make coroutine.c signal-safe, add cr_cid_info()

3 files changed, 311 insertions, 224 deletions

diff --git a/libcr/CMakeLists.txt b/libcr/CMakeLists.txt
index 14acba6..92a290a 100644
--- a/libcr/CMakeLists.txt
+++ b/libcr/CMakeLists.txt
@@ -8,3 +8,6 @@ target_include_directories(libcr SYSTEM INTERFACE ${CMAKE_CURRENT_LIST_DIR}/incl
 target_sources(libcr INTERFACE
 	coroutine.c
 )
+target_compile_options(libcr INTERFACE
+	-fno-split-stack
+)
diff --git a/libcr/coroutine.c b/libcr/coroutine.c
index b3dccff..8bfdb86 100644
--- a/libcr/coroutine.c
+++ b/libcr/coroutine.c
@@ -18,19 +18,19 @@
 #include "config.h"
 
 #ifndef CONFIG_COROUTINE_DEFAULT_STACK_SIZE
-#  error config.h must define CONFIG_COROUTINE_DEFAULT_STACK_SIZE
+#	error config.h must define CONFIG_COROUTINE_DEFAULT_STACK_SIZE
 #endif
 #ifndef CONFIG_COROUTINE_NUM
-#  error config.h must define CONFIG_COROUTINE_NUM
+#	error config.h must define CONFIG_COROUTINE_NUM
 #endif
 #ifndef CONFIG_COROUTINE_MEASURE_STACK
-#  error config.h must define CONFIG_COROUTINE_MEASURE_STACK
+#	error config.h must define CONFIG_COROUTINE_MEASURE_STACK
 #endif
 #ifndef CONFIG_COROUTINE_PROTECT_STACK
-#  error config.h must define CONFIG_COROUTINE_PROTECT_STACK
+#	error config.h must define CONFIG_COROUTINE_PROTECT_STACK
 #endif
 #ifndef CONFIG_COROUTINE_DEBUG
-#  error config.h must define CONFIG_COROUTINE_DEBUG
+#	error config.h must define CONFIG_COROUTINE_DEBUG
 #endif
 
 /* Implementation *************************************************************/
@@ -38,17 +38,17 @@
 /*
  * Portability notes:
  *
- * - It uses GCC `__attribute__`s, and the GNUC ({ ... }) statement exprs
-     etension.
+ * - It uses GCC `__attribute__`s, and the GNUC ({ ... }) statement
+ *   exprs extension.
  *
- * - It has a small bit of platform-specific code in the "platform support"
- *   section.  Other than this, it should be portable to other platforms CPUs.
- *   It currently contains implementations for __x86_64__ (assumes POSIX) and
- *   __arm__ (assumes ARM CMSIS), and should be fairly easy to add
- *   implementations for other platforms.
+ * - It has a small bit of platform-specific code in the "platform
+ *   support" section.  Other than this, it should be portable to
+ *   other platforms CPUs.  It currently contains implementations for
+ *   __x86_64__ (assumes POSIX) and __arm__ (assumes bare-metal), and
+ *   should be fairly easy to add implementations for other platforms.
  *
  * - It uses setjmp()/longjmp() in "unsafe" ways.  POSIX-2017
- *   longjmp(3) says
+ *   longjmp(3p) says
  *
  *   > If the most recent invocation of setjmp() with the
  *   > corresponding jmp_buf ... or if the function containing the
@@ -65,29 +65,29 @@
  *   choose to do something else, but I'm calling it out here because
  *   you never know.
  *
- *   Note that setjmp()/longjmp() are defined in 3 places: in the libc
- *   (glibc/newlib), as GCC intrinsics, and the lower-level GCC
- *   __builtin_{setjmp,newlib} which the libc and intrinsic versions
- *   likely use.  Our assumptions seem to be valid for
+ *   Our assumptions seem to be valid for
  *   x86_64-pc-linux-gnu/gcc-14.2.1/glibc-2.40 and
  *   arm-none-eabi/gcc-14.1.0/newlib-4.4.0.
  *
- * Why not use <ucontext.h>, the now-deprecated (was in POSIX.1-2001,
- * is gone in POSIX-2008) predecesor to <setjmp.h>?  It would let us
- * do this without any assembly or unsafe assumptions.  Simply:
- * because newlib does not provide it.
+ *   Why not use <ucontext.h>, the now-deprecated (was in
+ *   POSIX.1-2001, is gone in POSIX-2008) predecesor to <setjmp.h>?
+ *   It would let us do this without any assembly or unsafe
+ *   assumptions.  Simply: because newlib does not provide it.  But it
+ *   would let us avoid having an `sp` member in `struct coroutine`...
+ *   Maybe https://github.com/kaniini/libucontext ?  Or building a
+ *   ucontext-lib abstraction on top of setjmp/longjmp?
  */
 
 /*
  * Design decisions and notes:
  *
- * - Coroutines are launched with a stack stack that is filled with
- *   known values (zero or a pre-determined pattern, depending on the
- *   config above).  Because all stack variables should be initialized
- *   before they are read, this "shouldn't" make a difference, but:
- *   (1) Initializing it to a known value allows us to measure how
- *   much of the stack was written to, which is helpful to tune stack
- *   sizes.  (2) Leaving it uninitialized just gives me the willies.
+ * - Coroutines are launched with a stack that is filled with known
+ *   arbitrary values.  Because all stack variables should be
+ *   initialized before they are read, this "shouldn't" make a
+ *   difference, but: (1) Initializing it to a known value allows us
+ *   to measure how much of the stack was written to, which is helpful
+ *   to tune stack sizes.  (2) Leaving it uninitialized just gives me
+ *   the willies.
  *
  * - Because embedded programs should be adverse to using the heap,
  *   CONFIG_COROUTINE_NUM is fixed, instead of having coroutine_add()
@@ -111,12 +111,133 @@
  *   no longer exists.
  */
 
-#if __x86_64__
-#  include <unistd.h> /* for pause() */
-#  define wait_for_interrupt() pause()
-#elif __arm__
-#  include "hardware/sync.h" /* for __wfi(); */
-#  define wait_for_interrupt() __wfi()
+#define ALWAYS_INLINE inline __attribute__((always_inline))
+
+/* platform support ***********************************************************/
+
+/* As part of sbc-harness, this only really needs to support ARM-32, but being
+ * able to run it on my x86-64 GNU/Linux laptop is useful for debugging.  */
+
+#define CR_PLAT_STACK_ALIGNMENT \
+	({ __attribute__((aligned)) void fn(void) {}; __alignof__(fn); })
+
+#if 0
+{ /* to get Emacs indentation to work how I want */
+#endif
+#if __arm__ /*======================================================*/
+	/* Assume bare metal ARMv6-M.  */
+
+	#define CR_PLAT_STACK_GROWS_DOWNWARD 1
+
+	/* Wrappers for setjmp()/longjmp() that do *not* save the
+	 * interrupt mask.
+	 *
+	 * newlib setjmp()/longjmp() do not save the interrupt mask, *
+	 * so we can use them directly.  */
+	#define cr_plat_setjmp(env)             setjmp(env)
+	#define cr_plat_longjmp(env, val)       longjmp(env, val)
+
+	static ALWAYS_INLINE void cr_plat_wait_for_interrupt(void) {
+		asm volatile ("wfi":::"memory");
+	}
+	static ALWAYS_INLINE void cr_plat_disable_interrupts(void)  {
+		asm volatile ("cpsid i":::"memory");
+	}
+	static ALWAYS_INLINE void cr_plat_enable_interrupts(void)  {
+		asm volatile ("cpsie i":::"memory");
+	}
+
+	#if CONFIG_COROUTINE_MEASURE_STACK
+	static ALWAYS_INLINE uintptr_t cr_plat_get_sp(void) {
+		uintptr_t sp;
+		asm volatile ("mov %0, sp":"=r"(sp));
+		return sp;
+	}
+	#endif
+
+	static void cr_plat_call_with_stack(void *stack,
+	                                    cr_fn_t fn, void *args) {
+		static void *saved_sp = NULL;
+		/* str/ldr can only work with a "lo" register, which
+		 * sp is not, so we use r0 as an intermediate because
+		 * we're going to clobber it with args anyway.  */
+		asm volatile ("mov r0, sp\n\t" /* [saved_sp = sp */
+		              "str r0, %0\n\t" /* ] */
+		              "mov sp, %1\n\t" /* [sp = stack] */
+		              "mov r0, %3\n\t" /* [arg0 = args] */
+		              "blx %2\n\t"     /* [fn()] */
+		              "ldr r0, %0\n\t" /* [sp = staved_sp */
+		              "mov sp, r0"     /* ] */
+		              :
+		              : /* %0 */"m"(saved_sp),
+		                /* %1 */"r"(stack),
+		                /* %2 */"r"(fn),
+		                /* %3 */"r"(args)
+		              : "r0"
+		              );
+	}
+
+#elif __x86_64__ /*=================================================*/
+	/* Assume POSIX. */
+	#include <signal.h> /* for sig*, SIG_* */
+	#include <unistd.h> /* for pause() */
+
+	#define CR_PLAT_STACK_GROWS_DOWNWARD 1
+
+	/* Wrappers for setjmp()/longjmp() that do *not* save the
+	 * interrupt mask.
+	 *
+	 * POSIX leaves it implementation-defined whether they save
+	 * the signal mask; while glibc does not save it, let's not
+	 * rely on that.  */
+	#define cr_plat_setjmp(env)             sigsetjmp(env, 0)
+	#define cr_plat_longjmp(env, val)       siglongjmp(env, val)
+
+	static inline void cr_plat_wait_for_interrupt(void) {
+		pause();
+	}
+	static inline void cr_plat_disable_interrupts(void) {
+		sigset_t all;
+		sigfillset(&all);
+		sigprocmask(SIG_BLOCK, &all, NULL);
+	}
+	static inline void cr_plat_enable_interrupts(void) {
+		sigset_t all;
+		sigfillset(&all);
+		sigprocmask(SIG_UNBLOCK, &all, NULL);
+	}
+
+	#if CONFIG_COROUTINE_MEASURE_STACK
+	static ALWAYS_INLINE uintptr_t cr_plat_get_sp(void)  {
+		uintptr_t sp;
+		asm volatile ("movq %%rsp, %0":"=r"(sp));
+		return sp;
+	}
+	#endif
+
+	static void cr_plat_call_with_stack(void *stack,
+	                                    cr_fn_t fn, void *args) {
+		static void *saved_sp = NULL;
+		asm volatile ("movq %%rsp , %0\n\t"    /* saved_sp = sp */
+		              "movq %1    , %%rsp\n\t" /* sp = stack */
+		              "movq %3    , %%rdi\n\t" /* arg0 = args */
+		              "call *%2\n\t"           /* fn() */
+		              "movq %0    , %%rsp"     /* sp = saved_sp */
+		              :
+		              : /* %0 */"m"(saved_sp),
+		                /* %1 */"r"(stack),
+		                /* %2 */"r"(fn),
+		                /* %3 */"r"(args)
+		              : "rdi"
+		              );
+	}
+
+#else
+	#error unsupported platform
+
+#endif
+#if 0
+}
 #endif
 
 /* types **********************************************************************/
@@ -125,7 +246,8 @@ enum coroutine_state {
 	CR_NONE = 0,     /* this slot in the table is empty */
 	CR_INITIALIZING, /* running, before cr_begin() */
 	CR_RUNNING,      /* running, after cr_begin() */
-	CR_PRE_RUNNABLE, /* running, between cr_unpause_from_intrhandler() and cr_pause() */
+	CR_PRE_RUNNABLE, /* running, after cr_unpause_from_intrhandler()
+	                  *          but before cr_pause() */
 	CR_RUNNABLE,     /* not running, but runnable */
 	CR_PAUSED,       /* not running, and not runnable */
 };
@@ -133,6 +255,12 @@ enum coroutine_state {
 struct coroutine {
 	volatile enum coroutine_state  state;
 	jmp_buf                        env;
+#if CONFIG_COROUTINE_MEASURE_STACK
+	/* We aught to be able to get this out of .env, but libc
+	 * authors insist on jmp_buf being opaque, glibc going as far
+	 * as to xor it with a secret ot obfuscate it!  */
+	uintptr_t                      sp;
+#endif
 	size_t                         stack_size;
 	void                          *stack;
 };
@@ -156,6 +284,12 @@ static const uint8_t stack_pattern[] = {
 	0x4b, 0x14, 0xc4, 0xe0, 0xea, 0x62, 0x25, 0x63,
 };
 #endif
+#if CONFIG_COROUTINE_PROTECT_STACK
+#	define STACK_GUARD_SIZE \
+		round_up(sizeof(stack_pattern), CR_PLAT_STACK_ALIGNMENT)
+#else
+#	define STACK_GUARD_SIZE 0
+#endif
 
 /* global variables ***********************************************************/
 
@@ -193,24 +327,23 @@ static cid_t            coroutine_running                     = 0;
 #endif
 
 #ifdef __GLIBC__
-#  define assertf(expr, ...)                                          \
-  ((void) sizeof ((expr) ? 1 : 0), __extension__ ({                   \
-      if (expr)                                                       \
-        ; /* empty */                                                 \
-      else {                                                          \
-        errorf("assertion: " __VA_ARGS__);                            \
-        __assert_fail (#expr, __FILE__, __LINE__, __ASSERT_FUNCTION); \
-      }                                                               \
-    }))
+#	define assertf(expr, ...) ({                                        \
+		if (!(expr)) {                                              \
+			errorf("assertion: " __VA_ARGS__);                  \
+			__assert_fail(#expr, __FILE__, __LINE__, __func__); \
+		}                                                           \
+	})
 #else
-#  define assertf(expr, ...) assert(expr)
+#	define assertf(expr, ...) assert(expr)
 #endif
 
 /** Return `n` rounded up to the nearest multiple of `d` */
-#define round_up(n, d) ( ( ((n)+(d)-1) / (d) ) * (d) )
+#define round_up(n, d)          ( ( ((n)+(d)-1) / (d) ) * (d) )
+
+#define array_len(arr)          (sizeof(arr)/sizeof(arr[0]))
 
 static inline const char* coroutine_state_str(enum coroutine_state state) {
-	assert(state < (sizeof(coroutine_state_strs)/sizeof(coroutine_state_strs[0])));
+	assert(state < array_len(coroutine_state_strs));
 	return coroutine_state_strs[state];
 }
 
@@ -219,121 +352,22 @@ static inline void assert_cid(cid_t cid) {
 	assert(cid <= CONFIG_COROUTINE_NUM);
 #if CONFIG_COROUTINE_PROTECT_STACK
 	assert(coroutine_table[cid-1].stack_size);
-	assert(coroutine_table[cid-1].stack);
-	for (size_t i = 0; i < sizeof(stack_pattern); i++) {
+	uint8_t *stack = coroutine_table[cid-1].stack;
+	assert(stack);
+	for (size_t i = 0; i < STACK_GUARD_SIZE; i++) {
 		size_t j = coroutine_table[cid-1].stack_size - (i+1);
-		assert(((uint8_t*)coroutine_table[cid-1].stack)[i] == stack_pattern[i]);
-		assert(((uint8_t*)coroutine_table[cid-1].stack)[j] == stack_pattern[j%sizeof(stack_pattern)]);
+		assert(stack[i] == stack_pattern[i%sizeof(stack_pattern)]);
+		assert(stack[j] == stack_pattern[j%sizeof(stack_pattern)]);
 	}
 #endif
 }
 
-#define assert_cid_state(cid, expr) do {                      \
-		assert_cid(cid);                              \
+#define assert_cid_state(cid, expr) do {	  \
+		assert_cid(cid); \
 		cid_t state = coroutine_table[(cid)-1].state; \
-		assert(expr);                                 \
+		assert(expr); \
 	} while (0)
 
-/* platform support ***********************************************************/
-
-/* As part of sbc-harness, this only really needs to support ARM-32, but being
- * able to run it on my x86-64 GNU/Linux laptop is useful for debugging.  */
-
-#define CR_PLAT_STACK_ALIGNMENT ({ __attribute__((aligned)) void fn(void) {}; __alignof__(fn); })
-
-#if 0
-{ /* to get Emacs indentation to work how I want */
-#endif
-#if __arm__
-	/* Assume bare metal ARMv6-M.  */
-	#include "cmsis_compiler.h" /* for __WFI(), __disable_irq(), __enable_irq() */
-
-	#define CR_PLAT_STACK_GROWS_DOWNWARD 1
-
-	static void cr_plat_wait_for_interrupt(void) { asm volatile ("wfi":::"memory"); }
-	static void cr_plat_disable_interrupt(void)  { asm volatile ("cpsid i":::"memory"); }
-	static void cr_plat_enable_interrupts(void)  { asm volatile ("cpsie i":::"memory"); }
-
-#if CONFIG_COROUTINE_MEASURE_STACK
-	static size_t cr_plat_get_sp(void)  {
-		size_t sp;
-		asm volatile ("mov %0, sp":"=r"(sp));
-		return sp;
-	}
-#endif
-
-	static void cr_plat_call_with_stack(void *stack, cr_fn_t fn, void *args) {
-		static void *saved_sp = NULL;
-		/* str/ldr can only work with a "lo" register, which sp is
-		 * not, so we use r0 as an intermediate because we're going to
-		 * clobber it with args anyway.  */
-		asm volatile ("mov r0, sp\n\t" /* [saved_sp = sp */
-		              "str r0, %0\n\t" /* ] */
-		              "mov sp, %1\n\t" /* [sp = stack] */
-		              "mov r0, %3\n\t" /* [arg0 = args] */
-		              "blx %2\n\t"     /* [fn()] */
-		              "ldr r0, %0\n\t" /* [sp = staved_sp */
-		              "mov sp, r0"     /* ] */
-		              :
-		              : /* %0 */"m"(saved_sp),
-		                /* %1 */"r"(stack),
-		                /* %2 */"r"(fn),
-		                /* %3 */"r"(args)
-		              : "r0"
-		              );
-	}
-
-#elif __x86_64__
-	/* Assume POSIX. */
-	#include <signal.h> /* for sigset_t, sigprocmask(), SIG_*, sigfillset() */
-	#include <unistd.h> /* for pause() */
-
-	#define CR_PLAT_STACK_GROWS_DOWNWARD 1
-
-	static void cr_plat_wait_for_interrupt(void) {
-		pause();
-	}
-	static void cr_plat_disable_interrupts(void) {
-		sigset_t all;
-		sigfillset(&all);
-		sigprocmask(SIG_BLOCK, &all, NULL);
-	}
-	static void cr_plat_enable_interrupts(void) {
-		sigset_t all;
-		sigfillset(&all);
-		sigprocmask(SIG_UNBLOCK, &all, NULL);
-	}
-
-#if CONFIG_COROUTINE_MEASURE_STACK
-	static size_t cr_plat_get_sp(void)  {
-		size_t sp;
-		asm volatile ("movq %%rsp, %0":"=r"(sp));
-		return sp;
-	}
-#endif
-
-	static void cr_plat_call_with_stack(void *stack, cr_fn_t fn, void *args) {
-		static void *saved_sp = NULL;
-		asm volatile ("movq %%rsp , %0\n\t"    /* saved_sp = sp */
-		              "movq %1    , %%rsp\n\t" /* sp = stack */
-		              "movq %3    , %%rdi\n\t" /* arg0 = args */
-		              "call *%2\n\t"           /* fn() */
-		              "movq %0    , %%rsp"     /* sp = saved_sp */
-		              :
-		              : /* %0 */"m"(saved_sp),
-		                /* %1 */"r"(stack),
-		                /* %2 */"r"(fn),
-		                /* %3 */"r"(args)
-		              : "rdi"
-		              );
-
-#else
-	#error unsupported platform
-
-#endif
-#if 0
-}
-#endif
 
 /* coroutine_add() ************************************************************/
 
@@ -345,13 +379,14 @@ cid_t coroutine_add_with_stack_size(size_t stack_size, cr_fn_t fn, void *args) {
 		assert_cid_state(parent, state == CR_RUNNING);
 	assert(stack_size);
 	assert(fn);
-	debugf("coroutine_add_with_stack_size(%zu, %p, %p)...\n", stack_size, fn, args);
+	debugf("coroutine_add_with_stack_size(%zu, %p, %p)...\n",
+	       stack_size, fn, args);
 
 	cid_t child;
 	{
-		size_t idx_base = last_created;
-		for (size_t idx_shift = 0; idx_shift < CONFIG_COROUTINE_NUM; idx_shift++) {
-			child = ((idx_base + idx_shift) % CONFIG_COROUTINE_NUM) + 1;
+		size_t base = last_created;
+		for (size_t shift = 0; shift < CONFIG_COROUTINE_NUM; shift++) {
+			child = ((base + shift) % CONFIG_COROUTINE_NUM) + 1;
 			if (coroutine_table[child-1].state == CR_NONE)
 				goto found;
 		}
@@ -363,23 +398,25 @@ cid_t coroutine_add_with_stack_size(size_t stack_size, cr_fn_t fn, void *args) {
 	last_created = child;
 
 	coroutine_table[child-1].stack_size = stack_size;
-	coroutine_table[child-1].stack = aligned_alloc(CR_PLAT_STACK_ALIGNMENT, stack_size);
+	coroutine_table[child-1].stack =
+		aligned_alloc(CR_PLAT_STACK_ALIGNMENT, stack_size);
 #if CONFIG_COROUTINE_MEASURE_STACK || CONFIG_COROUTINE_PROTECT_STACK
 	for (size_t i = 0; i < stack_size; i++)
-		((uint8_t*)coroutine_table[child-1].stack)[i] = stack_pattern[i%sizeof(stack_pattern)];
+		((uint8_t *)coroutine_table[child-1].stack)[i] =
+			stack_pattern[i%sizeof(stack_pattern)];
 #endif
 
 	coroutine_running = child; 
 	coroutine_table[child-1].state = CR_INITIALIZING;
-	if (!setjmp(coroutine_add_env)) { /* point=a */
-		void *stack_base = coroutine_table[child-1].stack + (CR_PLAT_STACK_GROWS_DOWNWARD ? stack_size : 0);
-#if CONFIG_COROUTINE_PROTECT_STACK
-#  if CR_PLAT_STACK_GROWS_DOWNWARD
-		stack_base -= round_up(sizeof(stack_pattern), CR_PLAT_STACK_ALIGNMENT);
-#  else
-		stack_base += round_up(sizeof(stack_pattern), CR_PLAT_STACK_ALIGNMENT);
-#  endif
+	if (!cr_plat_setjmp(coroutine_add_env)) { /* point=a */
+		void *stack_base = coroutine_table[child-1].stack
+#if CR_PLAT_STACK_GROWS_DOWNWARD
+			+ stack_size
+			- STACK_GUARD_SIZE
+#else
+			+ STACK_GUARD_SIZE
 #endif
+			;
 		debugf("...stack     =%p\n", coroutine_table[child-1].stack);
 		debugf("...stack_base=%p\n", stack_base);
 		/* run until cr_begin() */
@@ -395,35 +432,12 @@ cid_t coroutine_add_with_stack_size(size_t stack_size, cr_fn_t fn, void *args) {
 }
 
 cid_t coroutine_add(cr_fn_t fn, void *args) {
-	return coroutine_add_with_stack_size(CONFIG_COROUTINE_DEFAULT_STACK_SIZE, fn, args);
+	return coroutine_add_with_stack_size(
+		CONFIG_COROUTINE_DEFAULT_STACK_SIZE, fn, args);
 }
 
 /* coroutine_main() ***********************************************************/
 
-#if CONFIG_COROUTINE_MEASURE_STACK
-struct stack_stats {
-	size_t cap;
-	size_t max;
-	//size_t cur;
-};
-
-static void measure_stack(cid_t cid, struct stack_stats *ret) {
-	ret->cap = coroutine_table[cid-1].stack_size - (CONFIG_COROUTINE_PROTECT_STACK ? 2*round_up(sizeof(stack_pattern), CR_PLAT_STACK_ALIGNMENT) : 0);
-
-	ret->max = ret->cap;
-	for (;;) {
-#if CR_PLAT_STACK_GROWS_DOWNWARD
-		size_t i = (CONFIG_COROUTINE_PROTECT_STACK ? round_up(sizeof(stack_pattern), CR_PLAT_STACK_ALIGNMENT) : 0) + ret->cap - ret->max;
-#else
-		size_t i = ret->max - 1 - (CONFIG_COROUTINE_PROTECT_STACK ? round_up(sizeof(stack_pattern), CR_PLAT_STACK_ALIGNMENT) : 0);
-#endif
-		if (ret->max == 0 || ((uint8_t*)coroutine_table[cid-1].stack)[i] != stack_pattern[i%sizeof(stack_pattern)])
-			break;
-		ret->max--;
-	}
-}
-#endif
-
 static inline cid_t next_coroutine() {
 	for (cid_t next = (coroutine_running%CONFIG_COROUTINE_NUM)+1;
 	     next != coroutine_running;
@@ -436,24 +450,23 @@ static inline cid_t next_coroutine() {
 
 void coroutine_main(void) {
 	debugf("coroutine_main()\n");
+	cr_plat_disable_interrupts();
 	coroutine_running = 0;
 	for (;;) {
 		cid_t next = next_coroutine();
 		if (!next) {
+			cr_plat_enable_interrupts();
 			errorf("no coroutines\n");
 			return;
 		}
-		if (!setjmp(coroutine_main_env)) { /* point=b */
+		if (!cr_plat_setjmp(coroutine_main_env)) { /* point=b */
 			coroutine_running = next;
 			coroutine_table[coroutine_running-1].state = CR_RUNNING;
-			longjmp(coroutine_table[coroutine_running-1].env, 1); /* jump to point=c */
+			cr_plat_longjmp(coroutine_table[coroutine_running-1].env, 1); /* jump to point=c */
 		}
+		/* This is where we jump to from cr_exit(), and from
+		 * nowhere else.  */
 		assert_cid_state(coroutine_running, state == CR_NONE);
-#if CONFIG_COROUTINE_MEASURE_STACK
-		struct stack_stats sizes;
-		measure_stack(coroutine_running, &sizes);
-		infof("cid=%zu: exited having used %zu B stack space\n", coroutine_running, sizes.max);
-#endif
 		free(coroutine_table[coroutine_running-1].stack);
 		coroutine_table[coroutine_running-1] = (struct coroutine){0};
 	}
@@ -462,81 +475,140 @@ void coroutine_main(void) {
 /* cr_*() *********************************************************************/
 
 void cr_begin(void) {
+	debugf("cid=%zu: cr_begin()\n", coroutine_running);
 	assert_cid_state(coroutine_running, state == CR_INITIALIZING);
 
 	coroutine_table[coroutine_running-1].state = CR_RUNNABLE;
-	if (!setjmp(coroutine_table[coroutine_running-1].env)) /* point=c1 */
-		longjmp(coroutine_add_env, 1); /* jump to point=a */
+	coroutine_table[coroutine_running-1].sp = cr_plat_get_sp();
+	if (!cr_plat_setjmp(coroutine_table[coroutine_running-1].env)) /* point=c1 */
+		cr_plat_longjmp(coroutine_add_env, 1); /* jump to point=a */
+	cr_plat_enable_interrupts();
 }
 
-static inline void _yield() {
+static inline void _cr_transition(enum coroutine_state state) {
+	coroutine_table[coroutine_running-1].state = state;
+
 	cid_t next;
 	for (;;) {
 		next = next_coroutine();
 		if (next)
+			/* Switch to `next`.  */
 			break;
 		if (coroutine_table[coroutine_running-1].state == CR_RUNNABLE) {
-			debugf("cid=%zu: no other runnable coroutines, not yielding\n", coroutine_running);
+			/* No other coroutine is runnable, don't
+			 * transition after all.  */
 			coroutine_table[coroutine_running-1].state = CR_RUNNING;
 			return;
-		} else { // XXX
-			debugf("cid=%zu: no runnable coroutines, sleeping\n", coroutine_running); // XXX
-			wait_for_interrupt(); // XXX
 		}
+		/* No coroutines are runnable, wait for an interrupt
+		 * to change that.  */
+		cr_plat_wait_for_interrupt();
 	}
-	if (!setjmp(coroutine_table[coroutine_running-1].env)) { /* point=c2 */
+
+	coroutine_table[coroutine_running-1].sp = cr_plat_get_sp();
+	if (!cr_plat_setjmp(coroutine_table[coroutine_running-1].env)) { /* point=c2 */
 		coroutine_running = next;
 		coroutine_table[coroutine_running-1].state = CR_RUNNING;
-		longjmp(coroutine_table[coroutine_running-1].env, 1); /* jump to point=b */
+		cr_plat_longjmp(coroutine_table[coroutine_running-1].env, 1); /* jump to point=b */
 	}
 }
 
 void cr_yield(void) {
+	debugf("cid=%zu: cr_yield()\n", coroutine_running);
 	assert_cid_state(coroutine_running, state == CR_RUNNING);
-	coroutine_table[coroutine_running-1].state = CR_RUNNABLE;
-	_yield();
+
+	cr_plat_disable_interrupts();
+	_cr_transition(CR_RUNNABLE);
+	cr_plat_enable_interrupts();
 }
 
 void cr_pause_and_yield(void) {
+	debugf("cid=%zu: cr_pause_and_yield()\n", coroutine_running);
 	assert_cid_state(coroutine_running, state == CR_RUNNING || state == CR_PRE_RUNNABLE);
-	coroutine_table[coroutine_running-1].state = CR_PAUSED;
-	_yield();
+
+	cr_plat_disable_interrupts();
+	if (coroutine_table[coroutine_running-1].state == CR_PRE_RUNNABLE)
+		_cr_transition(CR_RUNNABLE);
+	else
+		_cr_transition(CR_PAUSED);
+	cr_plat_enable_interrupts();
 }
 
 void cr_exit(void) {
+	debugf("cid=%zu: cr_exit()\n", coroutine_running);
 	assert_cid_state(coroutine_running, state == CR_RUNNING);
-	debugf("cid=%zu: exit\n", coroutine_running);
 
+	cr_plat_disable_interrupts();
 	coroutine_table[coroutine_running-1].state = CR_NONE;
-	longjmp(coroutine_main_env, 1); /* jump to point=b */
+	cr_plat_longjmp(coroutine_main_env, 1); /* jump to point=b */
 }
 
 void cr_unpause(cid_t cid) {
-	assert_cid_state(cid, state == CR_PAUSED);
 	debugf("cr_unpause(%zu)\n", cid);
+	assert_cid_state(cid, state == CR_PAUSED);
 
 	coroutine_table[cid-1].state = CR_RUNNABLE;
 }
 
 void cr_unpause_from_intrhandler(cid_t cid) {
-	assert_cid_state(cid, state == CR_RUNNING || state == CR_PAUSED);
 	debugf("cr_unpause_from_intrhandler(%zu)\n", cid);
+	assert_cid_state(cid, state == CR_RUNNING || state == CR_PAUSED);
 
-	switch (coroutine_table[cid-1].state) {
-	case CR_RUNNING:
+	if (coroutine_table[cid-1].state == CR_RUNNING) {
 		assert(cid == coroutine_running);
 		debugf("... raced, deferring unpause\n");
 		coroutine_table[cid-1].state = CR_PRE_RUNNABLE;
-		break;
-	case CR_PAUSED:
+	} else {
 		debugf("... actual unpause\n");
 		coroutine_table[cid-1].state = CR_RUNNABLE;
-		break;
-	default:
-		assert(false);
 	}
 }
 
 cid_t cr_getcid(void) {
 	return coroutine_running;
 }
+
+/* cr_cid_info() **************************************************************/
+
+#if CONFIG_COROUTINE_MEASURE_STACK
+
+void cr_cid_info(cid_t cid, struct cr_cid_info *ret) {
+	assert_cid(cid);
+	assert(ret);
+
+	/* stack_cap */
+	ret->stack_cap = coroutine_table[cid-1].stack_size - 2*STACK_GUARD_SIZE;
+
+	/* stack_max */
+	ret->stack_max = ret->stack_cap;
+	uint8_t *stack = (uint8_t *)coroutine_table[cid-1].stack;
+	for (;;) {
+		size_t i =
+#if CR_PLAT_STACK_GROWS_DOWNWARD
+			STACK_GUARD_SIZE + ret->stack_cap - ret->stack_max
+#else
+			ret->stack_max - 1 - STACK_GUARD_SIZE
+#endif
+			;
+		if (ret->stack_max == 0 ||
+		    stack[i] != stack_pattern[i%sizeof(stack_pattern)])
+			break;
+		ret->stack_max--;
+	}
+
+	/* stack_cur */
+	uintptr_t sp;
+	if (cid == coroutine_running)
+		sp = cr_plat_get_sp();
+	else
+		sp = coroutine_table[cid-1].sp;
+	assert(sp);
+	uintptr_t sb = (uintptr_t)coroutine_table[cid-1].stack;
+#if CR_PLAT_STACK_GROWS_DOWNWARD
+	ret->stack_cur = (sb - STACK_GUARD_SIZE) - sp;
+#else
+	ret->stack_cur = sp - (sb + STACK_GUARD_SIZE);
+#endif
+}
+
+#endif /* CONFIG_COROUTINE_MEASURE_STACK */
diff --git a/libcr/include/libcr/coroutine.h b/libcr/include/libcr/coroutine.h
index f580b09..0c5eac3 100644
--- a/libcr/include/libcr/coroutine.h
+++ b/libcr/include/libcr/coroutine.h
@@ -126,4 +126,16 @@ cid_t cr_getcid(void);
  * another as paused; a coroutine may only do those things to itself.
  */
 
+/* While the following are defined here unconditionally, the
+ * implementations are #if'd on CONFIG_COROUTINE_MEASURE_STACK.
+ */
+
+struct cr_cid_info {
+	size_t stack_cap;
+	size_t stack_max;
+	size_t stack_cur;
+};
+
+void cr_cid_info(cid_t cid, struct cr_cid_info *ret);
+
 #endif /* _COROUTINE_H_ */