On Mon, Feb 20, 2023 at 01:04:02PM +0100, Morten Brørup wrote: > > From: Bruce Richardson [mailto:bruce.richard...@intel.com] > > Sent: Monday, 20 February 2023 12.53 > > > > On Mon, Feb 20, 2023 at 12:12:50PM +0100, Morten Brørup wrote: > > > > From: Thomas Monjalon [mailto:tho...@monjalon.net] > > > > Sent: Monday, 20 February 2023 11.17 > > > > > > > > 20/02/2023 10:43, Morten Brørup: > > > > > > From: Thomas Monjalon [mailto:tho...@monjalon.net] > > > > > > Sent: Monday, 20 February 2023 08.45 > > > > > > > > > > > > 16/02/2023 09:36, Ruifeng Wang: > > > > > > > From: Chengwen Feng <fengcheng...@huawei.com> > > > > > > > > Subject: [PATCH] net/hns3: support disable IOVA as PA mode > > > > > > > > > > > > Could we change the title to "support IOVA as VA" ? > > > > > > > > > > The underlying problem is the meson configuration option name for > > > > this feature [1]: > > > > > > > > > > option('enable_iova_as_pa', type: 'boolean', value: true, > > > > description: > > > > > 'Support for IOVA as physical address. Disabling removes > > the > > > > buf_iova field of mbuf.') > > > > > > > > > > [1]: > > > > > > https://elixir.bootlin.com/dpdk/v22.11.1/source/meson_options.txt#L43 > > > > > > > > > > Formally, the patch provides the ability to set a boolean > > > > configuration value ("enable_iova_as_pa") to false, and thus the > > patch > > > > title is correct. > > > > > > > > > > Nonetheless, I agree that the title suggested by Thomas is an > > > > improvement. > > > > > > > > > > > > > > > Going back to the root cause, I think the configuration option > > should > > > > be an enum instead of a boolean, e.g. "iova_mode" with values > > "iova_pa" > > > > and "iova_va". > > > > > > > > We can enable both and have it decided at runtime. So I think the > > > > boolean is OK. > > > > > > I forgot that it could be changed at runtime. > > > > > > I'll share a few thoughts for consideration, but expect no further > > replies. Sorry about the noise. ;-) > > > > > > The documentation [2] says that IOVA as PA is always supported, and > > is the default mode. Support for IOVA as VA is optional. > > > > > > [2]: > > https://www.intel.com/content/www/us/en/developer/articles/technical/me > > mory-in-dpdk-part-2-deep-dive-into-iova.html > > > > > > IOVA as VA can be selected at runtime, as you mention, or at build > > time. But selecting IOVA as VA (at runtime or build time) requires > > support by the underlying environment/hardware. > > > > > > If IOVA as PA is always supported (and is the default), the name of > > this meson configuration option could be improved. Its current name > > says "enable feature X", but if feature X is already supported by > > default, the name seems meaningless. If we want to keep it boolean, it > > could be inverted, e.g.: "iova_as_va_only" with default value "false". > > > > > > However, if modifying the meson configuration option (name and/or > > type) doesn't reduce the risk of confusion with the various IOVA modes, > > it's not worth the effort. > > > > > I agree that this option is confusing, and thinking about it, I agree > > that > > a pair of named option is probably better than just a true/false > > booleans. > > My current thinking is that a combo option is best - maybe named: > > "supported_iova_modes", with possible values ["va_and_pa", "va_only"] > > may > > be clearest. However, that would be a change in how things are > > currently > > configured. > > > > A alternative if we want to keep compatibility, is to expand or clarify > > the > > help text for the existing "enable_iova_as_pa" option. The current help > > text reads: > > > > "Support for IOVA as physical address. Disabling removes the buf_iova > > field > > of mbuf." > > > > We could expand that to e.g.: > > > > "Support the use of physical addresses for IO addresses, such as used > > by > > VFIO in no-iommu mode, or UIO-based drivers. When disabled, DPDK can > > only > > run with IOMMU support for address mappings, but will have more space > > available in the mbuf structure". > > > > Such an explanation is quite a bit longer, but I see meson does a > > decent > > job of wrapping the output of "meson configure" in latest versions. > > > > /Bruce > > Updating the description of the meson configuration option would be an > improvement. > > But I'm thinking more about the ripple effect into the resulting #define's, > and the code using those. It would be nice getting this cleaned up. Which is > why I compare the IOVA mode to CPU Endianness, as an example of a Boolean > value represented by multiple #define's for code readability purposes. > > But I suppose such a change has too widely reaching side effects, regarding > backwards compatibility.
Actually, I think the current internal define is pretty ok right now. RTE_IOVA_AS_PA would probably be better as "RTE_SUPPORT_IOVA_AS_PA", but I don't think the lack of the alternative value "RTE_SUPPORT_IOVA_AS_VA" is an issue since a DPDK build always supports that - it's only at runtime it may not be supported e.g. if no IOMMU is present. /Bruce
