libhsakmt: Set CWSR SVM range MADV_DONTFORK

fork process copy-on-write MMU nitifier on CWSR range will evict user
queues, and then update GPU mapping and resume queues, use MADV_DONTFORK
to avoid COW MMU notifier callback on CWSR SVM range.

Use mmap to alloc SVM range for CWSR because posix_memalign don't alloc
new range in child process, this fails to register svm range as range is
invalid address in forked child process.

Change-Id: Ibaea56a691dd6f577ed2e1f2d43f4a3500b8316f
Signed-off-by: Philip Yang <Philip.Yang@amd.com>

src/fmm.c

@@ -706,9 +706,8 @@ static void *reserved_aperture_allocate_aligned(manageable_aperture_t *app,
 	return start;
 }
 
-static void *mmap_allocate_aligned(int prot, int flags, uint64_t size,
-				   uint64_t align,  uint64_t guard_size,
-				   void *aper_base, void *aper_limit)
+void *mmap_allocate_aligned(int prot, int flags, uint64_t size, uint64_t align,
+			    uint64_t guard_size, void *aper_base, void *aper_limit)
 {
 	void *addr, *aligned_addr, *aligned_end, *mapping_end;
 	uint64_t aligned_padded_size;

src/fmm.h

@@ -90,4 +90,7 @@ HSAKMT_STATUS fmm_map_to_gpu_nodes(void *address, uint64_t size,
 		uint32_t *nodes_to_map, uint64_t num_of_nodes, uint64_t *gpuvm_address);
 
 int open_drm_render_device(int minor);
+void *mmap_allocate_aligned(int prot, int flags, uint64_t size, uint64_t align,
+			    uint64_t guard_size, void *aper_base, void *aper_limit);
+
 #endif /* FMM_H_ */

src/queues.c

@@ -425,7 +425,8 @@ static void free_queue(struct queue *q)
 					 q->eop_buffer_size,
 					 PAGE_SIZE, q->use_ats);
 	if (q->unified_ctx_save_restore)
-		free(q->ctx_save_restore);
+		munmap(q->ctx_save_restore,
+		       PAGE_ALIGN_UP(q->ctx_save_restore_size + q->debug_memory_size));
 	else if (q->ctx_save_restore)
 		free_exec_aligned_memory(q->ctx_save_restore,
 					 q->ctx_save_restore_size + q->debug_memory_size,
@@ -500,9 +501,21 @@ static int handle_concrete_asic(struct queue *q,
 			void *addr;
 			HSAKMT_STATUS r = HSAKMT_STATUS_ERROR;
 
-			if (posix_memalign(&addr, GPU_HUGE_PAGE_SIZE, size))
-				pr_err("[%s] posix_memalign failed:\n", __func__);
-			else {
+			addr = mmap_allocate_aligned(PROT_READ | PROT_WRITE,
+						     MAP_ANONYMOUS | MAP_PRIVATE,
+						     size, GPU_HUGE_PAGE_SIZE, 0,
+						     0, (void *)LONG_MAX);
+			if (!addr) {
+				pr_err("mmap failed to alloc ctx area size 0x%x: %s\n",
+					size, strerror(errno));
+			} else {
+				/*
+				 * To avoid fork child process COW MMU notifier
+				 * callback evict parent process queues.
+				 */
+				if (madvise(addr, size, MADV_DONTFORK))
+					pr_err("madvise failed -%d\n", errno);
+
 				fill_cwsr_header(q, addr, Event, ErrPayload);
 
 				r = register_svm_range(addr, size,
@@ -511,8 +524,9 @@ static int handle_concrete_asic(struct queue *q,
 				if (r == HSAKMT_STATUS_SUCCESS) {
 					q->ctx_save_restore = addr;
 					q->unified_ctx_save_restore = true;
-				} else
-					free(addr);
+				} else {
+					munmap(addr, size);
+				}
 			}
 		}