On Tue, Sep 19, 2023 at 02:38:22PM +0200, Yann Sionneau wrote: > Hi, > > On 9/19/23 12:19, Wolfram Sang wrote: > > > I also agree that a wmb() in the i2c driver is not the more elegant fix. > > > For similar reasons, we hid barriers in the write*() macros, drivers > > > need to stay architecture-agnostic as much as possible. > > Exactly my thinking. I wanted to read this patch discussion later this > > week. But from glimpsing at it so far, I already wondered why there > > isn't a memory barrier in the final accessor to the register. > > The regmap accessors used by the designware driver end up calling > writel_relaxed() and readl_relaxed() : https://elixir.bootlin.com/linux/v6.6-rc2/source/drivers/i2c/busses/i2c-designware-common.c#L71 OK, since it ends up with the *_relaxed() accessors, there are no barriers here. I wonder whether the regmap API should have both standard and relaxed variants. If a regmap driver does not populate the .reg_write_relaxed etc. members, a regmap_write_relaxed() would just fall back to regmap_write(). We went through similar discussions many years ago around the I/O accessors and decided to add the barriers to readl/writel() even if they become more expensive, correctness should be first. The relaxed variants were added as optimisations if specific memory ordering was not required. I think the regmap API should follow the same semantics, go for correctness first as you can't tell what the side-effect of a regmap_write() is (e.g. kicking off DMA or causing an interrupt on another CPU). > In those cases I would say the smp_* barriers are what we are supposed to > use, isn't it? While smp_* is ok, it really depends on what the regmap_write() does. Is it a write to a shared peripheral (if not, you may need a DSB)? Does the regmap_write() caller know this? That's why I think having the barrier in dw_reg_write() is better. If you do want to stick to a fix in i2c_dw_xfer_init(), you could go for dma_wmb(). While this is not strictly DMA, it's sharing data with another coherent agent (a different CPU in this instance). The smp_wmb() is more about communication via memory not involving I/O. But this still assumes that the caller knows regmap_write() ends up with an I/O write*() (potentially relaxed). -- Catalin
