From: Andy Shevchenko > Sent: 23 February 2018 16:51 > On Fri, Feb 23, 2018 at 6:41 PM, David Laight <david.lai...@aculab.com> wrote: > > From: Arnd Bergmann > >> Sent: 23 February 2018 15:37 > >> > >> 32-bit architectures generally cannot use writeq(), so we now get a build > >> failure for the lpfc driver: > >> > >> drivers/scsi/lpfc/lpfc_sli.c: In function 'lpfc_sli4_wq_put': > >> drivers/scsi/lpfc/lpfc_sli.c:145:4: error: implicit declaration of > >> function 'writeq'; did you mean > >> 'writeb'? [-Werror=implicit-function-declaration] > >> > >> Another problem here is that writing out actual data (unlike accessing > >> mmio registers) means we must write the data with the same endianess > >> that we have read from memory, but writeq() will perform byte swaps > >> and add barriers inbetween accesses as we do for registers. > >> > >> Using memcpy_toio() should do the right thing here, using register > >> sized stores with correct endianess conversion and barriers (i.e. none), > >> but on some architectures might fall back to byte-size access. > > ... > > > > Have you looked at the performance impact of this on x86? > > Last time I looked memcpy_toio() aliased directly to memcpy(). > > memcpy() is run-time patched between several different algorithms. > > On recent Intel cpus memcpy() is implemented as 'rep movsb' relying > > on the hardware to DTRT. > > For uncached accesses (typical for io) the 'RT' has to be byte transfers. > > So instead of the 8 byte transfers (on 64 bit) you get single bytes. > > This won't be what is intended! > > memcpy_toio() should probably use 'rep movsd' for the bulk of the transfer. > > Maybe I'm wrong but it uses movsq on 64-bit and movsl on 32-bit.
(Let's not argue about the instruction mnemonic). You might expect that, but last time I looked at the bus cycles on a PCIe slave that wasn't what I saw. > The side-effect I referred previously is about tails, i.e. unaligned > bytes are transferred in portions > like > 7 on 64-bit will be 4 + 2 + 1, > 5 = 4 + 1 on 64bit memcpy() is allowed to do: (long *)(tgt+len)[-1] = (long *)(src+len)[-1]; rep_movsq(tgt, src, len >> 3); provided the length is at least 8. The misaligned PCIe transfer generates a single TLP covering 12 bytes with the relevant byte enables set for the first and last 32bit words. David