RE: [PATCH v6 03/24] rtw89: add core and trx files

[Pkshih <pkshih@realtek.com>]

> -----Original Message-----
> From: kvalo=codeaurora.org@mg.codeaurora.org <kvalo=codeaurora.org@mg.codeaurora.org> On
> Behalf Of Kalle Valo
> Sent: Saturday, October 2, 2021 12:26 AM
> To: Pkshih <pkshih@realtek.com>
> Cc: linux-wireless@vger.kernel.org
> Subject: Re: [PATCH v6 03/24] rtw89: add core and trx files
> 
> Ping-Ke Shih <pkshih@realtek.com> writes:
> 
> > Implement main flows that contains register/unregister mac80211 hw with
> > hardware capability, power on/off sequence, STA state actions, and
> > TX/RX path.
> >
> > The chip info is read from efuse while probing PCI, and then it can be
> > used to induce supported channel, band, bitrate, ht/vht/he capability,
> > and etc. Then, we register hardware with these capabilities.
> >
> > When network interface is up, driver does power-on sequence to enable MAC,
> > BB and RF function blocks. Oppositely, do power-off sequence when
> > interface is going to down.
> >
> > To maintain STA state, five callbacks are implemented -- add, assoc,
> > disassoc, disconnect and remove. In which state, driver tells firmware STA
> > info via H2C.
> >
> > TX flow:
> > When a SKB is going to be transmitted, we must know its type first. If
> > the type is mgmt or fwcmd made by driver, SKB is queued into corresponding
> > DMA channel and PCI ring. The other type is data frame that is more
> > complex, because it needs to establish BA session to have better throughput
> > with AMPDU and AMSDU.
> > In order to have better PCI DMA efficiency, we don't kick off DMA every
> > SKB. With wake TX queue, kick off DMA after a bunch of SKBs are written.
> > To achieve this, we have two HCI ops -- tx_write and tx_kick_off.
> >
> > BA establishment work:
> > For data frames, we start to establish BA session if the STA is associated
> > with APMDU capability and the TID session isn't established, and then the
> > BA work is used to ask mac80211 to start AMPDU actions. Driver implements
> > AMPDU action callbacks to know the session is established, so that we can
> > set AGG_EN bit in TX descriptor to enable AMPDU.
> >
> > RX flow:
> > When a RX SKB is delivered from PCI, rtw89_core_rx() process it depneds on
> > its type -- WIFI, C2H or PPDU. If type is C2H, it's queued into a C2H
> > queue, and wake a work to handle the C2H packet. If type is WIFI, it's a
> > normal RX packet. When mgmt or data frame is received, it is queued
> > into pending RX SKB queue to wait for corresponding PPDU packet (another
> > RX packet with PPDU type) to fill its rx_status, like RSSI. And, then
> > indicate this packet to mac80211. When control frame is received, indicate
> > it to mac80211 immediately.
> >
> > Track work:
> > Use track work to monitor PHY status to know the changes of environment,
> > and then update RA status or do RFK accordingly.
> >
> > Signed-off-by: Ping-Ke Shih <pkshih@realtek.com>
> 
> [...]
> 
> > +static __always_inline void RTW89_SET_TXWD(u8 *txdesc, u32 val, u8 offset, u32 mask)
> > +{
> > +	u32 *txd32 = (u32 *)txdesc;
> > +
> > +	le32p_replace_bits((__le32 *)(txd32 + offset), val, mask);
> > +}
> 
> I'm not convinced about this either, please just use inline.

This is because 'mask' argument of le32p_replace_bits() must be constant
only. If I use inline and build this driver with ccflags-y += -Os,
compiler reports errors:

In function 'field_multiplier',
    inlined from 'le32_encode_bits' at ./include/linux/bitfield.h:154:1,
    inlined from 'le32p_replace_bits' at ./include/linux/bitfield.h:154:1,
    inlined from 'RTW89_SET_FWCMD_UA32.constprop' at /work/git-root/rtwlan/rtw89/fw.h:1397:2:
./include/linux/bitfield.h:119:3: error: call to '__bad_mask' declared with attribute error: bad bitfield mask
  119 |   __bad_mask();
      |   ^~~~~~~~~~~~

I check the implement of le32p_replace_bits(), it looks like

static __always_inline void type##p_replace_bits(__##type *p,           \
                                        base val, base field)           \
{                                                                       \
        *p = (*p & ~to(field)) | type##_encode_bits(val, field);        \
}

So, I imitate the function to use __always_inline, and then it works.

Do you think I don't need to consider the case of Os?
But, -Os seems a standard option of Linux kernel.

ifdef CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE
KBUILD_CFLAGS += -O2
else ifdef CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE_O3
KBUILD_CFLAGS += -O3
else ifdef CONFIG_CC_OPTIMIZE_FOR_SIZE
KBUILD_CFLAGS += -Os
endif

--
Ping-Ke