Date:	Thu, 10 May 2007 00:47:59 -0400
From:	Michael Wu <flamingice@...rmilk.net>
To:	Jeff Garzik <jeff@...zik.org>
Cc:	"John W. Linville" <linville@...driver.com>,
	netdev@...r.kernel.org, linux-wireless@...r.kernel.org,
	David Miller <davem@...emloft.net>
Subject: Re: Please pull 'upstream-rtl8187' branch of wireless-2.6

On Thursday 10 May 2007 00:21, Jeff Garzik wrote:
> Michael Wu wrote:
> > On Wednesday 09 May 2007 19:12, Jeff Garzik wrote:
> >> John W. Linville wrote:
> >>> +static inline void eeprom_93cx6_pulse_high(struct eeprom_93cx6
> >>> *eeprom) +{
> >>> +	eeprom->reg_data_clock = 1;
> >>> +	eeprom->register_write(eeprom);
> >>> +	udelay(1);
> >>> +}
> >>> +
> >>> +static inline void eeprom_93cx6_pulse_low(struct eeprom_93cx6 *eeprom)
> >>> +{
> >>> +	eeprom->reg_data_clock = 0;
> >>> +	eeprom->register_write(eeprom);
> >>> +	udelay(1);
> >>> +}
> >>
> >> udelay-after-write normally indicates a PCI posting bug (or USB bus
> >> delay bug)
> >
> > Things may go bad if we try to bitbang the eeprom too fast.
>
> You are misunderstanding the point here.
>
> I was not asking why you needed the delay.  I was making an assertion
> about the unreliable nature of such delays.
>
> Sending a register write to a device is vastly different from knowing
> that the write has been received and processed by the device.
>
> On a PCI bus, this is called PCI write posting, where writes can be
> delayed and/or combined with other writes.  There is a similar concept
> with USB, because you are dealing with packets going to and from a USB
> device.
>
> You fundamentally cannot assume that the write has arrived at the device
> by the time the CPU is executing the next C statement (udelay) on the host.
>
> The normal procedure for PCI is to issue a read, which thereby
> guarantees that any writes have been flushed to the device.  I presume
> the same technique works with USB, but do not know for certain.
>
Ah, I see what you're talking about.

All register writes on rtl8187 are blocking. We know they've reached the 
device once the function returns.

I suppose there is a problem with the eeprom code not documenting the 
assumption that the write posted before returning.

> Regardless of any symbol issues, if it works without it, then remove it,
> because it is pointless code.  All pointless code should be removed.
>
> But nonetheless, for any "magic number" registers that remain, give them
> named constants that approximate their use as best as you can tell.
>
Well, I really don't know what this stuff does, so the best names would end up 
being something like RTL818X_INIT_MAGIC_REGISTER_1 and so on.

I'll ask Realtek instead.

> >> This seems an invalid use of skb->cb.  Don't use skb->cb.
> >
> > The driver owns these rx skbs, so it owns skb->cb. Why not?
>
> Wrong.  It no longer owns the skbs once it passes them to the stack.
>
Those skbs have not been passed up to the stack. They're empty skbs waiting 
for incoming frames. Any skbs that are passed up to the stack are unlinked 
from this list first.

> >>> +++ b/drivers/net/wireless/rtl818x.h
> >>> @@ -0,0 +1,212 @@
> >>> +#ifndef RTL818X_H
> >>> +#define RTL818X_H
> >>> +
> >>> +struct rtl818x_csr {
> >>> +	u8	MAC[6];
> >>> +	u8	reserved_0[2];
> >>> +	__le32	MAR[2];
> >>> +	u8	RX_FIFO_COUNT;
> >>> +	u8	reserved_1;
> >>> +	u8	TX_FIFO_COUNT;
> >>> +	u8	BQREQ;
> >>> +	u8	reserved_2[4];
> >>> +	__le32	TSFT[2];
> >>> +	__le32	TLPDA;
> >>> +	__le32	TNPDA;
> >>> +	__le32	THPDA;
> >>> +	__le16	BRSR;
> >>> +	u8	BSSID[6];
> >>> +	u8	RESP_RATE;
> >>> +	u8	EIFS;
> >>> +	u8	reserved_3[1];
> >>> +	u8	CMD;
> >>> +#define RTL818X_CMD_TX_ENABLE		(1 << 2)
> >>> +#define RTL818X_CMD_RX_ENABLE		(1 << 3)
> >>> +#define RTL818X_CMD_RESET		(1 << 4)
> >>> +	u8	reserved_4[4];
> >>> +	__le16	INT_MASK;
> >>> +	__le16	INT_STATUS;
> >>> +#define RTL818X_INT_RX_OK		(1 <<  0)
> >>> +#define RTL818X_INT_RX_ERR		(1 <<  1)
> >>> +#define RTL818X_INT_TXL_OK		(1 <<  2)
> >>> +#define RTL818X_INT_TXL_ERR		(1 <<  3)
> >>> +#define RTL818X_INT_RX_DU		(1 <<  4)
> >>> +#define RTL818X_INT_RX_FO		(1 <<  5)
> >>> +#define RTL818X_INT_TXN_OK		(1 <<  6)
> >>> +#define RTL818X_INT_TXN_ERR		(1 <<  7)
> >>> +#define RTL818X_INT_TXH_OK		(1 <<  8)
> >>> +#define RTL818X_INT_TXH_ERR		(1 <<  9)
> >>> +#define RTL818X_INT_TXB_OK		(1 << 10)
> >>> +#define RTL818X_INT_TXB_ERR		(1 << 11)
> >>> +#define RTL818X_INT_ATIM		(1 << 12)
> >>> +#define RTL818X_INT_BEACON		(1 << 13)
> >>> +#define RTL818X_INT_TIME_OUT		(1 << 14)
> >>> +#define RTL818X_INT_TX_FO		(1 << 15)
> >>> +	__le32	TX_CONF;
> >>> +#define RTL818X_TX_CONF_LOOPBACK_MAC	(1 << 17)
> >>> +#define RTL818X_TX_CONF_NO_ICV		(1 << 19)
> >>> +#define RTL818X_TX_CONF_DISCW		(1 << 20)
> >>> +#define RTL818X_TX_CONF_R8180_ABCD	(2 << 25)
> >>> +#define RTL818X_TX_CONF_R8180_F		(3 << 25)
> >>> +#define RTL818X_TX_CONF_R8185_ABC	(4 << 25)
> >>> +#define RTL818X_TX_CONF_R8185_D		(5 << 25)
> >>> +#define RTL818X_TX_CONF_HWVER_MASK	(7 << 25)
> >>> +#define RTL818X_TX_CONF_CW_MIN		(1 << 31)
> >>> +	__le32	RX_CONF;
> >>> +#define RTL818X_RX_CONF_MONITOR		(1 <<  0)
> >>> +#define RTL818X_RX_CONF_NICMAC		(1 <<  1)
> >>> +#define RTL818X_RX_CONF_MULTICAST	(1 <<  2)
> >>> +#define RTL818X_RX_CONF_BROADCAST	(1 <<  3)
> >>> +#define RTL818X_RX_CONF_DATA		(1 << 18)
> >>> +#define RTL818X_RX_CONF_CTRL		(1 << 19)
> >>> +#define RTL818X_RX_CONF_MGMT		(1 << 20)
> >>> +#define RTL818X_RX_CONF_BSSID		(1 << 23)
> >>> +#define RTL818X_RX_CONF_RX_AUTORESETPHY	(1 << 28)
> >>> +#define RTL818X_RX_CONF_ONLYERLPKT	(1 << 31)
> >>> +	__le32	INT_TIMEOUT;
> >>> +	__le32	TBDA;
> >>> +	u8	EEPROM_CMD;
> >>> +#define RTL818X_EEPROM_CMD_READ		(1 << 0)
> >>> +#define RTL818X_EEPROM_CMD_WRITE	(1 << 1)
> >>> +#define RTL818X_EEPROM_CMD_CK		(1 << 2)
> >>> +#define RTL818X_EEPROM_CMD_CS		(1 << 3)
> >>> +#define RTL818X_EEPROM_CMD_NORMAL	(0 << 6)
> >>> +#define RTL818X_EEPROM_CMD_LOAD		(1 << 6)
> >>> +#define RTL818X_EEPROM_CMD_PROGRAM	(2 << 6)
> >>> +#define RTL818X_EEPROM_CMD_CONFIG	(3 << 6)
> >>> +	u8	CONFIG0;
> >>> +	u8	CONFIG1;
> >>> +	u8	CONFIG2;
> >>> +	__le32	ANAPARAM;
> >>> +	u8	MSR;
> >>> +#define RTL818X_MSR_NO_LINK		(0 << 2)
> >>> +#define RTL818X_MSR_ADHOC		(1 << 2)
> >>> +#define RTL818X_MSR_INFRA		(2 << 2)
> >>> +	u8	CONFIG3;
> >>> +#define RTL818X_CONFIG3_ANAPARAM_WRITE	(1 << 6)
> >>> +	u8	CONFIG4;
> >>> +#define RTL818X_CONFIG4_POWEROFF	(1 << 6)
> >>> +#define RTL818X_CONFIG4_VCOOFF		(1 << 7)
> >>> +	u8	TESTR;
> >>> +	u8	reserved_9[2];
> >>> +	__le16	PGSELECT;
> >>> +	__le32	ANAPARAM2;
> >>> +	u8	reserved_10[12];
> >>> +	__le16	BEACON_INTERVAL;
> >>> +	__le16	ATIM_WND;
> >>> +	__le16	BEACON_INTERVAL_TIME;
> >>> +	__le16	ATIMTR_INTERVAL;
> >>> +	u8	reserved_11[4];
> >>> +	u8	PHY[4];
> >>> +	__le16	RFPinsOutput;
> >>> +	__le16	RFPinsEnable;
> >>> +	__le16	RFPinsSelect;
> >>> +	__le16	RFPinsInput;
> >>> +	__le32	RF_PARA;
> >>> +	__le32	RF_TIMING;
> >>> +	u8	GP_ENABLE;
> >>> +	u8	GPIO;
> >>> +	u8	reserved_12[10];
> >>> +	u8	TX_AGC_CTL;
> >>> +#define RTL818X_TX_AGC_CTL_PERPACKET_GAIN_SHIFT		(1 << 0)
> >>> +#define RTL818X_TX_AGC_CTL_PERPACKET_ANTSEL_SHIFT	(1 << 1)
> >>> +#define RTL818X_TX_AGC_CTL_FEEDBACK_ANT			(1 << 2)
> >>> +	u8	TX_GAIN_CCK;
> >>> +	u8	TX_GAIN_OFDM;
> >>> +	u8	TX_ANTENNA;
> >>> +	u8	reserved_13[16];
> >>> +	u8	WPA_CONF;
> >>> +	u8	reserved_14[3];
> >>> +	u8	SIFS;
> >>> +	u8	DIFS;
> >>> +	u8	SLOT;
> >>> +	u8	reserved_15[5];
> >>> +	u8	CW_CONF;
> >>> +#define RTL818X_CW_CONF_PERPACKET_CW_SHIFT	(1 << 0)
> >>> +#define RTL818X_CW_CONF_PERPACKET_RETRY_SHIFT	(1 << 1)
> >>> +	u8	CW_VAL;
> >>> +	u8	RATE_FALLBACK;
> >>> +	u8	reserved_16[25];
> >>> +	u8	CONFIG5;
> >>> +	u8	TX_DMA_POLLING;
> >>> +	u8	reserved_17[2];
> >>> +	__le16	CWR;
> >>> +	u8	RETRY_CTR;
> >>> +	u8	reserved_18[5];
> >>> +	__le32	RDSAR;
> >>> +	u8	reserved_19[18];
> >>> +	u16	TALLY_CNT;
> >>> +	u8	TALLY_SEL;
> >>> +} __attribute__((packed));
> >>
> >> enums have more visibility to the compiler and debugging tools
> >
> > enums don't have the same kind of typechecking this has.
>
> You're right.  They actually HAVE typechecking, whereas cpp #defines do
> not.
>
> We like typechecking in Linux, but your #defines have none.
>
Don't most drivers use #defines?

At any rate, all the registers are fully typechecked. The #defines sprinkled 
through there are bits for individual registers.. you want those converted to 
enums? (if so.. how does that help?)

-Michael Wu

Content of type "application/pgp-signature" skipped

Powered by blists - more mailing lists