[lkml]   [2006]   [Sep]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subjectquestion regarding cacheline size

This is for PCMCIA (cardbus) version of Silicon Image 3124 SerialATA
controller. When cacheline size is configured, the controller uses Read
Multiple commands.

• Bit [07:00]: Cache Line Size (R/W). This bit field is used to specify
the system cacheline size in terms of 32-bit words. The SiI3124, when
initiating a read transaction, will issue the Read Multiple PCI command
if empty space in its FIFO is greater than the value programmed in this

As the BIOS doesn't run after hotplugging cardbus card, the cache line
isn't configured and the controller ends up having 0 cache line size and
always using Read command. When that happens, write performance drops
hard - the throughput is < 2Mbytes/s.

So, sata_sil24 driver has to program CLS if it's not already set, but
I'm not sure which number to punch in. FWIW, sil3124 doesn't seem to
put restrictions on the values which can be used for CLS. There are
several candidates...

* L1_CACHE_BYTES / 4 : this is used by init routine in yenta_socket.c.
It seems to be a sane default but I'm not sure whether L1 cache line
size always coincides with the size as seen from PCI bus.

* pci_cache_line_size in drivers/pci/pci.c : this is used for
pci_generic_prep_mwi() and can be overridden by arch specific code.
this seems more appropriate but is not exported.

For all involved commands - memory read line, memory read multiple and
memory write and invalidate - a value larger than actual cacheline size
doesn't hurt but a smaller value may.

I'm thinking of implementing a query function for pci_cache_line_size,
say, int pci_cacheline_size(struct pci_dev *pdev), and use it in the
device init routine. Does this sound sane?


To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2006-09-07 10:33    [W:0.059 / U:1.448 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site