Add a Linux user space backend for the CAAM port

configure.ac

@@ -4717,6 +4717,14 @@ then
             AM_CFLAGS="$AM_CFLAGS -DWOLFSSL_IMX6UL_CAAM"
             ;;
 
+        linux)
+            # Drive a job ring from Linux user space. Needs a reserved
+            # physical DMA pool (boot with mem=) and the in-tree caam driver
+            # unbound; see wolfcrypt/src/port/caam/README.md.
+            AM_CFLAGS="$AM_CFLAGS -DWOLFSSL_CAAM_LINUX -DWOLFSSL_CAAM_NO_SM"
+            ENABLED_CAAM_LINUX="yes"
+            ;;
+
         seco)
             SECO_DIR=$trylibsecodir
             AM_CPPFLAGS="$AM_CPPFLAGS -I$SECO_DIR/include"
@@ -13417,6 +13425,7 @@ AM_CONDITIONAL([BUILD_RC2],[test "x$ENABLED_RC2" = "xyes"])
 AM_CONDITIONAL([BUILD_CUDA],[test "x$ENABLED_CUDA" = "xyes"])
 AM_CONDITIONAL([BUILD_CAAM],[test "x$ENABLED_CAAM" != "xno"])
 AM_CONDITIONAL([BUILD_QNXCAAM],[test "x$ENABLED_CAAM_QNX" = "xyes"])
+AM_CONDITIONAL([BUILD_CAAM_LINUX],[test "x$ENABLED_CAAM_LINUX" = "xyes"])
 AM_CONDITIONAL([BUILD_IOTSAFE],[test "x$ENABLED_IOTSAFE" = "xyes"])
 AM_CONDITIONAL([BUILD_IOTSAFE_HWRNG],[test "x$ENABLED_IOTSAFE_HWRNG" = "xyes"])
 AM_CONDITIONAL([BUILD_VAULTIC],[test "x$ENABLED_VAULTIC" = "xyes"])

wolfcrypt/src/include.am

@@ -99,6 +99,8 @@ EXTRA_DIST += wolfcrypt/src/port/ti/ti-aes.c \
               wolfcrypt/src/port/caam/caam_driver.c \
               wolfcrypt/src/port/caam/caam_error.c \
               wolfcrypt/src/port/caam/caam_qnx.c \
+              wolfcrypt/src/port/caam/caam_linux.c \
+              wolfcrypt/src/port/caam/wolfcaam_linux.c \
               wolfcrypt/src/port/caam/caam_integrity.c \
               wolfcrypt/src/port/caam/caam_sha.c \
               wolfcrypt/src/port/caam/caam_doc.pdf \
@@ -243,6 +245,16 @@ endif
 EXTRA_DIST += wolfcrypt/src/port/sealsq/README.md
 
 
+if BUILD_CAAM_LINUX
+# The driver core is built out of tree for QNX (a resource manager) and into
+# the kernel for INTEGRITY. On Linux it runs in process, so it is built into
+# the library along with its port layer and request dispatch.
+src_libwolfssl@LIBSUFFIX@_la_SOURCES += wolfcrypt/src/port/caam/caam_driver.c
+src_libwolfssl@LIBSUFFIX@_la_SOURCES += wolfcrypt/src/port/caam/caam_error.c
+src_libwolfssl@LIBSUFFIX@_la_SOURCES += wolfcrypt/src/port/caam/caam_linux.c
+src_libwolfssl@LIBSUFFIX@_la_SOURCES += wolfcrypt/src/port/caam/wolfcaam_linux.c
+endif
+
 if BUILD_CAAM
 src_libwolfssl@LIBSUFFIX@_la_SOURCES += wolfcrypt/src/port/caam/wolfcaam_init.c
 src_libwolfssl@LIBSUFFIX@_la_SOURCES += wolfcrypt/src/port/caam/wolfcaam_qnx.c

wolfcrypt/src/port/caam/README.md

@@ -1 +1,90 @@
-See caam_doc.pdf for documentation about building and using.
+# wolfSSL CAAM Port
+
+See `caam_doc.pdf` for documentation about building and using the driver on
+i.MX under INTEGRITY, and `IDE/QNX/README.md` for the QNX build.
+
+## Linux user space (`--enable-caam=linux`)
+
+Runs the same driver core (`caam_driver.c`) from a Linux user space process,
+developed against the SEC on a QorIQ T1040. QNX reaches the hardware through a
+resource manager; here the driver runs in the calling process and a request is
+a direct call.
+
+    ./configure --host=powerpc-linux-gnu CC=powerpc-linux-gnu-gcc \
+                --enable-caam=linux
+    make
+
+Accelerated: **AES-CBC, AES-CTR, AES-ECB** and the **TRNG**. Everything else
+returns `CRYPTOCB_UNAVAILABLE` and runs in software, so an unsupported
+operation is a performance question, not a correctness one. Hashing has no
+descriptors in this driver (as on QNX), there is no i.MX style secure memory
+block on this part so blobs and black keys are out, and public key is not
+dispatched yet even though the driver carries the descriptors.
+
+### The engine has to be yours
+
+The in-tree `caam` driver claims all four job rings at boot and, on a part with
+more than 4 GB of DDR, sets `MCFGR[PS]` for 64-bit descriptor pointers.
+`MCFGR` is global to the block, so sharing it is not an option: this driver
+writes 32-bit pointer words. Take the whole engine, then put it back in 32-bit
+mode:
+
+    for j in ffe301000.jr ffe302000.jr ffe303000.jr ffe304000.jr; do
+        echo $j > /sys/bus/platform/drivers/caam_jr/unbind; done
+    echo ffe300000.crypto > /sys/bus/platform/drivers/caam/unbind
+
+Unbinding leaves that driver's interrupt handler registered, and the first job
+to complete would otherwise raise an IRQ it services against freed state and
+panic the kernel. `CAAM_SET_JOBRING_ADDR()` masks the ring interrupt when it
+claims the ring; the driver polls anyway.
+
+### Reserved DMA memory
+
+The engine cannot use ordinary user pages: they are not physically contiguous,
+and on a 36-bit part they sit above what a 32-bit descriptor pointer can
+address (measured on a T1040: virtual `0x100b4f10` -> physical
+`0x1_EF9C2F10`). Boot Linux with `mem=` so it stops managing the top of DDR,
+and the port carves engine buffers out of that reserved range instead:
+
+    setenv othbootargs 'ramdisk_size=1000000 mem=2048M'
+
+That leaves physical `[2 GB, 4 GB)` unmanaged, contiguous, and below the
+32-bit limit. It also makes `CONFIG_STRICT_DEVMEM` allow the mapping, since
+the range is no longer reported as System RAM.
+
+Note CCSR is at the full physical address under Linux (`0xF_FE000000` on a
+T1040), not the 32-bit view bare metal sees.
+
+| Macro | Default | Meaning |
+|---|---|---|
+| `CAAM_LINUX_CCSR_PHYS` | `0xFFE000000ULL` | Physical base of the CCSR window |
+| `CAAM_LINUX_SEC_OFFSET` | `0x300000` | SEC block within CCSR |
+| `CAAM_LINUX_JR_OFFSET` | `0x1000` | Which job ring to claim |
+| `CAAM_LINUX_POOL_PHYS` | `0x80000000ULL` | Base of the reserved DMA pool |
+| `CAAM_LINUX_POOL_SZ` | 256 KB | Size of that pool |
+| `CAAM_LINUX_AES_MAX` | 16 KB | Largest AES request taken; bigger goes to software |
+
+### Performance: read this before enabling it
+
+On a T1040 at 1.4 GHz the engine is **slower than software AES**, measured
+with the same binary over the same buffers:
+
+| Buffer | Software | CAAM | Ratio |
+|---:|---:|---:|---:|
+| 512 B | 42.69 MiB/s | 21.61 MiB/s | 0.51x |
+| 1 KB | 44.61 MiB/s | 24.88 MiB/s | 0.56x |
+| 4 KB | 45.67 MiB/s | 27.09 MiB/s | 0.59x |
+| 8 KB | 46.30 MiB/s | 27.50 MiB/s | 0.59x |
+| 16 KB | 46.31 MiB/s | 27.75 MiB/s | 0.60x |
+
+Hardware throughput is flat at roughly 28 MiB/s because the per-job cost
+dominates: two `memcpy`s to stage operands in and results out of the pool, a
+descriptor build, and a polled completion. Software rises with buffer size and
+the e5500's AES is quick. The gap does narrow with key size, since the engine
+barely notices AES-256 while software slows down.
+
+So the reason to enable this is not throughput. It is access to the TRNG as a
+real entropy source, and having the crypto happen somewhere other than the
+core. Closing the gap means removing the bounce buffers (letting callers
+allocate from the pool directly), raising the transfer size toward the SEC's
+64 KB per-descriptor limit, and completing on an interrupt rather than a poll.

wolfcrypt/src/port/caam/caam_driver.c

@@ -24,7 +24,8 @@
 #endif
 
 #if (defined(__INTEGRITY) || defined(INTEGRITY)) || \
-    (defined(__QNX__) || defined(__QNXNTO__))
+    (defined(__QNX__) || defined(__QNXNTO__)) || \
+    defined(WOLFSSL_CAAM_LINUX)
 
 #if defined(__QNX__) || defined(__QNXNTO__)
     #include <sys/mman.h>
@@ -240,6 +241,7 @@ static Error caamFreePage(unsigned int page)
 }
 
 /* free the partition and dealloc */
+#ifndef WOLFSSL_CAAM_NO_SM
 Error caamFreePart(unsigned int part)
 {
     unsigned int status;
@@ -263,9 +265,11 @@ Error caamFreePart(unsigned int part)
     WOLFSSL_MSG("free'd partition");
     return Success;
 }
+#endif /* !WOLFSSL_CAAM_NO_SM */
 
 
 /* find all partitions we own and free them */
+#ifndef WOLFSSL_CAAM_NO_SM
 static Error caamFreeAllPart()
 {
     unsigned int SMPO;
@@ -281,11 +285,13 @@ static Error caamFreeAllPart()
 
     return 0;
 }
+#endif /* !WOLFSSL_CAAM_NO_SM */
 
 
 /* search through the partitions to find an unused one
  * returns negative value on failure, on success returns 0 or greater
  */
+#ifndef WOLFSSL_CAAM_NO_SM
 int caamFindUnusedPartition()
 {
     unsigned int SMPO;
@@ -302,6 +308,7 @@ int caamFindUnusedPartition()
 
     return ret;
 }
+#endif /* !WOLFSSL_CAAM_NO_SM */
 
 
 /* flag contains how the partition is set i.e CSP flag and read/write access
@@ -366,6 +373,7 @@ static Error caamCreatePartition(unsigned int* page, unsigned int par,
 
 
 /* return a partitions physical address on success, returns 0 on fail */
+#ifndef WOLFSSL_CAAM_NO_SM
 CAAM_ADDRESS caamGetPartition(unsigned int part, int partSz, unsigned int flag)
 {
     int err;
@@ -382,6 +390,7 @@ CAAM_ADDRESS caamGetPartition(unsigned int part, int partSz, unsigned int flag)
 
     return (CAAM_ADDRESS)(CAAM_PAGE + (part << 12));
 }
+#endif /* !WOLFSSL_CAAM_NO_SM */
 
 
 /* Gets the status of a job. Returns CAAM_WAITING if no output jobs ready to be
@@ -1977,7 +1986,9 @@ int InitCAAM(void)
 
 int CleanupCAAM()
 {
+#ifndef WOLFSSL_CAAM_NO_SM
     caamFreeAllPart();
+#endif
     CAAM_UNSET_JOBRING_ADDR(caam.ring.BaseAddr, caam.ring.JobIn,
         caam.ring.VirtualIn);
     CAAM_FREE_MUTEX(&caam.ring.jr_lock);

wolfcrypt/src/port/caam/caam_error.c

@@ -24,7 +24,8 @@
 #endif
 
 #if (defined(__INTEGRITY) || defined(INTEGRITY)) || \
-    (defined(__QNX__) || defined(__QNXNTO__))
+    (defined(__QNX__) || defined(__QNXNTO__)) || \
+    defined(WOLFSSL_CAAM_LINUX)
 
 #include <wolfssl/wolfcrypt/port/caam/caam_driver.h>
 #include <wolfssl/wolfcrypt/port/caam/caam_error.h>