libc: kernel: Update msm_kgsl.h header

Change-Id: I8b22bcb4816fca5e0b92fee4b2ccb0ef135be5a9
Signed-off-by: Dima Zavin <dima@android.com>

libc/arch-arm/bionic/memcpy.S

@@ -37,8 +37,9 @@
         .type memcpy, %function
         .align 4
 
-/* a prefetch distance of 32*4 works best experimentally */
-#define PREFETCH_DISTANCE   (32*4)
+/* a prefetch distance of 4 cache-lines works best experimentally */
+#define CACHE_LINE_SIZE     64
+#define PREFETCH_DISTANCE   (CACHE_LINE_SIZE*4)
 
 memcpy:
         .fnstart
@@ -46,8 +47,8 @@ memcpy:
         stmfd       sp!, {r0, lr}
 
         /* start preloading as early as possible */
-        pld         [r1, #0]
-        pld         [r1, #32]
+        pld         [r1, #(CACHE_LINE_SIZE*0)]
+        pld         [r1, #(CACHE_LINE_SIZE*1)]
 
         /* do we have at least 16-bytes to copy (needed for alignment below) */
         cmp         r2, #16
@@ -79,13 +80,11 @@ memcpy:
 2:
 
 0:      /* preload immediately the next cache line, which we may need */
-        pld         [r1, #(32*0)]
-        pld         [r1, #(32*1)]
-        pld         [r1, #(32*2)]
-        pld         [r1, #(32*3)]
+        pld         [r1, #(CACHE_LINE_SIZE*0)]
+        pld         [r1, #(CACHE_LINE_SIZE*1)]
 
-        /* make sure we have at least 128 bytes to copy */
-        subs        r2, r2, #128
+        /* make sure we have at least 64 bytes to copy */
+        subs        r2, r2, #64
         blo         2f
 
         /* preload all the cache lines we need.
@@ -94,29 +93,21 @@ memcpy:
          * avoid the goofy code below. In practice this doesn't seem to make
          * a big difference.
          */
-        pld         [r1, #(PREFETCH_DISTANCE + 32*0)]
-        pld         [r1, #(PREFETCH_DISTANCE + 32*1)]
-        pld         [r1, #(PREFETCH_DISTANCE + 32*2)]
-        pld         [r1, #(PREFETCH_DISTANCE + 32*3)]
+        pld         [r1, #(CACHE_LINE_SIZE*2)]
+        pld         [r1, #(CACHE_LINE_SIZE*3)]
+        pld         [r1, #(PREFETCH_DISTANCE)]
 
-1:      /* The main loop copies 128 bytes at a time */
+1:      /* The main loop copies 64 bytes at a time */
         vld1.8      {d0  - d3},   [r1]!
         vld1.8      {d4  - d7},   [r1]!
-        vld1.8      {d16 - d19},  [r1]!
-        vld1.8      {d20 - d23},  [r1]!
-        pld         [r1, #(PREFETCH_DISTANCE + 32*0)]
-        pld         [r1, #(PREFETCH_DISTANCE + 32*1)]
-        pld         [r1, #(PREFETCH_DISTANCE + 32*2)]
-        pld         [r1, #(PREFETCH_DISTANCE + 32*3)]
-        subs        r2, r2, #128
+        pld         [r1, #(PREFETCH_DISTANCE)]
+        subs        r2, r2, #64
         vst1.8      {d0  - d3},   [r0, :128]!
         vst1.8      {d4  - d7},   [r0, :128]!
-        vst1.8      {d16 - d19},  [r0, :128]!
-        vst1.8      {d20 - d23},  [r0, :128]!
         bhs         1b
 
 2:      /* fix-up the remaining count and make sure we have >= 32 bytes left */
-        add         r2, r2, #128
+        add         r2, r2, #64
         subs        r2, r2, #32
         blo         4f
 

libc/arch-arm/bionic/memset.S

@@ -80,7 +80,7 @@ memset:
         
 		rsb         r3, r0, #0
 		ands		r3, r3, #0x1C
-		beq         aligned32
+		beq         3f
 		cmp         r3, r2
 		andhi		r3, r2, #0x1C
 		sub         r2, r2, r3
@@ -93,7 +93,7 @@ memset:
 		movs		r3, r3, lsl #2
         strcs       r1, [r0], #4
 
-aligned32:
+3:
         subs        r2, r2, #32
         mov         r3, r1
         bmi         2f

libc/kernel/common/linux/msm_kgsl.h

@@ -188,5 +188,12 @@ struct kgsl_sharedmem_from_vmalloc {
 
 #define IOCTL_KGSL_SHAREDMEM_FLUSH_CACHE   _IOW(KGSL_IOC_TYPE, 0x24, struct kgsl_sharedmem_free)
 
+struct kgsl_drawctxt_set_bin_base_offset {
+ unsigned int drawctxt_id;
+ unsigned int offset;
+};
+
+#define IOCTL_KGSL_DRAWCTXT_SET_BIN_BASE_OFFSET   _IOW(KGSL_IOC_TYPE, 0x25, struct kgsl_drawctxt_set_bin_base_offset)
+
 #endif
 

libc/unistd/abort.c

@@ -39,8 +39,13 @@
 #define debug_log(format, ...)  \
     __libc_android_log_print(ANDROID_LOG_DEBUG, "libc-abort", (format), ##__VA_ARGS__ )
 
+#ifdef __arm__
+void
+__libc_android_abort(void)
+#else
 void
 abort(void)
+#endif
 {
 	struct atexit *p = __atexit;
 	static int cleanup_called = 0;
@@ -97,3 +102,29 @@ abort(void)
 	(void)kill(getpid(), SIGABRT);
 	_exit(1);
 }
+
+#ifdef __arm__
+/*
+ * abort() does not return, which gcc interprets to mean that it doesn't
+ * need to preserve any of the callee-save registers.  Unfortunately this
+ * includes the link register, so if LR is used there is no way to determine
+ * which function called abort().
+ *
+ * We work around this by inserting a trivial stub that doesn't alter
+ * any of the "interesting" registers and thus doesn't need to save them.
+ * We can't just call __libc_android_abort from C because gcc uses "bl"
+ * without first saving LR, so we use an asm statement.  This also has
+ * the side-effect of replacing abort() with __libc_android_abort() in
+ * the stack trace.
+ *
+ * Ideally __libc_android_abort would be static, but I haven't figured out
+ * how to tell gcc to call a static function from an asm statement.
+ */
+void
+abort(void)
+{
+    asm ("b __libc_android_abort");
+    _exit(1);       /* suppress gcc noreturn warnings */
+}
+#endif
+

linker/ba.c

@@ -30,65 +30,41 @@
 #include "linker_debug.h"
 #include "ba.h"
 
-struct ba_bits {
-    unsigned allocated:1;           /* 1 if allocated, 0 if free */
-    unsigned order:7;               /* size of the region in ba space */
-};
-
-struct ba_info {
-    /* start address of the ba space */
-    unsigned long base;
-    /* total size of the ba space */
-    unsigned long size;
-    /* number of entries in the ba space */
-    int num_entries;
-    /* the bitmap for the region indicating which entries are allocated
-     * and which are free */
-    struct ba_bits *bitmap;
-};
-
 #undef min
 #define min(a,b) ((a)<(b)?(a):(b))
 
-#define BA_MIN_ALLOC LIBINC
-#define BA_MAX_ORDER 128
-#define BA_START LIBBASE
-#define BA_SIZE (LIBLAST - LIBBASE)
-
-#define BA_IS_FREE(index) (!(ba.bitmap[index].allocated))
-#define BA_ORDER(index) ba.bitmap[index].order
+#define BA_IS_FREE(index) (!(ba->bitmap[index].allocated))
+#define BA_ORDER(index) ba->bitmap[index].order
 #define BA_BUDDY_INDEX(index) ((index) ^ (1 << BA_ORDER(index)))
 #define BA_NEXT_INDEX(index) ((index) + (1 << BA_ORDER(index)))
-#define BA_OFFSET(index) ((index) * BA_MIN_ALLOC)
-#define BA_START_ADDR(index) (BA_OFFSET(index) + ba.base)
-#define BA_LEN(index) ((1 << BA_ORDER(index)) * BA_MIN_ALLOC)
+#define BA_OFFSET(index) ((index) * ba->min_alloc)
+#define BA_START_ADDR(index) (BA_OFFSET(index) + ba->base)
+#define BA_LEN(index) ((1 << BA_ORDER(index)) * ba->min_alloc)
 
-static struct ba_bits ba_bitmap[BA_SIZE / BA_MIN_ALLOC];
+static unsigned long ba_order(struct ba *ba, unsigned long len);
 
-static struct ba_info ba = {
-    .base = BA_START,
-    .size = BA_SIZE,
-    .bitmap = ba_bitmap,
-    .num_entries = sizeof(ba_bitmap)/sizeof(ba_bitmap[0]),
-};
-
-void ba_init(void)
+void ba_init(struct ba *ba)
 {
     int i, index = 0;
-    for (i = sizeof(ba.num_entries) * 8 - 1; i >= 0; i--) {
-        if (ba.num_entries &  1<<i) {
+
+    unsigned long max_order = ba_order(ba, ba->size);
+    if (ba->max_order == 0 || ba->max_order > max_order)
+        ba->max_order = max_order;
+
+    for (i = sizeof(ba->num_entries) * 8 - 1; i >= 0; i--) {
+        if (ba->num_entries &  1<<i) {
             BA_ORDER(index) = i;
             index = BA_NEXT_INDEX(index);
         }
     }
 }
 
-int ba_free(int index)
+int ba_free(struct ba *ba, int index)
 {
     int buddy, curr = index;
 
     /* clean up the bitmap, merging any buddies */
-    ba.bitmap[curr].allocated = 0;
+    ba->bitmap[curr].allocated = 0;
     /* find a slots buddy Buddy# = Slot# ^ (1 << order)
      * if the buddy is also free merge them
      * repeat until the buddy is not free or end of the bitmap is reached
@@ -103,16 +79,16 @@ int ba_free(int index)
         } else {
             break;
         }
-    } while (curr < ba.num_entries);
+    } while (curr < ba->num_entries);
 
     return 0;
 }
 
-static unsigned long ba_order(unsigned long len)
+static unsigned long ba_order(struct ba *ba, unsigned long len)
 {
     unsigned long i;
 
-    len = (len + BA_MIN_ALLOC - 1) / BA_MIN_ALLOC;
+    len = (len + ba->min_alloc - 1) / ba->min_alloc;
     len--;
     for (i = 0; i < sizeof(len)*8; i++)
         if (len >> i == 0)
@@ -120,14 +96,14 @@ static unsigned long ba_order(unsigned long len)
     return i;
 }
 
-int ba_allocate(unsigned long len)
+int ba_allocate(struct ba *ba, unsigned long len)
 {
     int curr = 0;
-    int end = ba.num_entries;
+    int end = ba->num_entries;
     int best_fit = -1;
-    unsigned long order = ba_order(len);
+    unsigned long order = ba_order(ba, len);
 
-    if (order > BA_MAX_ORDER)
+    if (order > ba->max_order)
         return -1;
 
     /* look through the bitmap:
@@ -165,16 +141,16 @@ int ba_allocate(unsigned long len)
         buddy = BA_BUDDY_INDEX(best_fit);
         BA_ORDER(buddy) = BA_ORDER(best_fit);
     }
-    ba.bitmap[best_fit].allocated = 1;
+    ba->bitmap[best_fit].allocated = 1;
     return best_fit;
 }
 
-unsigned long ba_start_addr(int index)
+unsigned long ba_start_addr(struct ba *ba, int index)
 {
     return BA_START_ADDR(index);
 }
 
-unsigned long ba_len(int index)
+unsigned long ba_len(struct ba *ba, int index)
 {
     return BA_LEN(index);
 }

linker/ba.h

@@ -29,10 +29,31 @@
 #ifndef __LINKER_BA_H
 #define __LINKER_BA_H
 
-extern void ba_init(void);
-extern int ba_allocate(unsigned long len);
-extern int ba_free(int index);
-extern unsigned long ba_start_addr(int index);
-extern unsigned long ba_len(int index);
+struct ba_bits {
+    unsigned allocated:1;           /* 1 if allocated, 0 if free */
+    unsigned order:7;               /* size of the region in ba space */
+};
+
+struct ba {
+    /* start address of the ba space */
+    unsigned long base;
+    /* total size of the ba space */
+    unsigned long size;
+    /* the smaller allocation that can be made */
+    unsigned long min_alloc;
+    /* the order of the largest allocation that can be made */
+    unsigned long max_order;
+    /* number of entries in the ba space */
+    int num_entries;
+    /* the bitmap for the region indicating which entries are allocated
+     * and which are free */
+    struct ba_bits *bitmap;
+};
+
+extern void ba_init(struct ba *ba);
+extern int ba_allocate(struct ba *ba, unsigned long len);
+extern int ba_free(struct ba *ba, int index);
+extern unsigned long ba_start_addr(struct ba *ba, int index);
+extern unsigned long ba_len(struct ba *ba, int index);
 
 #endif

linker/linker.c

@@ -91,6 +91,18 @@ static soinfo *sonext = &libdl_info;
 static soinfo *somain; /* main process, always the one after libdl_info */
 #endif
 
+
+/* Set up for the buddy allocator managing the prelinked libraries. */
+static struct ba_bits ba_prelink_bitmap[(LIBLAST - LIBBASE) / LIBINC];
+static struct ba ba_prelink = {
+    .base = LIBBASE,
+    .size = LIBLAST - LIBBASE,
+    .min_alloc = LIBINC,
+    /* max_order will be determined automatically */
+    .bitmap = ba_prelink_bitmap,
+    .num_entries = sizeof(ba_prelink_bitmap)/sizeof(ba_prelink_bitmap[0]),
+};
+
 static inline int validate_soinfo(soinfo *si)
 {
     return (si >= sopool && si < sopool + SO_MAX) ||
@@ -783,14 +795,14 @@ alloc_mem_region(soinfo *si)
        for it from the buddy allocator, which manages the area between
        LIBBASE and LIBLAST.
     */
-    si->ba_index = ba_allocate(si->size);
+    si->ba_index = ba_allocate(&ba_prelink, si->size);
     if(si->ba_index >= 0) {
-        si->base = ba_start_addr(si->ba_index);
+        si->base = ba_start_addr(&ba_prelink, si->ba_index);
         PRINT("%5d mapping library '%s' at %08x (index %d) " \
               "through buddy allocator.\n",
               pid, si->name, si->base, si->ba_index);
         if (reserve_mem_region(si) < 0) {
-            ba_free(si->ba_index);
+            ba_free(&ba_prelink, si->ba_index);
             si->ba_index = -1;
             si->base = 0;
             goto err;
@@ -1086,7 +1098,7 @@ load_library(const char *name)
     /* Now actually load the library's segments into right places in memory */
     if (load_segments(fd, &__header[0], si) < 0) {
         if (si->ba_index >= 0) {
-            ba_free(si->ba_index);
+            ba_free(&ba_prelink, si->ba_index);
             si->ba_index = -1;
         }
         goto fail;
@@ -1189,7 +1201,7 @@ unsigned unload_library(soinfo *si)
             PRINT("%5d releasing library '%s' address space at %08x "\
                   "through buddy allocator.\n",
                   pid, si->name, si->base);
-            ba_free(si->ba_index);
+            ba_free(&ba_prelink, si->ba_index);
         }
         notify_gdb_of_unload(si);
         free_info(si);
@@ -1240,9 +1252,13 @@ static int reloc_library(soinfo *si, Elf32_Rel *rel, unsigned count)
                 return -1;
             }
 #endif
-            if ((s->st_shndx == SHN_UNDEF) && (s->st_value != 0)) {
-                DL_ERR("%5d In '%s', shndx=%d && value=0x%08x. We do not "
-                      "handle this yet", pid, si->name, s->st_shndx,
+            // st_shndx==SHN_UNDEF means an undefined symbol.
+            // st_value should be 0 then, except that the low bit of st_value is
+            // used to indicate whether the symbol points to an ARM or thumb function,
+            // and should be ignored in the following check.
+            if ((s->st_shndx == SHN_UNDEF) && ((s->st_value & ~1) != 0)) {
+                DL_ERR("%5d In '%s', symbol=%s shndx=%d && value=0x%08x. We do not "
+                      "handle this yet", pid, si->name, sym_name, s->st_shndx,
                       s->st_value);
                 return -1;
             }
@@ -1893,7 +1909,7 @@ unsigned __linker_init(unsigned **elfdata)
         vecs += 2;
     }
 
-    ba_init();
+    ba_init(&ba_prelink);
 
     si->base = 0;
     si->dynamic = (unsigned *)-1;