lima: implement BO cache

Allocating BOs is expensive, so we should avoid doing that by caching
freed BOs.

BO cache is modelled after one in v3d driver and works as follows:

- in lima_bo_create() check if we have matching BO in cache and return
  it if there's one, allocate new BO otherwise.
- in lima_bo_unreference() (renamed from lima_bo_free()): put BO in
  cache instead of freeing it and remove all stale BOs from cache

Reviewed-by: Qiang Yu <yuq825@gmail.com>
Signed-off-by: Vasily Khoruzhick <anarsoul@gmail.com>

8 files changed, 212 insertions, 30 deletions

diff --git a/src/gallium/drivers/lima/lima_bo.c b/src/gallium/drivers/lima/lima_bo.c
index b082d32afd8..22263ea0054 100644
--- a/src/gallium/drivers/lima/lima_bo.c
+++ b/src/gallium/drivers/lima/lima_bo.c
@@ -30,6 +30,7 @@
 #include "drm-uapi/lima_drm.h"
 
 #include "util/u_hash_table.h"
+#include "util/u_math.h"
 #include "util/os_time.h"
 #include "os/os_mman.h"
 
@@ -37,6 +38,7 @@
 
 #include "lima_screen.h"
 #include "lima_bo.h"
+#include "lima_util.h"
 
 #define PTR_TO_UINT(x) ((unsigned)((intptr_t)(x)))
 
@@ -68,6 +70,16 @@ err_out0:
    return false;
 }
 
+bool lima_bo_cache_init(struct lima_screen *screen)
+{
+   mtx_init(&screen->bo_cache_lock, mtx_plain);
+   list_inithead(&screen->bo_cache_time);
+   for (int i = 0; i < NR_BO_CACHE_BUCKETS; i++)
+      list_inithead(&screen->bo_cache_buckets[i]);
+
+   return true;
+}
+
 void lima_bo_table_fini(struct lima_screen *screen)
 {
    mtx_destroy(&screen->bo_table_lock);
@@ -75,6 +87,13 @@ void lima_bo_table_fini(struct lima_screen *screen)
    util_hash_table_destroy(screen->bo_flink_names);
 }
 
+static void
+lima_bo_cache_remove(struct lima_bo *bo)
+{
+   list_del(&bo->size_list);
+   list_del(&bo->time_list);
+}
+
 static void lima_close_kms_handle(struct lima_screen *screen, uint32_t handle)
 {
    struct drm_gem_close args = {
@@ -84,6 +103,36 @@ static void lima_close_kms_handle(struct lima_screen *screen, uint32_t handle)
    drmIoctl(screen->fd, DRM_IOCTL_GEM_CLOSE, &args);
 }
 
+static void
+lima_bo_free(struct lima_bo *bo)
+{
+   struct lima_screen *screen = bo->screen;
+   mtx_lock(&screen->bo_table_lock);
+   util_hash_table_remove(screen->bo_handles,
+                          (void *)(uintptr_t)bo->handle);
+   if (bo->flink_name)
+      util_hash_table_remove(screen->bo_flink_names,
+                             (void *)(uintptr_t)bo->flink_name);
+   mtx_unlock(&screen->bo_table_lock);
+
+   if (bo->map)
+      lima_bo_unmap(bo);
+
+   lima_close_kms_handle(screen, bo->handle);
+   free(bo);
+}
+
+void lima_bo_cache_fini(struct lima_screen *screen)
+{
+   mtx_destroy(&screen->bo_cache_lock);
+
+   list_for_each_entry_safe(struct lima_bo, entry,
+                            &screen->bo_cache_time, time_list) {
+      lima_bo_cache_remove(entry);
+      lima_bo_free(entry);
+   }
+}
+
 static bool lima_bo_get_info(struct lima_bo *bo)
 {
    struct drm_lima_gem_info req = {
@@ -98,10 +147,112 @@ static bool lima_bo_get_info(struct lima_bo *bo)
    return true;
 }
 
+static unsigned
+lima_bucket_index(unsigned size)
+{
+   /* Round down to POT to compute a bucket index */
+
+   unsigned bucket_index = util_logbase2(size);
+
+   /* Clamp the bucket index; all huge allocations will be
+    * sorted into the largest bucket */
+   bucket_index = CLAMP(bucket_index, MIN_BO_CACHE_BUCKET,
+                        MAX_BO_CACHE_BUCKET);
+
+   /* Reindex from 0 */
+   return (bucket_index - MIN_BO_CACHE_BUCKET);
+}
+
+static struct list_head *
+lima_bo_cache_get_bucket(struct lima_screen *screen, unsigned size)
+{
+   return &screen->bo_cache_buckets[lima_bucket_index(size)];
+}
+
+static void
+lima_bo_cache_free_stale_bos(struct lima_screen *screen, time_t time)
+{
+   list_for_each_entry_safe(struct lima_bo, entry,
+                            &screen->bo_cache_time, time_list) {
+      /* Free BOs that are sitting idle for longer than 5 seconds */
+      if (time - entry->free_time > 6) {
+         lima_bo_cache_remove(entry);
+         lima_bo_free(entry);
+      } else
+         break;
+   }
+}
+
+static bool
+lima_bo_cache_put(struct lima_bo *bo)
+{
+   if (!bo->cacheable)
+      return false;
+
+   struct lima_screen *screen = bo->screen;
+
+   mtx_lock(&screen->bo_cache_lock);
+   struct list_head *bucket = lima_bo_cache_get_bucket(screen, bo->size);
+
+   if (!bucket) {
+      mtx_unlock(&screen->bo_cache_lock);
+      return false;
+   }
+
+   struct timespec time;
+   clock_gettime(CLOCK_MONOTONIC, &time);
+   bo->free_time = time.tv_sec;
+   list_addtail(&bo->size_list, bucket);
+   list_addtail(&bo->time_list, &screen->bo_cache_time);
+   lima_bo_cache_free_stale_bos(screen, time.tv_sec);
+   mtx_unlock(&screen->bo_cache_lock);
+
+   return true;
+}
+
+static struct lima_bo *
+lima_bo_cache_get(struct lima_screen *screen, uint32_t size, uint32_t flags)
+{
+   struct lima_bo *bo = NULL;
+   mtx_lock(&screen->bo_cache_lock);
+   struct list_head *bucket = lima_bo_cache_get_bucket(screen, size);
+
+   if (!bucket) {
+      mtx_unlock(&screen->bo_cache_lock);
+      return false;
+   }
+
+   list_for_each_entry_safe(struct lima_bo, entry, bucket, size_list) {
+      if (entry->size >= size &&
+          entry->flags == flags) {
+         /* Check if BO is idle. If it's not it's better to allocate new one */
+         if (!lima_bo_wait(entry, LIMA_GEM_WAIT_WRITE, 0))
+            break;
+
+         lima_bo_cache_remove(entry);
+         p_atomic_set(&entry->refcnt, 1);
+         bo = entry;
+         break;
+      }
+   }
+
+   mtx_unlock(&screen->bo_cache_lock);
+
+   return bo;
+}
+
 struct lima_bo *lima_bo_create(struct lima_screen *screen,
                                uint32_t size, uint32_t flags)
 {
    struct lima_bo *bo;
+
+   /* Try to get bo from cache first */
+   bo = lima_bo_cache_get(screen, size, flags);
+   if (bo)
+      return bo;
+
+   size = align(size, LIMA_PAGE_SIZE);
+
    struct drm_lima_gem_create req = {
       .size = size,
       .flags = flags,
@@ -110,12 +261,17 @@ struct lima_bo *lima_bo_create(struct lima_screen *screen,
    if (!(bo = calloc(1, sizeof(*bo))))
       return NULL;
 
+   list_inithead(&bo->time_list);
+   list_inithead(&bo->size_list);
+
    if (drmIoctl(screen->fd, DRM_IOCTL_LIMA_GEM_CREATE, &req))
       goto err_out0;
 
    bo->screen = screen;
    bo->size = req.size;
+   bo->flags = req.flags;
    bo->handle = req.handle;
+   bo->cacheable = !(lima_debug & LIMA_DEBUG_NO_BO_CACHE);
    p_atomic_set(&bo->refcnt, 1);
 
    if (!lima_bo_get_info(bo))
@@ -130,25 +286,16 @@ err_out0:
    return NULL;
 }
 
-void lima_bo_free(struct lima_bo *bo)
+void lima_bo_unreference(struct lima_bo *bo)
 {
    if (!p_atomic_dec_zero(&bo->refcnt))
       return;
 
-   struct lima_screen *screen = bo->screen;
-   mtx_lock(&screen->bo_table_lock);
-   util_hash_table_remove(screen->bo_handles,
-                          (void *)(uintptr_t)bo->handle);
-   if (bo->flink_name)
-      util_hash_table_remove(screen->bo_flink_names,
-                             (void *)(uintptr_t)bo->flink_name);
-   mtx_unlock(&screen->bo_table_lock);
-
-   if (bo->map)
-      lima_bo_unmap(bo);
+   /* Try to put it into cache */
+   if (lima_bo_cache_put(bo))
+      return;
 
-   lima_close_kms_handle(screen, bo->handle);
-   free(bo);
+   lima_bo_free(bo);
 }
 
 void *lima_bo_map(struct lima_bo *bo)
@@ -175,6 +322,9 @@ bool lima_bo_export(struct lima_bo *bo, struct winsys_handle *handle)
 {
    struct lima_screen *screen = bo->screen;
 
+   /* Don't cache exported BOs */
+   bo->cacheable = false;
+
    switch (handle->type) {
    case WINSYS_HANDLE_TYPE_SHARED:
       if (!bo->flink_name) {
@@ -271,6 +421,8 @@ struct lima_bo *lima_bo_import(struct lima_screen *screen,
 
    if (bo) {
       p_atomic_inc(&bo->refcnt);
+      /* Don't cache imported BOs */
+      bo->cacheable = false;
       mtx_unlock(&screen->bo_table_lock);
       return bo;
    }
@@ -282,6 +434,10 @@ struct lima_bo *lima_bo_import(struct lima_screen *screen,
       return NULL;
    }
 
+   /* Don't cache imported BOs */
+   bo->cacheable = false;
+   list_inithead(&bo->time_list);
+   list_inithead(&bo->size_list);
    bo->screen = screen;
    p_atomic_set(&bo->refcnt, 1);
 
diff --git a/src/gallium/drivers/lima/lima_bo.h b/src/gallium/drivers/lima/lima_bo.h
index 3f440b3b014..7bbd0063602 100644
--- a/src/gallium/drivers/lima/lima_bo.h
+++ b/src/gallium/drivers/lima/lima_bo.h
@@ -28,12 +28,18 @@
 #include <stdint.h>
 
 #include "util/u_atomic.h"
+#include "util/list.h"
 
 struct lima_bo {
    struct lima_screen *screen;
+   struct list_head time_list;
+   struct list_head size_list;
    int refcnt;
+   bool cacheable;
+   time_t free_time;
 
    uint32_t size;
+   uint32_t flags;
    uint32_t handle;
    uint64_t offset;
    uint32_t flink_name;
@@ -44,10 +50,12 @@ struct lima_bo {
 
 bool lima_bo_table_init(struct lima_screen *screen);
 void lima_bo_table_fini(struct lima_screen *screen);
+bool lima_bo_cache_init(struct lima_screen *screen);
+void lima_bo_cache_fini(struct lima_screen *screen);
 
 struct lima_bo *lima_bo_create(struct lima_screen *screen, uint32_t size,
                                uint32_t flags);
-void lima_bo_free(struct lima_bo *bo);
+void lima_bo_unreference(struct lima_bo *bo);
 
 static inline void lima_bo_reference(struct lima_bo *bo)
 {
diff --git a/src/gallium/drivers/lima/lima_context.c b/src/gallium/drivers/lima/lima_context.c
index 3c3887001bf..813e87361bb 100644
--- a/src/gallium/drivers/lima/lima_context.c
+++ b/src/gallium/drivers/lima/lima_context.c
@@ -138,13 +138,13 @@ lima_context_destroy(struct pipe_context *pctx)
 
    for (int i = 0; i < LIMA_CTX_PLB_MAX_NUM; i++) {
       if (ctx->plb[i])
-         lima_bo_free(ctx->plb[i]);
+         lima_bo_unreference(ctx->plb[i]);
       if (ctx->gp_tile_heap[i])
-         lima_bo_free(ctx->gp_tile_heap[i]);
+         lima_bo_unreference(ctx->gp_tile_heap[i]);
    }
 
    if (ctx->plb_gp_stream)
-      lima_bo_free(ctx->plb_gp_stream);
+      lima_bo_unreference(ctx->plb_gp_stream);
 
    if (ctx->plb_pp_stream)
       assert(!_mesa_hash_table_num_entries(ctx->plb_pp_stream));
diff --git a/src/gallium/drivers/lima/lima_program.c b/src/gallium/drivers/lima/lima_program.c
index e3863feb80e..5e639748bfc 100644
--- a/src/gallium/drivers/lima/lima_program.c
+++ b/src/gallium/drivers/lima/lima_program.c
@@ -302,7 +302,7 @@ lima_delete_fs_state(struct pipe_context *pctx, void *hwcso)
    struct lima_fs_shader_state *so = hwcso;
 
    if (so->bo)
-      lima_bo_free(so->bo);
+      lima_bo_unreference(so->bo);
 
    ralloc_free(so);
 }
@@ -396,7 +396,7 @@ lima_delete_vs_state(struct pipe_context *pctx, void *hwcso)
    struct lima_vs_shader_state *so = hwcso;
 
    if (so->bo)
-      lima_bo_free(so->bo);
+      lima_bo_unreference(so->bo);
 
    ralloc_free(so);
 }
diff --git a/src/gallium/drivers/lima/lima_resource.c b/src/gallium/drivers/lima/lima_resource.c
index faa129998c3..e01e60c0465 100644
--- a/src/gallium/drivers/lima/lima_resource.c
+++ b/src/gallium/drivers/lima/lima_resource.c
@@ -259,7 +259,7 @@ lima_resource_destroy(struct pipe_screen *pscreen, struct pipe_resource *pres)
    struct lima_resource *res = lima_resource(pres);
 
    if (res->bo)
-      lima_bo_free(res->bo);
+      lima_bo_unreference(res->bo);
 
    if (res->scanout)
       renderonly_scanout_destroy(res->scanout, screen->ro);
@@ -528,7 +528,7 @@ lima_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf)
          struct lima_ctx_plb_pp_stream *s = entry->data;
          if (--s->refcnt == 0) {
             if (s->bo)
-               lima_bo_free(s->bo);
+               lima_bo_unreference(s->bo);
             _mesa_hash_table_remove(ctx->plb_pp_stream, entry);
             ralloc_free(s);
          }
diff --git a/src/gallium/drivers/lima/lima_screen.c b/src/gallium/drivers/lima/lima_screen.c
index 5e6ac1ffb08..3c1288c897f 100644
--- a/src/gallium/drivers/lima/lima_screen.c
+++ b/src/gallium/drivers/lima/lima_screen.c
@@ -61,8 +61,9 @@ lima_screen_destroy(struct pipe_screen *pscreen)
       free(screen->ro);
 
    if (screen->pp_buffer)
-      lima_bo_free(screen->pp_buffer);
+      lima_bo_unreference(screen->pp_buffer);
 
+   lima_bo_cache_fini(screen);
    lima_bo_table_fini(screen);
    ralloc_free(screen);
 }
@@ -418,6 +419,8 @@ static const struct debug_named_value debug_options[] = {
           "dump GPU command stream to $PWD/lima.dump" },
         { "shaderdb", LIMA_DEBUG_SHADERDB,
           "print shader information for shaderdb" },
+        { "nobocache", LIMA_DEBUG_NO_BO_CACHE,
+          "disable BO cache" },
         { NULL }
 };
 
@@ -478,16 +481,20 @@ lima_screen_create(int fd, struct renderonly *ro)
    if (!lima_screen_query_info(screen))
       goto err_out0;
 
-   if (!lima_bo_table_init(screen))
+   if (!lima_bo_cache_init(screen))
       goto err_out0;
 
+   if (!lima_bo_table_init(screen))
+      goto err_out1;
+
    screen->pp_ra = ppir_regalloc_init(screen);
    if (!screen->pp_ra)
-      goto err_out1;
+      goto err_out2;
 
    screen->pp_buffer = lima_bo_create(screen, pp_buffer_size, 0);
    if (!screen->pp_buffer)
-      goto err_out1;
+      goto err_out2;
+   screen->pp_buffer->cacheable = false;
 
    /* fs program for clear buffer?
     * const0 1 0 0 -1.67773, mov.v0 $0 ^const0.xxxx, stop
@@ -534,7 +541,7 @@ lima_screen_create(int fd, struct renderonly *ro)
       screen->ro = renderonly_dup(ro);
       if (!screen->ro) {
          fprintf(stderr, "Failed to dup renderonly object\n");
-         goto err_out2;
+         goto err_out3;
       }
    }
 
@@ -559,10 +566,12 @@ lima_screen_create(int fd, struct renderonly *ro)
 
    return &screen->base;
 
+err_out3:
+   lima_bo_unreference(screen->pp_buffer);
 err_out2:
-   lima_bo_free(screen->pp_buffer);
-err_out1:
    lima_bo_table_fini(screen);
+err_out1:
+   lima_bo_cache_fini(screen);
 err_out0:
    ralloc_free(screen);
    return NULL;
diff --git a/src/gallium/drivers/lima/lima_screen.h b/src/gallium/drivers/lima/lima_screen.h
index 547d083ecd0..62fa480738c 100644
--- a/src/gallium/drivers/lima/lima_screen.h
+++ b/src/gallium/drivers/lima/lima_screen.h
@@ -37,6 +37,7 @@
 #define LIMA_DEBUG_PP      (1 << 1)
 #define LIMA_DEBUG_DUMP    (1 << 2)
 #define LIMA_DEBUG_SHADERDB (1 << 3)
+#define LIMA_DEBUG_NO_BO_CACHE (1 << 4)
 
 extern uint32_t lima_debug;
 extern FILE *lima_dump_command_stream;
@@ -46,6 +47,11 @@ extern int lima_ppir_force_spilling;
 
 struct ra_regs;
 
+#define MIN_BO_CACHE_BUCKET (12) /* 2^12 = 4KB */
+#define MAX_BO_CACHE_BUCKET (22) /* 2^22 = 4MB */
+
+#define NR_BO_CACHE_BUCKETS (MAX_BO_CACHE_BUCKET - MIN_BO_CACHE_BUCKET + 1)
+
 struct lima_screen {
    struct pipe_screen base;
    struct renderonly *ro;
@@ -60,8 +66,11 @@ struct lima_screen {
 
    /* bo table */
    mtx_t bo_table_lock;
+   mtx_t bo_cache_lock;
    struct util_hash_table *bo_handles;
    struct util_hash_table *bo_flink_names;
+   struct list_head bo_cache_buckets[NR_BO_CACHE_BUCKETS];
+   struct list_head bo_cache_time;
 
    struct slab_parent_pool transfer_pool;
 
diff --git a/src/gallium/drivers/lima/lima_submit.c b/src/gallium/drivers/lima/lima_submit.c
index 3977af8078e..acc73d08054 100644
--- a/src/gallium/drivers/lima/lima_submit.c
+++ b/src/gallium/drivers/lima/lima_submit.c
@@ -145,7 +145,7 @@ bool lima_submit_start(struct lima_submit *submit, void *frame, uint32_t size)
    bool ret = drmIoctl(submit->screen->fd, DRM_IOCTL_LIMA_GEM_SUBMIT, &req) == 0;
 
    util_dynarray_foreach(&submit->bos, struct lima_bo *, bo) {
-      lima_bo_free(*bo);
+      lima_bo_unreference(*bo);
    }
 
    util_dynarray_clear(&submit->gem_bos);