]> git.saurik.com Git - apple/xnu.git/blobdiff - osfmk/ppc/commpage/commpage.c
xnu-517.tar.gz
[apple/xnu.git] / osfmk / ppc / commpage / commpage.c
index 907ae701d00dd40b5e2453eaf84b77d3767e1c3b..e611402465cfef0d7e61de8bb638acc4587d2df2 100644 (file)
  */
 
 /*
- *  This is a simplifed version of the commpage support from 10.3.
- *  The supported feature is the tuning of _cpu_capabilities.
- *  There is no shared page for user processes.
+ *     Here's what to do if you want to add a new routine to the comm page:
+ *
+ *             1. Add a definition for it's address in osfmk/ppc/cpu_capabilities.h,
+ *                being careful to reserve room for future expansion.
+ *
+ *             2. Write one or more versions of the routine, each with it's own
+ *                commpage_descriptor.  The tricky part is getting the "special",
+ *                "musthave", and "canthave" fields right, so that exactly one
+ *                version of the routine is selected for every machine.
+ *                The source files should be in osfmk/ppc/commpage/.
+ *
+ *             3. Add a ptr to your new commpage_descriptor(s) in the "routines"
+ *                array in commpage_populate().  Of course, you'll also have to
+ *                declare them "extern" in commpage_populate().
+ *
+ *             4. Write the code in Libc to use the new routine.
  */
 
 #include <mach/mach_types.h>
 #include <ppc/machine_routines.h>
 #include <machine/cpu_capabilities.h>
 #include <machine/commpage.h>
+#include <machine/pmap.h>
+#include <vm/vm_kern.h>
+#include <mach/vm_map.h>
+
+static char    *next = NULL;                   // next available byte in comm page
+static int             cur_routine = 0;                // comm page address of "current" routine
+static int             matched;                                // true if we've found a match for "current" routine
 
 int            _cpu_capabilities = 0;                  // define the capability vector
 
+char   *commPagePtr = NULL;                    // virtual address of comm page in kernel map
+
+
+/* Allocate the commpages and add to the shared submap created by vm:
+ *     1. allocate pages in the kernel map (RW)
+ *     2. wire them down
+ *     3. make a memory entry out of them
+ *     4. map that entry into the shared comm region map (R-only)
+ */
+static void*
+commpage_allocate( void )
+{
+    extern     vm_map_t        com_region_map;                         // the shared submap, set up in vm init
+    vm_offset_t                        kernel_addr;                            // address of commpage in kernel map
+    vm_offset_t                        zero = 0;
+    vm_size_t                  size = _COMM_PAGE_AREA_USED;    // size actually populated
+    ipc_port_t                 handle;
+    
+    if (com_region_map == NULL)
+        panic("commpage map is null");
+    
+    if (vm_allocate(kernel_map,&kernel_addr,_COMM_PAGE_AREA_USED,VM_FLAGS_ANYWHERE))
+        panic("cannot allocate commpage");
+        
+    if (vm_map_wire(kernel_map,kernel_addr,kernel_addr+_COMM_PAGE_AREA_USED,VM_PROT_DEFAULT,FALSE))
+        panic("cannot wire commpage");
+    
+    if (mach_make_memory_entry(        kernel_map,                     // target map
+                                &size,                         // size
+                                kernel_addr,           // offset (address in kernel map)
+                                VM_PROT_DEFAULT,       // map it RW
+                                &handle,                       // this is the object handle we get
+                                NULL ))                                // parent_entry
+        panic("cannot make entry for commpage");
+    
+    if (vm_map_64(     com_region_map,                                 // target map (shared submap)
+                    &zero,                                                     // address (map into 1st page in submap)
+                    _COMM_PAGE_AREA_USED,                      // size
+                    0,                                                         // mask
+                    VM_FLAGS_FIXED,                                    // flags (it must be 1st page in submap)
+                    handle,                                                    // port is the memory entry we just made
+                    0,                                                         // offset (map 1st page in memory entry)
+                    FALSE,                                                     // copy
+                    VM_PROT_READ,                                      // cur_protection (R-only in user map)
+                    VM_PROT_READ,                                      // max_protection
+                    VM_INHERIT_SHARE ))                                // inheritance
+        panic("cannot map commpage");
+        
+    ipc_port_release(handle);
+        
+    return (void*) kernel_addr;                                                // return address in kernel map
+}
+
+
+/* Get address (in kernel map) of a commpage field. */
+
+static void*
+commpage_addr_of(
+    int        addr_at_runtime )
+{
+    return     (void*) (commPagePtr + addr_at_runtime - _COMM_PAGE_BASE_ADDRESS);
+}
+
+
 /* Determine number of CPUs on this system.  We cannot rely on
  * machine_info.max_cpus this early in the boot.
  */
@@ -86,6 +170,186 @@ commpage_init_cpu_capabilities( void )
 }
 
 
+/* Copy data into commpage. */
+
+ void
+commpage_stuff(
+    int        address,
+    void       *source,
+    int        length  )
+{    
+    char       *dest = commpage_addr_of(address);
+    
+    if (dest < next)
+        panic("commpage overlap: %08 - %08X", dest, next);
+    
+    bcopy((char*)source,dest,length);
+    
+    next = (dest + length);
+}
+
+
+/* Modify commpage code in-place for this specific platform. */
+
+static void
+commpage_change(
+    uint32_t   *ptr,
+    int                bytes,
+    uint32_t   search_mask, 
+    uint32_t   search_pattern,
+    uint32_t   new_mask,
+    uint32_t   new_pattern,
+    int                        (*check)(uint32_t instruction)  )
+{
+    int                        words = bytes >> 2;
+    uint32_t   word;
+    int                        found_one = 0;
+
+    while( (--words) >= 0 ) {
+        word = *ptr;
+        if ((word & search_mask)==search_pattern) {
+            if ((check==NULL) || (check(word))) {      // check instruction if necessary
+                found_one = 1;
+                word &= ~new_mask;
+                word |= new_pattern;
+                *ptr = word;
+            }
+        }
+        ptr++;
+    }
+    
+    if (!found_one)
+        panic("commpage opcode not found");
+}
+
+
+/* Check to see if exactly one bit is set in a MTCRF instruction's FXM field.
+ */
+static int
+commpage_onebit(
+    uint32_t   mtcrf )
+{
+    int x = (mtcrf >> 12) & 0xFF;              // isolate the FXM field of the MTCRF
+    
+    if (x==0)
+        panic("commpage bad mtcrf");
+        
+    return     (x & (x-1))==0 ? 1 : 0;         // return 1 iff exactly 1 bit set in FXM field
+}
+
+
+/* Handle kCommPageDCBA bit: this routine uses DCBA.  If the machine we're
+ * running on doesn't benefit from use of that instruction, map them to NOPs
+ * in the commpage.
+ */
+static void
+commpage_handle_dcbas(
+    int        address,
+    int        length  )
+{
+    uint32_t   *ptr, search_mask, search, replace_mask, replace;
+    
+    if ((_cpu_capabilities & kDcbaAvailable) == 0) {
+        ptr = commpage_addr_of(address);
+        
+        search_mask =  0xFC0007FE;             // search x-form opcode bits
+        search =               0x7C0005EC;             // for a DCBA
+        replace_mask =         0xFFFFFFFF;             // replace all bits...
+        replace =              0x60000000;             // ...with a NOP
+    
+        commpage_change(ptr,length,search_mask,search,replace_mask,replace,NULL);
+    }
+}
+
+
+/* Handle kCommPageSYNC bit: this routine uses SYNC or LWSYNC.  If we're
+ * running on a UP machine, map them to NOPs.
+ */
+static void
+commpage_handle_syncs(
+    int        address, 
+    int        length  )
+{
+    uint32_t   *ptr, search_mask, search, replace_mask, replace;
+    
+    if (_NumCPUs() == 1) {
+        ptr = commpage_addr_of(address);
+        
+        search_mask =  0xFC0007FE;             // search x-form opcode bits
+        search =               0x7C0004AC;             // for a SYNC or LWSYNC
+        replace_mask =         0xFFFFFFFF;             // replace all bits...
+        replace =              0x60000000;             // ...with a NOP
+    
+        commpage_change(ptr,length,search_mask,search,replace_mask,replace,NULL);
+    }
+}
+
+
+/* Handle kCommPageMTCRF bit.  When this was written (3/03), the assembler did not
+ * recognize the special form of MTCRF instructions, in which exactly one bit is set
+ * in the 8-bit mask field.  Bit 11 of the instruction should be set in this case,
+ * since the 970 and probably other 64-bit processors optimize it.  Once the assembler
+ * has been updated this code can be removed, though it need not be.
+ */
+static void
+commpage_handle_mtcrfs(
+    int        address, 
+    int        length  )
+{
+    uint32_t   *ptr, search_mask, search, replace_mask, replace;
+    
+    if (_cpu_capabilities & k64Bit) {
+        ptr = commpage_addr_of(address);
+        
+        search_mask =  0xFC0007FE;             // search x-form opcode bits
+        search =               0x7C000120;             // for a MTCRF
+        replace_mask =         0x00100000;             // replace bit 11...
+        replace =              0x00100000;             // ...with a 1-bit
+    
+        commpage_change(ptr,length,search_mask,search,replace_mask,replace,commpage_onebit);
+    }
+}
+
+
+/* Copy a routine into comm page if it matches running machine.
+ */
+static void
+commpage_stuff_routine(
+    commpage_descriptor        *rd     )
+{
+    char       *routine_code;
+    int                must,cant;
+    
+    if (rd->commpage_address != cur_routine) {
+        if ((cur_routine!=0) && (matched==0))
+            panic("commpage no match");
+        cur_routine = rd->commpage_address;
+        matched = 0;
+    }
+    
+    must = _cpu_capabilities & rd->musthave;
+    cant = _cpu_capabilities & rd->canthave;
+    
+    if ((must == rd->musthave) && (cant == 0)) {
+        if (matched)
+            panic("commpage duplicate matches");
+        matched = 1;
+        routine_code = ((char*)rd) + rd->code_offset;
+        
+        commpage_stuff(rd->commpage_address,routine_code,rd->code_length);
+        
+        if (rd->special & kCommPageDCBA)
+            commpage_handle_dcbas(rd->commpage_address,rd->code_length);
+            
+        if (rd->special & kCommPageSYNC)
+            commpage_handle_syncs(rd->commpage_address,rd->code_length);
+            
+        if (rd->special & kCommPageMTCRF)
+            commpage_handle_mtcrfs(rd->commpage_address,rd->code_length);
+    }
+}
+
+
 /* Fill in commpage: called once, during kernel initialization, from the
  * startup thread before user-mode code is running.
  * See the top of this file for a list of what you have to do to add
@@ -94,5 +358,136 @@ commpage_init_cpu_capabilities( void )
 void
 commpage_populate( void )
 {
+    char       c1;
+    short      c2;
+    addr64_t c8;
+    static double      two52 = 1048576.0 * 1048576.0 * 4096.0; // 2**52
+    static double      ten6 = 1000000.0;                                               // 10**6
+    commpage_descriptor        **rd;
+    short      version = _COMM_PAGE_THIS_VERSION;
+    
+    
+    commPagePtr = (char*) commpage_allocate();
+    
     commpage_init_cpu_capabilities();
+
+
+    /* Stuff in the constants.  We move things into the comm page in strictly
+     * ascending order, so we can check for overlap and panic if so.
+     */
+     
+    commpage_stuff(_COMM_PAGE_VERSION,&version,2);
+
+    commpage_stuff(_COMM_PAGE_CPU_CAPABILITIES,&_cpu_capabilities,sizeof(int));
+    
+    c1 = (_cpu_capabilities & kHasAltivec) ? -1 : 0;
+    commpage_stuff(_COMM_PAGE_ALTIVEC,&c1,1);
+    
+    c1 = (_cpu_capabilities & k64Bit) ? -1 : 0;
+    commpage_stuff(_COMM_PAGE_64_BIT,&c1,1);
+    
+    if (_cpu_capabilities & kCache32)
+        c2 = 32;
+    else if (_cpu_capabilities & kCache64)
+        c2 = 64;
+    else if (_cpu_capabilities & kCache128)
+        c2 = 128;
+    commpage_stuff(_COMM_PAGE_CACHE_LINESIZE,&c2,2);
+    
+    commpage_stuff(_COMM_PAGE_2_TO_52,&two52,8);
+    
+    commpage_stuff(_COMM_PAGE_10_TO_6,&ten6,8);
+    
+    c8 = 0;                                                                                                    // 0 timestamp means "disabled"
+    commpage_stuff(_COMM_PAGE_TIMEBASE,&c8,8);
+    commpage_stuff(_COMM_PAGE_TIMESTAMP,&c8,8);
+    commpage_stuff(_COMM_PAGE_SEC_PER_TICK,&c8,8);
+
+
+    /* Now the routines.  We try each potential routine in turn,
+     * and copy in any that "match" the platform we are running on.
+     * We require that exactly one routine match for each slot in the
+     * comm page, and panic if not.
+     *
+     * The check for overlap assumes that these routines are
+     * in strictly ascending order, sorted by address in the
+     * comm page.
+     */
+
+    extern     commpage_descriptor     mach_absolute_time_32;
+    extern     commpage_descriptor     mach_absolute_time_64;
+    extern     commpage_descriptor     spinlock_32_try_mp;
+    extern     commpage_descriptor     spinlock_32_try_up;
+    extern     commpage_descriptor     spinlock_64_try_mp;
+    extern     commpage_descriptor     spinlock_64_try_up;
+    extern     commpage_descriptor     spinlock_32_lock_mp;
+    extern     commpage_descriptor     spinlock_32_lock_up;
+    extern     commpage_descriptor     spinlock_64_lock_mp;
+    extern     commpage_descriptor     spinlock_64_lock_up;
+    extern     commpage_descriptor     spinlock_32_unlock_mp;
+    extern     commpage_descriptor     spinlock_32_unlock_up;
+    extern     commpage_descriptor     spinlock_64_unlock_mp;
+    extern     commpage_descriptor     spinlock_64_unlock_up;
+    extern     commpage_descriptor     pthread_getspecific_sprg3;
+    extern     commpage_descriptor     pthread_getspecific_uftrap;
+    extern     commpage_descriptor     gettimeofday_32;
+    extern     commpage_descriptor     gettimeofday_64;
+    extern     commpage_descriptor     commpage_flush_dcache;
+    extern     commpage_descriptor     commpage_flush_icache;
+    extern     commpage_descriptor     pthread_self_sprg3;
+    extern     commpage_descriptor     pthread_self_uftrap;
+    extern     commpage_descriptor     spinlock_relinquish;
+    extern     commpage_descriptor     bzero_32;
+    extern     commpage_descriptor     bzero_128;
+    extern     commpage_descriptor     bcopy_g3;
+    extern     commpage_descriptor     bcopy_g4;
+    extern     commpage_descriptor     bcopy_970;
+    extern     commpage_descriptor     bcopy_64;
+    extern     commpage_descriptor     bigcopy_970;
+    
+    static     commpage_descriptor     *routines[] = {
+        &mach_absolute_time_32,
+        &mach_absolute_time_64,
+        &spinlock_32_try_mp,
+        &spinlock_32_try_up,
+        &spinlock_64_try_mp,
+        &spinlock_64_try_up,
+        &spinlock_32_lock_mp,
+        &spinlock_32_lock_up,
+        &spinlock_64_lock_mp,
+        &spinlock_64_lock_up,
+        &spinlock_32_unlock_mp,
+        &spinlock_32_unlock_up,
+        &spinlock_64_unlock_mp,
+        &spinlock_64_unlock_up,
+        &pthread_getspecific_sprg3,
+        &pthread_getspecific_uftrap,
+        &gettimeofday_32,
+        &gettimeofday_64,
+        &commpage_flush_dcache,
+        &commpage_flush_icache,
+        &pthread_self_sprg3,
+        &pthread_self_uftrap,
+        &spinlock_relinquish,
+        &bzero_32,
+        &bzero_128,
+        &bcopy_g3,
+        &bcopy_g4,
+        &bcopy_970,
+        &bcopy_64,
+        &bigcopy_970,
+        NULL };
+        
+    for( rd = routines; *rd != NULL ; rd++ ) 
+        commpage_stuff_routine(*rd);
+        
+    if (!matched)
+        panic("commpage no match on last routine");
+    
+    if (next > (commPagePtr + _COMM_PAGE_AREA_USED))
+        panic("commpage overflow");
+        
+    sync_cache_virtual((vm_offset_t) commPagePtr,_COMM_PAGE_AREA_USED);        // make all that new code executable
+
 }
+