In D19226#411354, @markj wrote:

It looks like this is assuming that the old value is not valid, but that's not always true. For example, pmap_promote_pde() simply overwrites the old PDE.

I don't really see how pmap_kextract() is safe wrt torn writes.

Alan' notes.

Why cannot the same happen for read side ?

More Mark' notes.

Editing.

Grammar.

In D18893#410913, @markj wrote:

The code changes look ok to me. Do you plan to upload man page changes here too?

In D18423#410851, @emaste wrote:

I believe this is ready to commit now (disabled by default); commit message:

Do static binaries work ? I suspect they do not, or they work by a chance. Most scary is init(8) of course, but other static binaries esp. rescue(8) are also questionable.

Provide implementations for all arches.

s/-/=/

Fix quite a bug in the loading of depended objects.
Restructure to avoid coding current thread offset calculation twice.

The buffer is not statically allocated, it is statically sized (in the revision summary).

Rename rangeset_get() to rangeset_lookup().
Remove one more _NOFREE.
Clear dst after failed copy.

In D19096#408759, @ngie wrote:

Thanks so much!

@emaste: MFC note: this breaks struct namei layout. Do we want the merge ?

Should this field be at the end of the struct to avoid breaking ABI?

In D19063#408640, @gallatin wrote:

The reason to have this in the driver is for performance. The plan for pfil is to eventually just use a pointer to a memory blob for filtering. The current fake mbuf stuff is a step in that direction.

By passing just a pointer from inside the driver, filtering and dropping floods of unwanted traffic can become incredibly cheap. For each packet examined and dropped, we avoid the cache misses inherent in allocating an mbuf and filling it in (at least 2 cachelines just for the mbuf), as well as the memory copy that most drivers do for small packets (likely 2 more cachelines). For larger packets, or drivers that do not employ the copy-for-small packets strategy, instead of the copy, we avoid the virtual to physical (or worse, IOMMU) translation for new rx buffers. Avoiding non-IO cache misses is especially beneficial on Intel Xeon w/DDIO, when recently received traffic may still be resident in the DDIO cache ways.

I have a generic question about the patch. Why this needs to be done inside the driver, and why cannot pfil hooks be applied e.g. in ether_input() ? It is strange that all drivers would need this patch.

In D19072#407798, @emaste wrote:

Should we encourage lld to revive the https://reviews.llvm.org/D33041 patch?