Mark all ccalls as GC safe (#2262)

This allows the GC to run while potentially blocking in a CUDA library.
To make this safe, callbacks into Julia should again transition to
GC-unsafe mode.

It should be reimplemented when JuliaLang/julia#49933 lands.

Co-authored-by: Tim Besard <tim.besard@gmail.com>

Author: vchuravy

.buildkite/pipeline.yml

@@ -132,7 +132,12 @@ steps:
             println("--- :julia: Instantiating project")
             withenv("JULIA_PKG_PRECOMPILE_AUTO" => 0) do
               Pkg.activate(joinpath(pwd(), "lib", lowercase("{{matrix.package}}")))
-              Pkg.instantiate()
+              try
+                Pkg.instantiate()
+              catch
+                # if we fail to instantiate, assume that we need a newer CUDA.jl
+                Pkg.develop(path=".")
+              end
 
               Pkg.add("CUDA_Runtime_jll")
               write("LocalPreferences.toml", "[CUDA_Runtime_jll]\nversion = \"{{matrix.cuda}}\"")

Project.toml

@@ -1,6 +1,6 @@
 name = "CUDA"
 uuid = "052768ef-5323-5732-b1bb-66c8b64840ba"
-version = "5.2.0"
+version = "5.3.0"
 
 [deps]
 AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"

lib/cublas/CUBLAS.jl

@@ -210,7 +210,7 @@ function log_message(ptr)
     return
 end
 
-function _log_message(blob)
+@gcunsafe_callback function _log_message(blob)
     # the message format isn't documented, but it looks like a message starts with a capital
     # and the severity (e.g. `I!`), and subsequent lines start with a lowercase mark (`!i`)
     #

lib/cublas/libcublas.jl

@@ -36,7 +36,9 @@ end
 # HACK: callback function for `launch_configuration` on platforms without support for
 #       trampolines as used by `@cfunction` (JuliaLang/julia#27174, JuliaLang/julia#32154)
 _shmem_cb = nothing
-_shmem_cint_cb(x::Cint) = Cint(something(_shmem_cb)(x))
+@gcunsafe_callback function _shmem_cint_cb(x::Cint)
+    Cint(something(_shmem_cb)(x))
+end
 _shmem_cb_lock = Threads.ReentrantLock()
 
 """
@@ -58,7 +60,9 @@ function launch_configuration(fun::CuFunction; shmem::Union{Integer,Base.Callabl
     if isa(shmem, Integer)
         cuOccupancyMaxPotentialBlockSize(blocks_ref, threads_ref, fun, C_NULL, shmem, max_threads)
     elseif Sys.ARCH == :x86 || Sys.ARCH == :x86_64
-        shmem_cint = threads -> Cint(shmem(threads))
+        @gcunsafe_callback function shmem_cint(threads)
+            Cint(shmem(threads))
+        end
         cb = @cfunction($shmem_cint, Cint, (Cint,))
         cuOccupancyMaxPotentialBlockSize(blocks_ref, threads_ref, fun, cb, 0, max_threads)
     else

lib/cudadrv/libcuda.jl

@@ -1,7 +1,7 @@
 name = "cuDNN"
 uuid = "02a925ec-e4fe-4b08-9a7e-0d78e3d38ccd"
 authors = ["Tim Besard <tim.besard@gmail.com>"]
-version = "1.3.0"
+version = "1.3.1"
 
 [deps]
 CEnum = "fa961155-64e5-5f13-b03f-caf6b980ea82"
@@ -11,7 +11,7 @@ CUDNN_jll = "62b44479-cb7b-5706-934f-f13b2eb2e645"
 
 [compat]
 CEnum = "0.2, 0.3, 0.4, 0.5"
-CUDA = "~5.1, ~5.2"
+CUDA = "~5.3"
 CUDA_Runtime_Discovery = "0.2"
 CUDNN_jll = "~8.9"
 julia = "1.8"

lib/cudadrv/occupancy.jl

@@ -137,7 +137,7 @@ function log_message(sev, udata, dbg_ptr, ptr)
     return
 end
 
-function _log_message(sev, dbg, str)
+@gcunsafe_callback function _log_message(sev, dbg, str)
     lines = split(str, '\0')
     msg = join(lines, '\n')
     if sev == CUDNN_SEV_INFO