(The inter-version diffs are messed up due to the zstd version bump that happened during this review, unfortunately, but the full patch isn't too big.)

The current patch looks good to me, modulo all of the M_NOWAIT. I think allocations should generally be M_WAITOK unless we are truly in a sleep-free context; and I think it's unlikely compression should be done from sleep-free contexts anyway.

It also occurs to me that moving this to the boot loader will mean we'll likely need to have the same sed'ing I suggested for the kernel, so there's opportunity for reuse of that. I'd be happy to help with that since that's clearly the next stop for these patches.

I am on board with the idea of moving contrib/zstd to sys/contrib/zstd. (I think the kernel build already reaches outside of sys/ for some headers, if not .c files, though.) Maybe bde would complain about the abomination.

I would do that as a separate patch so that the SVN moves are tracked correctly.

In D10407#271258, @cem wrote:

I am on board with the idea of moving contrib/zstd to sys/contrib/zstd. (I think the kernel build already reaches outside of sys/ for some headers, if not .c files, though.) Maybe bde would complain about the abomination.

As far as I know, the only thing outside sys that the kernel build uses is header files from the compiler (not the userland source tree) for intrinsic intel aes instructions, but nothing else. And that's a compiler issue, not grabbing something form userland. If I've missed something, I'd be on board with fixing that asap. OK, there is a 'gdbinit' target that uses debug scripts from src/tools/debugscripts/gdbinit.* to generate a .gdbinit when you do an make install.debug...

bde would complain, but he's likely not the only one...

It is true that sys/boot reaches outside extensively (even before my hacking on it), but that's not the kernel and likely will be moving to src/stand or similar soon anyway...

I would do that as a separate patch so that the SVN moves are tracked correctly.

Agreed.

A more complete grep shows the kernel will also use ${S}/../COPYRIGHT if it exists, but won't if it doesn't. And that's only for a copyright notice inside a comment for vers.c. Otherwise, all other relative path things in the kernel and module builds land inside the kernel.
There is one oddball. In libtekken which lives in sys/tekken/ there is a reference libc, but it's a library that doesn't look to be built at all and looks to be userland code.

In D10407#271305, @imp wrote:

As far as I know, the only thing outside sys that the kernel build uses is header files from the compiler (not the userland source tree) for intrinsic intel aes instructions, but nothing else.

Yeah, that one was all I had in mind.

Thanks for the feedback Warner. I'll work on that stuff later today or tomorrow.

FYI, I turned on ZSTD assertions in DEBUG kernels with this patch:

diff --git a/contrib/zstd/lib/common/bitstream.h b/contrib/zstd/lib/common/bitstream.h
index e2de370..052dc6c 100644
--- a/contrib/zstd/lib/common/bitstream.h
+++ b/contrib/zstd/lib/common/bitstream.h
@@ -53,17 +53,21 @@ extern "C" {


 /*-*************************************
 *  Debug
 ***************************************/
+#ifdef _KERNEL
+#  define assert(condition)    KASSERT(condition, (#condition))
+#else
 #  if defined(BIT_DEBUG) && (BIT_DEBUG>=1)
 #    include <assert.h>
 #  else
 #    ifndef assert
 #      define assert(condition) ((void)0)
 #    endif
 #  endif
+#endif


 /*=========================================
 *  Target specific
 =========================================*/
diff --git a/contrib/zstd/lib/common/zstd_internal.h b/contrib/zstd/lib/common/zstd_internal.h
index 1431d68..36c8d82 100644
--- a/contrib/zstd/lib/common/zstd_internal.h
+++ b/contrib/zstd/lib/common/zstd_internal.h
@@ -42,17 +42,21 @@ extern "C" {


 /*-*************************************
 *  Debug
 ***************************************/
+#ifdef _KERNEL
+#  define assert(condition)    KASSERT(condition, (#condition))
+#else
 #  if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=1)
 #    include <assert.h>
 #  else
 #    ifndef assert
 #      define assert(condition) ((void)0)
 #    endif
 #  endif
+#endif

 #define ZSTD_STATIC_ASSERT(c) { enum { ZSTD_static_assert = 1/(int)(!!(c)) }; }

 #if defined(ZSTD_DEBUG) && (ZSTD_DEBUG>=2)
 #  include <stdio.h>
diff --git a/contrib/zstd/lib/compress/zstd_compress.c b/contrib/zstd/lib/compress/zstd_compress.c
index 3cf625a..89fcbcb 100644
--- a/contrib/zstd/lib/compress/zstd_compress.c
+++ b/contrib/zstd/lib/compress/zstd_compress.c
@@ -1338,11 +1338,11 @@ MEM_STATIC size_t ZSTD_buildCTable(void* dst, size_t dstCapacity,
             if (FSE_isError(NCountSize)) return NCountSize;
             CHECK_F(FSE_buildCTable_wksp(CTable, norm, max, tableLog, workspace, workspaceSize));
             return NCountSize;
         }
     }
-    default: return assert(0), ERROR(GENERIC);
+    default: assert(0); return ERROR(GENERIC);
     }
 }

 MEM_STATIC size_t ZSTD_encodeSequences(void* dst, size_t dstCapacity,
     FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,

(Whitespace changes due to indent elided.)

I think it might be even easier to do this right.

If you create stddef.h, stdlib.h, string.h, and stdio.h that just includes zstd_kfreebsd.h somewhere in sys/contrib/zstd, then you could add -I for that directory and ditch all the #ifdef_KERNEL changes. No fancy 'sed' is needed like I thought before. stand/libsa will be migrating to this way as well to facilitate the lua import in the coming weeks.

Please go ahead and commit this, after fixing the easy issues brought up by Warner. :-)

Please commit this :-). If you don't have time to make Warner's changes and commit, I'm happy to assist.