@ static void sfp_set_state(struct sfp *sfp, unsigned int state)static int sfp_read (struct sfp *sfp,bool a2,u8 addr,void *buf,size_t len) { -return sfp->read (sfp, a2, addr, buf, len); +const struct i2c_adapter_quirks *q = sfp->i2c->quirks; +int ret; +size_t rx_bytes =0 ; +
drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, unsigned int state) · line 383
Quoted 10 lines
drivers/net/phy/sfp.c +10
-0+/* Many i2c hw have limited rx buffers, split-up request when needed. */ +while ((q->max_read_len) && (len > q->max_read_len)) { + ret = sfp->read (sfp, a2, addr, buf, q->max_read_len); +if (ret <0 ) +return ret; + rx_bytes += ret; + addr += q->max_read_len; + buf += q->max_read_len; + len -= q->max_read_len; + }
sfp->read defaults to sfp_i2c_read() but it is possible to override that by registering a custom sfp_bus instance (nothing does it today, but that could obviously change), so there is no guarantee that sfp->i2c->quirks is applicable unless sfp_i2c_read() is used.
sfp_i2c_read() is presumably where the max_read_len splitting should occur, or better yet, should not i2c_transfer() take care of that on its own? That way there would be no layering violation of having to fetch the quirk bitmask for the i2c adapter being used, that is something that should belong in the core i2c framework.
Quoted 12 lines
drivers/net/phy/sfp.c +8
-0+ + ret = sfp->read (sfp, a2, addr, buf, len); +if (ret <0 ) +return ret; + + rx_bytes += ret; + +return rx_bytes; }static int sfp_write (struct sfp *sfp,bool a2,u8 addr,void *buf,size_t len)
Signature · 1 line
Florian
Quoting Florian Fainelli <[email protected]>:
Quoted 48 lines · René van Dorst
Quoted 31 lines
Without this patch sfp code retries to read the full struct
sfp_eeprom_idid out of the SFP eeprom. Sizeof(id) is 96 bytes.
My i2c hardware, Mediatek mt7621, has a rx buffer of 64 bytes. Sosfp_readgets -NOSUPPORTED back on his turn return-EAGAIN. Same issue is with theSFP_EXT_STATUSdata which is 92 bytes.By split-up the request in multiple smaller requests with a max size of i2c
max_read_len, we can readout the SFP module successfully.Tested with MT7621 and two Fiberstore modules SFP-GB-GE-T and SFP-GE-BX.
Signed-off-by René van Dorst <[email protected]>--- drivers/net/phy/sfp.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)drivers/net/phy/sfp.c +4
-1@@ -367,7 +367,28 @@ static void sfp_set_state(struct sfp *sfp, unsigned int state)static int sfp_read (struct sfp *sfp,bool a2,u8 addr,void *buf,size_t len) { -return sfp->read (sfp, a2, addr, buf, len); +const struct i2c_adapter_quirks *q = sfp->i2c->quirks; +int ret; +size_t rx_bytes =0 ; +drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, · line 385
Quoted 10 lines
drivers/net/phy/sfp.c +10
-0+/* Many i2c hw have limited rx buffers, split-up request when needed. */ +while ((q->max_read_len) && (len > q->max_read_len)) { + ret = sfp->read (sfp, a2, addr, buf, q->max_read_len); +if (ret <0 ) +return ret; + rx_bytes += ret; + addr += q->max_read_len; + buf += q->max_read_len; + len -= q->max_read_len; + }
sfp->readdefaults tosfp_i2c_read()but it is possible to override that by registering a customsfp_businstance (nothing does it today, but that could obviously change), so there is no guarantee thatsfp->i2c->quirksis applicable unlesssfp_i2c_read()is used.
Quoted 5 lines
sfp_i2c_read()is presumably where themax_read_lensplitting should occur, or better yet, should noti2c_transfer()take care of that on its own? That way there would be no layering violation of having to fetch the quirk bitmask for the i2c adapter being used, that is something that should belong in the core i2c framework.
Yes it is better to put it in sfp_i2c_read().
I think it is best to handle the split-up within the driver. The driver knows how to talk to the device and may apply device quirks.
Also tda1004x [0] and TPM [1] driver also handles it within the driver itself.
TPM driver just try to send the want size and split-up request to I2C_SMBUS_BLOCK_MAX when a -EOPNOTSUPP returns, just retries it a number of times.
I can do the same but I have to pick a minimum size.
Looking in SSF-8472rev12.2.1 they don't limit the way you access the device. So use I2C_SMBUS_BLOCK_MAX of 32 bytes is sufficient or lookup the i2c quirk max_read_len is also an option.
Grepping thru the i2c busses I see only 2 devices which has less then 32 bytes of buffer. i2c-nvidia-gpu (Nvidia GPU) and i2c-pmcmsp (microcontroller MSP71xx). Both are unlikely to be used for these kind of applications.
I think I2C_SMBUS_BLOCK_MAX is safe to use.
Greats,
René
[0] https://elixir.bootlin.com/linux/v5.0-rc3/source/drivers/media/dvb-frontends/tda1004x.c#L319
[1] https://elixir.bootlin.com/linux/v5.0-rc3/source/drivers/char/tpm/tpm_i2c_infineon.c#L158
Quoting René van Dorst <[email protected]>:
Quoted 62 lines
Quoting Florian Fainelli <[email protected]>:
Quoted 47 lines · René van Dorst
Quoted 30 lines
Without this patch sfp code retries to read the full struct
sfp_eeprom_idid out of the SFP eeprom. Sizeof(id) is 96 bytes.
My i2c hardware, Mediatek mt7621, has a rx buffer of 64 bytes. Sosfp_readgets -NOSUPPORTED back on his turn return-EAGAIN. Same issue is with theSFP_EXT_STATUSdata which is 92 bytes.By split-up the request in multiple smaller requests with a max size of i2c
max_read_len, we can readout the SFP module successfully.Tested with MT7621 and two Fiberstore modules SFP-GB-GE-T and SFP-GE-BX.
Signed-off-by René van Dorst <[email protected]>---
drivers/net/phy/sfp.c| 23 ++++++++++++++++++++++-1 file changed, 22 insertions(+), 1 deletion(-)drivers/net/phy/sfp.c +4
-1@@ -367,7 +367,28 @@ static void sfp_set_state(struct sfp *sfp, unsigned int state)static int sfp_read (struct sfp *sfp,bool a2,u8 addr,void *buf,size_t len) { -return sfp->read (sfp, a2, addr, buf, len); +const struct i2c_adapter_quirks *q = sfp->i2c->quirks; +int ret; +size_t rx_bytes =0 ; +drivers/net/phy/sfp.c static void sfp_set_state(struct sfp *sfp, · line 384
Quoted 10 lines
drivers/net/phy/sfp.c +10
-0+/* Many i2c hw have limited rx buffers, split-up request when needed. */ +while ((q->max_read_len) && (len > q->max_read_len)) { + ret = sfp->read (sfp, a2, addr, buf, q->max_read_len); +if (ret <0 ) +return ret; + rx_bytes += ret; + addr += q->max_read_len; + buf += q->max_read_len; + len -= q->max_read_len; + }
sfp->readdefaults tosfp_i2c_read()but it is possible to override that by registering a customsfp_businstance (nothing does it today, but that could obviously change), so there is no guarantee thatsfp->i2c->quirksis applicable unlesssfp_i2c_read()is used.Quoted 5 lines
sfp_i2c_read()is presumably where themax_read_lensplitting should occur, or better yet, should noti2c_transfer()take care of that on its own? That way there would be no layering violation of having to fetch the quirk bitmask for the i2c adapter being used, that is something that should belong in the core i2c framework.Yes it is better to put it in
sfp_i2c_read().I think it is best to handle the split-up within the driver. The driver knows how to talk to the device and may apply device quirks.
Also tda1004x [0] and TPM [1] driver also handles it within the driver itself.
Quoted 8 lines
TPM driver just try to send the want size and split-up request to
I2C_SMBUS_BLOCK_MAXwhen a-EOPNOTSUPPreturns, just retries it a number of times.I can do the same but I have to pick a minimum size.
Looking in SSF-8472rev12.2.1 they don't limit the way you access the device. So useI2C_SMBUS_BLOCK_MAXof 32 bytes is sufficient or lookup the i2c quirkmax_read_lenis also an option.
Hi Florian,
I did a test in sfp_i2c_read() with reducing the read size when -EOPNOTSUPP returns like TPM driver. But this always produces noise in kernel log about the "msg too long" if the device is init at boot or when plugged-in. Devices with SFP_EXT_STATUS generates also an 2nd error.
So I prefer to use the quirk max_read_len.
dmesg output reduce read size when -EOPNOTSUPP returns.
[ 134.220000] i2c i2c-0: adapter quirk: msg too long (addr 0x0050, size 96, read) [ 135.250000] sfp sfp-lan5: module FiberStore SFP-GB-GE-T rev B sn <snip> dc <snip> [ 141.150000] sfp sfp-lan5: module removed [ 150.950000] i2c i2c-0: adapter quirk: msg too long (addr 0x0050, size 96, read) [ 151.980000] sfp sfp-lan5: module FiberStore SFP-GE-BX rev A0 sn <snip> dc <snip> [ 151.990000] i2c i2c-0: adapter quirk: msg too long (addr 0x0051, size 92, read)
dmesg output with read size always reduced to quiek max_read_len.
[ 27.350000] sfp sfp-lan5: module FiberStore SFP-GE-BX rev A0 sn <snip> dc <snip> [ 34.540000] sfp sfp-lan5: module removed [ 39.360000] sfp sfp-lan5: module FiberStore SFP-GB-GE-T rev B sn <snip> dc <snip>
So shall I spin v2 with quirk max_read_len as max read size?
Greats,
René
Quoted 12 lines
Grepping thru the i2c busses I see only 2 devices which has less then 32 bytes of buffer. i2c-nvidia-gpu (Nvidia GPU) and i2c-pmcmsp (microcontroller MSP71xx). Both are unlikely to be used for these kind of applications.
I think
I2C_SMBUS_BLOCK_MAXis safe to use.Greats,
René
[0] https://elixir.bootlin.com/linux/v5.0-rc3/source/drivers/media/dvb-frontends/tda1004x.c#L319
[1] https://elixir.bootlin.com/linux/v5.0-rc3/source/drivers/char/tpm/tpm_i2c_infineon.c#L158
Quoted 81 lines · René van Dorst
Quoting René van Dorst <[email protected]>:
Quoted 73 lines
Quoting Florian Fainelli <[email protected]>:
Quoted 52 lines · René van Dorst
Quoted 37 lines
Without this patch sfp code retries to read the full struct
sfp_eeprom_id
id out of the SFP eeprom. Sizeof(id) is 96 bytes.
My i2c hardware, Mediatek mt7621, has a rx buffer of 64 bytes. Sosfp_readgets -NOSUPPORTED back on his turn return-EAGAIN. Same issue is with theSFP_EXT_STATUSdata which is 92 bytes.By split-up the request in multiple smaller requests with a max size of i2c
max_read_len, we can readout the SFP module successfully.Tested with MT7621 and two Fiberstore modules SFP-GB-GE-T and SFP-GE-BX.
Signed-off-by René van Dorst <[email protected]>---
drivers/net/phy/sfp.c| 23 ++++++++++++++++++++++-1 file changed, 22 insertions(+), 1 deletion(-)drivers/net/phy/sfp.c +5
-1@@ -367,7 +367,28 @@ static void sfp_set_state(struct sfp *sfp, unsigned int state)static int sfp_read (struct sfp *sfp,bool a2,u8 addr,void *buf,size_t len) { - return sfp->read(sfp, a2, addr, buf, len); + const struct i2c_adapter_quirks *q = sfp->i2c->quirks; + int ret; + size_t rx_bytes = 0; + + /* Many i2c hw have limited rx buffers, split-up request whenneeded. */
+ while ((q->max_read_len) && (len >q->max_read_len)) {Quoted 8 lines
+ ret =
sfp->read(sfp, a2, addr, buf, q->max_read_len); + if (ret < 0)
+ return ret;
+rx_bytes+= ret;
+ addr +=q->max_read_len;
+ buf +=q->max_read_len;
+ len -=q->max_read_len;
+ }
sfp->readdefaults tosfp_i2c_read()but it is possible to override that by registering a customsfp_businstance (nothing does it today, but that could obviously change), so there is no guarantee thatsfp->i2c->quirksis applicable unlesssfp_i2c_read()is used.Quoted 5 lines
sfp_i2c_read()is presumably where themax_read_lensplitting should occur, or better yet, should noti2c_transfer()take care of that on its own? That way there would be no layering violation of having to fetch the quirk bitmask for the i2c adapter being used, that is something that should belong in the core i2c framework.Yes it is better to put it in
sfp_i2c_read().I think it is best to handle the split-up within the driver. The driver knows how to talk to the device and may apply device quirks.
Also tda1004x [0] and TPM [1] driver also handles it within the driver itself.
TPM driver just try to send the want size and split-up request to
I2C_SMBUS_BLOCK_MAXwhen a-EOPNOTSUPPreturns, just retries it a number of
times.Quoted 5 lines
I can do the same but I have to pick a minimum size. Looking in SSF-8472rev12.2.1 they don't limit the way you access the device.
So useI2C_SMBUS_BLOCK_MAXof 32 bytes is sufficient or lookup the i2c quirkmax_read_lenis also an option.
Quoted 7 lines
Hi Florian,
I did a test in
sfp_i2c_read()with reducing the read size when-EOPNOTSUPPreturns like TPM driver. But this always produces noise in kernel log about the "msg too long" if the device is init at boot or when plugged-in. Devices withSFP_EXT_STATUSgenerates also an 2nd error.
So I prefer to use the quirkmax_read_len.
My point still stands, we have an abstraction layer through the i2c core which sits between clients and adapters.
If you have a simple read into a buffer transaction (like what we have here), then you can provide a safe accessor function that takes care of looking at the max_read_len quirk and splitting things into the appropriate number of transaction. Then we can also eliminate the "msg too long" annoying message since we are now using an appropriate function.
The fact that there are multiple drivers that need to be patched to split i2c transaction is an indication that this is not IMHO addressed at the right level, especially if what they do is as simple as an i2c_transfer() into a single buffer and require no quirky transaction.
Andrew, Heiner and Russell might have a different view on this.
Quoted 22 lines
dmesg output reduce read size when
-EOPNOTSUPPreturns.
[ 134.220000] i2c i2c-0: adapter quirk: msg too long (addr0x0050, size 96, read)
[ 135.250000] sfp sfp-lan5: module FiberStore SFP-GB-GE-T rev B sn <snip> dc <snip>
[ 141.150000] sfp sfp-lan5: module removed
[ 150.950000] i2c i2c-0: adapter quirk: msg too long (addr0x0050, size 96, read)
[ 151.980000] sfp sfp-lan5: module FiberStore SFP-GE-BX rev A0 sn <snip> dc <snip>
[ 151.990000] i2c i2c-0: adapter quirk: msg too long (addr0x0051, size 92, read)dmesg output with read size always reduced to quiek
max_read_len. [ 27.350000] sfp sfp-lan5: module FiberStore SFP-GE-BX rev A0 sn <snip> dc <snip>
[ 34.540000] sfp sfp-lan5: module removed
[ 39.360000] sfp sfp-lan5: module FiberStore SFP-GB-GE-T rev B sn <snip> dc <snip>So shall I spin v2 with quirk
max_read_lenas max read size?
Signature · 1 line
Florian
Andrew Lunn
maintainerQuoted 9 lines
My point still stands, we have an abstraction layer through the i2c core which sits between clients and adapters.
If you have a simple read into a buffer transaction (like what we have here), then you can provide a safe accessor function that takes care of looking at the
max_read_lenquirk and splitting things into the appropriate number of transaction. Then we can also eliminate the "msg too long" annoying message since we are now using an appropriate function.
Quoted 6 lines
The fact that there are multiple drivers that need to be patched to split i2c transaction is an indication that this is not IMHO addressed at the right level, especially if what they do is as simple as an
i2c_transfer()into a single buffer and require no quirky transaction.Andrew, Heiner and Russell might have a different view on this.
Hi Florian
I suspect the issue is, only the driver knows if the device supports splitting a read into multiple reads without having to do something in between. Some device might require you do an address setup between each read, for example.
However, the core could offer an i2c_transfer_segmentable(), which does the segmentation as required here, and the driver can then elect to use that, not i2c_transfer().
Andrew