Skip to contentSkip to conversation

radeon: preallocate memory for command stream parsing

PATCH In review Jerome Glisse 4 participants 6 messages Jun 23, 2009 – Jun 25, 2009

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 radeon: preallocate memory for command stream parsing

Command stream parsing is the most common operation and can happen hundred of times per second, we don't want to allocate/free memory each time this ioctl is call. This rework the ioctl to avoid doing so by allocating temporary memory along the ib pool.

---
 drivers/gpu/drm/radeon/r100.c        |   27 +++---
 drivers/gpu/drm/radeon/r300.c        |    6 +-
 drivers/gpu/drm/radeon/radeon.h      |  110 ++++++++++----------
 drivers/gpu/drm/radeon/radeon_cs.c   |  186 +++++++++++++++-------------------
 drivers/gpu/drm/radeon/radeon_ring.c |   23 ++++
 5 files changed, 177 insertions(+), 175 deletions(-)
drivers/gpu/drm/radeon/r100.c +13 -14
index c550932..56ef9f7 100644
@@ -699,7 +699,7 @@ void r100_cs_dump_packet(struct radeon_cs_parser *p,
 	unsigned idx;
 
 	ib = p->ib->ptr;
-	ib_chunk = &p->chunks[p->chunk_ib_idx];
+	ib_chunk = &p->ibc;
 	idx = pkt->idx;
 	for (i = 0; i <= (pkt->count + 1); i++, idx++) {
 		DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
@@ -718,14 +718,15 @@ int r100_cs_packet_parse(struct radeon_cs_parser *p,
 			 struct radeon_cs_packet *pkt,
 			 unsigned idx)
 {
-	struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
-	uint32_t header = ib_chunk->kdata[idx];
+	struct radeon_cs_chunk *ib_chunk = &p->ibc;
+	uint32_t header;
 
 	if (idx >= ib_chunk->length_dw) {
 		DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
 			  idx, ib_chunk->length_dw);
 		return -EINVAL;
 	}
+	header = ib_chunk->kdata[idx];
 	pkt->idx = idx;
 	pkt->type = CP_PACKET_GET_TYPE(header);
 	pkt->count = CP_PACKET_GET_COUNT(header);
@@ -767,18 +768,16 @@ int r100_cs_packet_next_reloc(struct radeon_cs_parser *p,
 			      struct radeon_cs_reloc **cs_reloc)
 {
 	struct radeon_cs_chunk *ib_chunk;
-	struct radeon_cs_chunk *relocs_chunk;
 	struct radeon_cs_packet p3reloc;
 	unsigned idx;
 	int r;
 
-	if (p->chunk_relocs_idx == -1) {
+	if (p->relocs == NULL) {
 		DRM_ERROR("No relocation chunk !\n");
 		return -EINVAL;
 	}
 	*cs_reloc = NULL;
-	ib_chunk = &p->chunks[p->chunk_ib_idx];
-	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
+	ib_chunk = &p->ibc;
 	r = r100_cs_packet_parse(p, &p3reloc, p->idx);
 	if (r) {
 		return r;
@@ -791,14 +790,14 @@ int r100_cs_packet_next_reloc(struct radeon_cs_parser *p,
 		return -EINVAL;
 	}
 	idx = ib_chunk->kdata[p3reloc.idx + 1];
-	if (idx >= relocs_chunk->length_dw) {
+	if ((idx / 4) >= p->nrelocs) {
 		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
-			  idx, relocs_chunk->length_dw);
+			  idx, p->nrelocs * 4);
 		r100_cs_dump_packet(p, &p3reloc);
 		return -EINVAL;
 	}
 	/* FIXME: we assume reloc size is 4 dwords */
-	*cs_reloc = p->relocs_ptr[(idx / 4)];
+	*cs_reloc = p->relocs_ptr[idx / 4];
 	return 0;
 }
 
@@ -816,7 +815,7 @@ static int r100_packet0_check(struct radeon_cs_parser *p,
 	int r;
 
 	ib = p->ib->ptr;
-	ib_chunk = &p->chunks[p->chunk_ib_idx];
+	ib_chunk = &p->ibc;
 	idx = pkt->idx + 1;
 	reg = pkt->reg;
 	onereg = false;
@@ -897,7 +896,7 @@ int r100_cs_track_check_pkt3_indx_buffer(struct radeon_cs_parser *p,
 	struct radeon_cs_chunk *ib_chunk;
 	unsigned idx;
 
-	ib_chunk = &p->chunks[p->chunk_ib_idx];
+	ib_chunk = &p->ibc;
 	idx = pkt->idx + 1;
 	if ((ib_chunk->kdata[idx+2] + 1) > radeon_object_size(robj)) {
 		DRM_ERROR("[drm] Buffer too small for PACKET3 INDX_BUFFER "
@@ -920,7 +919,7 @@ static int r100_packet3_check(struct radeon_cs_parser *p,
 	int r;
 
 	ib = p->ib->ptr;
-	ib_chunk = &p->chunks[p->chunk_ib_idx];
+	ib_chunk = &p->ibc;
 	idx = pkt->idx + 1;
 	switch (pkt->opcode) {
 	case PACKET3_3D_LOAD_VBPNTR:
@@ -1027,7 +1026,7 @@ int r100_cs_parse(struct radeon_cs_parser *p)
 		if (r) {
 			return r;
 		}
-	} while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
+	} while (p->idx < p->ibc.length_dw);
 	return 0;
 }
 
drivers/gpu/drm/radeon/r300.c +3 -3
index e2ed5bc..2b2199e 100644
@@ -1024,7 +1024,7 @@ static int r300_packet0_check(struct radeon_cs_parser *p,
 	int r;
 
 	ib = p->ib->ptr;
-	ib_chunk = &p->chunks[p->chunk_ib_idx];
+	ib_chunk = &p->ibc;
 	track = (struct r300_cs_track*)p->track;
 	switch(reg) {
 	case RADEON_DST_PITCH_OFFSET:
@@ -1361,7 +1361,7 @@ static int r300_packet3_check(struct radeon_cs_parser *p,
 	int r;
 
 	ib = p->ib->ptr;
-	ib_chunk = &p->chunks[p->chunk_ib_idx];
+	ib_chunk = &p->ibc;
 	idx = pkt->idx + 1;
 	track = (struct r300_cs_track*)p->track;
 	switch(pkt->opcode) {
@@ -1520,7 +1520,7 @@ int r300_cs_parse(struct radeon_cs_parser *p)
 		if (r) {
 			return r;
 		}
-	} while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
+	} while (p->idx < p->ibc.length_dw);
 	return 0;
 }
 
drivers/gpu/drm/radeon/radeon.h +56 -54

134 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_cs.c +82 -104

264 lines not on this page · raw message

drivers/gpu/drm/radeon/radeon_ring.c +23 -0
index a853261..0026a11 100644
@@ -177,6 +177,20 @@ int radeon_ib_pool_init(struct radeon_device *rdev)
 
 	/* Allocate 1M object buffer */
 	INIT_LIST_HEAD(&rdev->ib_pool.scheduled_ibs);
+	rdev->ib_pool.ib = drm_calloc_large(RADEON_IB_POOL_SIZE, 64 * 1024);
+	if (rdev->ib_pool.ib == NULL) {
+		return -ENOMEM;
+	}
+	rdev->ib_pool.relocs = drm_calloc_large(RADEON_IB_POOL_SIZE * 1024,
+						sizeof(struct radeon_cs_reloc));
+	if (rdev->ib_pool.relocs == NULL) {
+		return -ENOMEM;
+	}
+	rdev->ib_pool.relocs_ptr = drm_calloc_large(RADEON_IB_POOL_SIZE * 1024,
+						    sizeof(void *));
+	if (rdev->ib_pool.relocs == NULL) {
+		return -ENOMEM;
+	}
 	r = radeon_object_create(rdev, NULL,  RADEON_IB_POOL_SIZE*64*1024,
 				 true, RADEON_GEM_DOMAIN_GTT,
 				 false, &rdev->ib_pool.robj);
@@ -202,6 +216,9 @@ int radeon_ib_pool_init(struct radeon_device *rdev)
 		rdev->ib_pool.ibs[i].ptr = ptr + offset;
 		rdev->ib_pool.ibs[i].idx = i;
 		rdev->ib_pool.ibs[i].length_dw = 0;
+		rdev->ib_pool.ibs[i].ib = &rdev->ib_pool.ib[i * 16 * 1024];
+		rdev->ib_pool.ibs[i].relocs = &rdev->ib_pool.relocs[i * 1024];
+		rdev->ib_pool.ibs[i].relocs_ptr = &rdev->ib_pool.relocs_ptr[i * 1024];
 		INIT_LIST_HEAD(&rdev->ib_pool.ibs[i].list);
 	}
 	bitmap_zero(rdev->ib_pool.alloc_bm, RADEON_IB_POOL_SIZE);
@@ -225,6 +242,12 @@ void radeon_ib_pool_fini(struct radeon_device *rdev)
 		radeon_object_unref(&rdev->ib_pool.robj);
 		rdev->ib_pool.robj = NULL;
 	}
+	if (rdev->ib_pool.ib)
+		drm_free_large(rdev->ib_pool.ib);
+	if (rdev->ib_pool.relocs)
+		drm_free_large(rdev->ib_pool.relocs);
+	if (rdev->ib_pool.relocs_ptr)
+		drm_free_large(rdev->ib_pool.relocs_ptr);
 	mutex_unlock(&rdev->ib_pool.mutex);
 }
 
Signature · 1 line
1.6.2.2
#1 lore raw

Hi Jerome,

Quoted 7 lines · Jerome Glisse

Command stream parsing is the most common operation and can happen hundred of times per second, we don't want to allocate/free memory each time this ioctl is call. This rework the ioctl to avoid doing so by allocating temporary memory along the ib pool.

So how much does this help (i.e. where are the numbers)? I am bit surprised "hundred of times per second" is an issue for our slab allocators. Hmm?

Quoted 616 lines

---
 drivers/gpu/drm/radeon/r100.c        |   27 +++--- drivers/gpu/drm/radeon/r300.c        |    6 +- drivers/gpu/drm/radeon/radeon.h      |  110 ++++++++++---------- drivers/gpu/drm/radeon/radeon_cs.c   |  186 +++++++++++++++------------------- drivers/gpu/drm/radeon/radeon_ring.c |   23 ++++

 5 files changed, 177 insertions(+), 175 deletions(-)
drivers/gpu/drm/radeon/r100.c +11 -12
index c550932..56ef9f7 100644
@@ -699,7 +699,7 @@ void r100_cs_dump_packet(struct radeon_cs_parser *p,
        unsigned idx;
 
        ib = p->ib->ptr;
-       ib_chunk = &p->chunks[p->chunk_ib_idx];
+       ib_chunk = &p->ibc;
        idx = pkt->idx;
        for (i = 0; i <= (pkt->count + 1); i++, idx++) {
                DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
@@ -718,14 +718,15 @@ int r100_cs_packet_parse(struct radeon_cs_parser *p,
                         struct radeon_cs_packet *pkt,
                         unsigned idx)
  {
-       struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
-       uint32_t header = ib_chunk->kdata[idx];
+       struct radeon_cs_chunk *ib_chunk = &p->ibc;
+       uint32_t header;
 
        if (idx >= ib_chunk->length_dw) {
                DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
                          idx, ib_chunk->length_dw);
                return -EINVAL;
        }
+       header = ib_chunk->kdata[idx];
        pkt->idx = idx;
        pkt->type = CP_PACKET_GET_TYPE(header);
        pkt->count = CP_PACKET_GET_COUNT(header);
@@ -767,18 +768,16 @@ int r100_cs_packet_next_reloc(struct radeon_cs_parser *p,
                              struct radeon_cs_reloc **cs_reloc)
  {
        struct radeon_cs_chunk *ib_chunk;
-       struct radeon_cs_chunk *relocs_chunk;
        struct radeon_cs_packet p3reloc;
        unsigned idx;
        int r;
 
-       if (p->chunk_relocs_idx == -1) {
+       if (p->relocs == NULL) {
                DRM_ERROR("No relocation chunk !\n");
                return -EINVAL;
        }
        *cs_reloc = NULL;
-       ib_chunk = &p->chunks[p->chunk_ib_idx];
-       relocs_chunk = &p->chunks[p->chunk_relocs_idx];
+       ib_chunk = &p->ibc;
        r = r100_cs_packet_parse(p, &p3reloc, p->idx);
        if (r) {
                return r;
@@ -791,14 +790,14 @@ int r100_cs_packet_next_reloc(struct radeon_cs_parser *p,
                return -EINVAL;
        }
        idx = ib_chunk->kdata[p3reloc.idx + 1];
-       if (idx >= relocs_chunk->length_dw) {
+       if ((idx / 4) >= p->nrelocs) {
                DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
-                         idx, relocs_chunk->length_dw);
+                         idx, p->nrelocs * 4);
                r100_cs_dump_packet(p, &p3reloc);
                return -EINVAL;
        }
        /* FIXME: we assume reloc size is 4 dwords */
-       *cs_reloc = p->relocs_ptr[(idx / 4)];
+       *cs_reloc = p->relocs_ptr[idx / 4];
        return 0;
  }
 
@@ -816,7 +815,7 @@ static int r100_packet0_check(struct radeon_cs_parser *p,
        int r;
 
        ib = p->ib->ptr;
-       ib_chunk = &p->chunks[p->chunk_ib_idx];
+       ib_chunk = &p->ibc;
        idx = pkt->idx + 1;
        reg = pkt->reg;
        onereg = false;
@@ -897,7 +896,7 @@ int r100_cs_track_check_pkt3_indx_buffer(struct radeon_cs_parser *p,
        struct radeon_cs_chunk *ib_chunk;
        unsigned idx;
 
-       ib_chunk = &p->chunks[p->chunk_ib_idx];
+       ib_chunk = &p->ibc;
        idx = pkt->idx + 1;
        if ((ib_chunk->kdata[idx+2] + 1) > radeon_object_size(robj)) {

               DRM_ERROR("[drm] Buffer too small for PACKET3 INDX_BUFFER "

drivers/gpu/drm/radeon/r100.c +1 -1
@@ -920,7 +919,7 @@ static int r100_packet3_check(struct radeon_cs_parser *p,
        int r;
 
        ib = p->ib->ptr;
-       ib_chunk = &p->chunks[p->chunk_ib_idx];
+       ib_chunk = &p->ibc;
        idx = pkt->idx + 1;
        switch (pkt->opcode) {

       case PACKET3_3D_LOAD_VBPNTR:

drivers/gpu/drm/radeon/r100.c +1 -1
@@ -1027,7 +1026,7 @@ int r100_cs_parse(struct radeon_cs_parser *p)
                if (r) {
                        return r;
                }
-       } while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
+       } while (p->idx < p->ibc.length_dw);
        return 0;
  }
 
drivers/gpu/drm/radeon/r300.c +1 -1
index e2ed5bc..2b2199e 100644
@@ -1024,7 +1024,7 @@ static int r300_packet0_check(struct radeon_cs_parser *p,
        int r;
 
        ib = p->ib->ptr;
-       ib_chunk = &p->chunks[p->chunk_ib_idx];
+       ib_chunk = &p->ibc;
        track = (struct r300_cs_track*)p->track;
        switch(reg) {

       case RADEON_DST_PITCH_OFFSET:

drivers/gpu/drm/radeon/r300.c +2 -2
@@ -1361,7 +1361,7 @@ static int r300_packet3_check(struct radeon_cs_parser *p,
        int r;
 
        ib = p->ib->ptr;
-       ib_chunk = &p->chunks[p->chunk_ib_idx];
+       ib_chunk = &p->ibc;
        idx = pkt->idx + 1;
        track = (struct r300_cs_track*)p->track;
        switch(pkt->opcode) {
@@ -1520,7 +1520,7 @@ int r300_cs_parse(struct radeon_cs_parser *p)
                if (r) {
                        return r;
                }
-       } while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
+       } while (p->idx < p->ibc.length_dw);
        return 0;
  }
 
drivers/gpu/drm/radeon/radeon.h +56 -54
index d61f2fc..1bea78a 100644
@@ -352,6 +352,56 @@ void radeon_irq_kms_fini(struct radeon_device *rdev);
 
 
  /*
+ * CS.
+ */
+struct radeon_ib;
+
+struct radeon_cs_reloc {
+       struct drm_gem_object           *gobj;
+       struct radeon_object            *robj;
+       struct radeon_object_list       lobj;
+       uint32_t                        handle;
+       uint32_t                        flags;
+};
+
+struct radeon_cs_chunk {
+       uint32_t                chunk_id;
+       uint32_t                length_dw;
+       uint32_t                *kdata;
+};
+
+struct radeon_cs_parser {
+       struct radeon_device    *rdev;
+       struct drm_file         *filp;
+       struct radeon_cs_chunk  ibc;
+       /* IB */
+       unsigned                idx;
+       /* relocations */
+       unsigned                nrelocs;
+       struct radeon_cs_reloc  *relocs;
+       struct radeon_cs_reloc  **relocs_ptr;
+       struct list_head        validated;
+       struct radeon_ib        *ib;
+       void                    *track;
+};
+
+struct radeon_cs_packet {
+       unsigned        idx;
+       unsigned        type;
+       unsigned        reg;
+       unsigned        opcode;
+       int             count;
+       unsigned        one_reg_wr;
+};
+
+typedef int (*radeon_packet0_check_t)(struct radeon_cs_parser *p,
+                                     struct radeon_cs_packet *pkt,
+                                     unsigned idx, unsigned reg);
+typedef int (*radeon_packet3_check_t)(struct radeon_cs_parser *p,
+                                     struct radeon_cs_packet *pkt);
+
+
+/*
  * CP & ring.
  */
  struct radeon_ib {
@@ -360,12 +410,18 @@ struct radeon_ib {
        uint64_t                gpu_addr;
        struct radeon_fence     *fence;
        volatile uint32_t       *ptr;
+       uint32_t                *ib;
+       struct radeon_cs_reloc  *relocs;
+       struct radeon_cs_reloc  **relocs_ptr;
        uint32_t                length_dw;
  };
 
  struct radeon_ib_pool {
        struct mutex            mutex;
        struct radeon_object    *robj;
+       uint32_t                *ib;
+       struct radeon_cs_reloc  *relocs;
+       struct radeon_cs_reloc  **relocs_ptr;
        struct list_head        scheduled_ibs;
        struct radeon_ib        ibs[RADEON_IB_POOL_SIZE];
        bool                    ready;
@@ -405,60 +461,6 @@ void radeon_ring_fini(struct radeon_device *rdev);
 
 
  /*
- * CS.
- */
-struct radeon_cs_reloc {
-       struct drm_gem_object           *gobj;
-       struct radeon_object            *robj;
-       struct radeon_object_list       lobj;
-       uint32_t                        handle;
-       uint32_t                        flags;
-};
-
-struct radeon_cs_chunk {
-       uint32_t                chunk_id;
-       uint32_t                length_dw;
-       uint32_t                *kdata;
-};
-
-struct radeon_cs_parser {
-       struct radeon_device    *rdev;
-       struct drm_file         *filp;
-       /* chunks */
-       unsigned                nchunks;
-       struct radeon_cs_chunk  *chunks;
-       uint64_t                *chunks_array;
-       /* IB */
-       unsigned                idx;
-       /* relocations */
-       unsigned                nrelocs;
-       struct radeon_cs_reloc  *relocs;
-       struct radeon_cs_reloc  **relocs_ptr;
-       struct list_head        validated;
-       /* indices of various chunks */
-       int                     chunk_ib_idx;
-       int                     chunk_relocs_idx;
-       struct radeon_ib        *ib;
-       void                    *track;
-};
-
-struct radeon_cs_packet {
-       unsigned        idx;
-       unsigned        type;
-       unsigned        reg;
-       unsigned        opcode;
-       int             count;
-       unsigned        one_reg_wr;
-};
-
-typedef int (*radeon_packet0_check_t)(struct radeon_cs_parser *p,
-                                     struct radeon_cs_packet *pkt,
-                                     unsigned idx, unsigned reg);
-typedef int (*radeon_packet3_check_t)(struct radeon_cs_parser *p,
-                                     struct radeon_cs_packet *pkt);
-
-
-/*
  * AGP
  */
  int radeon_agp_init(struct radeon_device *rdev);
drivers/gpu/drm/radeon/radeon_cs.c +82 -104
index b843f9b..8fb08fa 100644
@@ -29,61 +29,52 @@
  #include "radeon_reg.h"
  #include "radeon.h"
 
-void r100_cs_dump_packet(struct radeon_cs_parser *p,
-                        struct radeon_cs_packet *pkt);
-
-int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
+static int radeon_cs_parser_reloc(struct radeon_cs_parser *p,
+                                 struct drm_radeon_cs_chunk *chunk)
  {
-       struct drm_device *ddev = p->rdev->ddev;
-       struct radeon_cs_chunk *chunk;
-       unsigned i, j;
+       struct drm_radeon_cs_reloc __user *reloc_ptr;
+       struct drm_radeon_cs_reloc reloc;
+       unsigned i, j, c;
        bool duplicate;
 
-       if (p->chunk_relocs_idx == -1) {
-               return 0;
-       }
-       chunk = &p->chunks[p->chunk_relocs_idx];
-       /* FIXME: we assume that each relocs use 4 dwords */
        p->nrelocs = chunk->length_dw / 4;
-       p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
-       if (p->relocs_ptr == NULL) {
-               return -ENOMEM;
-       }
-       p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
-       if (p->relocs == NULL) {
-               return -ENOMEM;
+       if (p->nrelocs > 1024) {
+               DRM_ERROR("Too many relocations %d\n", p->nrelocs);
+               return -EINVAL;
        }
-       for (i = 0; i < p->nrelocs; i++) {
-               struct drm_radeon_cs_reloc *r;
-
+       for (i = 0, c = 0; i < p->nrelocs; i++) {
+               reloc_ptr = (void __user*)(unsigned long)(chunk->chunk_data + i * 16);
+               if (DRM_COPY_FROM_USER(&reloc, reloc_ptr, 16)) {
+                       return -EFAULT;
+               }
                duplicate = false;
-               r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
-               for (j = 0; j < p->nrelocs; j++) {
-                       if (r->handle == p->relocs[j].handle) {
+               for (j = 0; j < c; j++) {
+                       if (reloc.handle == p->relocs[j].handle) {
                                p->relocs_ptr[i] = &p->relocs[j];
                                duplicate = true;
                                break;
                        }
                }
                if (!duplicate) {
-                       p->relocs[i].gobj = drm_gem_object_lookup(ddev,
+                       p->relocs[i].gobj = drm_gem_object_lookup(p->rdev->ddev,
                                                                  p->filp,
-                                                                 r->handle);
+                                                                 reloc.handle);
                        if (p->relocs[i].gobj == NULL) {
                                DRM_ERROR("gem object lookup failed 0x%x\n",
-                                         r->handle);
+                                         reloc.handle);
                                return -EINVAL;
                        }
                        p->relocs_ptr[i] = &p->relocs[i];
                        p->relocs[i].robj = p->relocs[i].gobj->driver_private;
                        p->relocs[i].lobj.robj = p->relocs[i].robj;
-                       p->relocs[i].lobj.rdomain = r->read_domains;
-                       p->relocs[i].lobj.wdomain = r->write_domain;
-                       p->relocs[i].handle = r->handle;
-                       p->relocs[i].flags = r->flags;
+                       p->relocs[i].lobj.rdomain = reloc.read_domains;
+                       p->relocs[i].lobj.wdomain = reloc.write_domain;
+                       p->relocs[i].handle = reloc.handle;
+                       p->relocs[i].flags = reloc.flags;
                        INIT_LIST_HEAD(&p->relocs[i].lobj.list);
                        radeon_object_list_add_object(&p->relocs[i].lobj,
                                                      &p->validated);
+                       c++;
                }
        }
        return radeon_object_list_validate(&p->validated, p->ib->fence);
@@ -92,65 +83,53 @@ int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
  int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
  {
        struct drm_radeon_cs *cs = data;
-       uint64_t *chunk_array_ptr;
-       unsigned size, i;
+       unsigned i;
+       int r;
+       bool found_ib = false;
 
-       if (!cs->num_chunks) {
-               return 0;
-       }
-       /* get chunks */
-       INIT_LIST_HEAD(&p->validated);
-       p->idx = 0;
-       p->chunk_ib_idx = -1;
-       p->chunk_relocs_idx = -1;
-       p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
-       if (p->chunks_array == NULL) {
-               return -ENOMEM;
-       }
-       chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
-       if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
-                              sizeof(uint64_t)*cs->num_chunks)) {
-               return -EFAULT;
-       }
-       p->nchunks = cs->num_chunks;
-       p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
-       if (p->chunks == NULL) {
-               return -ENOMEM;
-       }
-       for (i = 0; i < p->nchunks; i++) {
-               struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
-               struct drm_radeon_cs_chunk user_chunk;
+       for (i = 0; i < cs->num_chunks; i++) {
+               struct drm_radeon_cs_chunk __user *chunk_ptr = NULL;
+               struct drm_radeon_cs_chunk chunk;
                uint32_t __user *cdata;
+               unsigned tmp;
+               u64 ptr;
 
-               chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
-               if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
-                                      sizeof(struct drm_radeon_cs_chunk))) {
+               chunk_ptr = (void __user*)(unsigned long)(cs->chunks + i * 8);
+               if (DRM_COPY_FROM_USER(&ptr, chunk_ptr,  sizeof(u64))) {
                        return -EFAULT;
                }
-               p->chunks[i].chunk_id = user_chunk.chunk_id;
-               if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
-                       p->chunk_relocs_idx = i;
-               }
-               if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
-                       p->chunk_ib_idx = i;
-               }
-               p->chunks[i].length_dw = user_chunk.length_dw;
-               cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
-
-               p->chunks[i].kdata = NULL;
-               size = p->chunks[i].length_dw * sizeof(uint32_t);
-               p->chunks[i].kdata = kzalloc(size, GFP_KERNEL);
-               if (p->chunks[i].kdata == NULL) {
-                       return -ENOMEM;
-               }
-               if (DRM_COPY_FROM_USER(p->chunks[i].kdata, cdata, size)) {
+               chunk_ptr = (void __user*)(unsigned long)ptr;
+               if (DRM_COPY_FROM_USER(&chunk, chunk_ptr, sizeof(struct drm_radeon_cs_chunk))) {
                        return -EFAULT;
                }
-       }
-       if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
-               DRM_ERROR("cs IB too big: %d\n",
-                         p->chunks[p->chunk_ib_idx].length_dw);
-               return -EINVAL;
+               switch (chunk.chunk_id) {
+               case RADEON_CHUNK_ID_RELOCS:
+                       r = radeon_cs_parser_reloc(p, &chunk);
+                       if (r) {
+                               return r;
+                       }
+                       break;
+               case RADEON_CHUNK_ID_IB:
+                       if (found_ib) {
+                               DRM_ERROR("Multiple IB chunks not supported\n");
+                               return -EINVAL;
+                       }
+                       found_ib = true;
+                       if (chunk.length_dw > 16 * 1024) {
+                               DRM_ERROR("cs IB too big: %d\n", chunk.length_dw);
+                               return -EINVAL;
+                       }
+                       p->ibc.chunk_id = chunk.chunk_id;
+                       p->ibc.length_dw = chunk.length_dw;
+                       tmp = p->ibc.length_dw * sizeof(u32);
+                       cdata = (uint32_t *)(unsigned long)chunk.chunk_data;
+                       if (DRM_COPY_FROM_USER(p->ibc.kdata, cdata, tmp)) {
+                               return -EFAULT;
+                       }
+                       break;
+               default:
+                       break;
+               }
        }
        return 0;
  }
@@ -179,13 +158,6 @@ static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
                        mutex_unlock(&parser->rdev->ddev->struct_mutex);
                }
        }
-       kfree(parser->relocs);
-       kfree(parser->relocs_ptr);
-       for (i = 0; i < parser->nchunks; i++) {
-               kfree(parser->chunks[i].kdata);
-       }
-       kfree(parser->chunks);
-       kfree(parser->chunks_array);
        radeon_ib_free(parser->rdev, &parser->ib);
  }
 
@@ -193,7 +165,6 @@ int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
  {
        struct radeon_device *rdev = dev->dev_private;
        struct radeon_cs_parser parser;
-       struct radeon_cs_chunk *ib_chunk;
        int r;
 
        mutex_lock(&rdev->cs_mutex);
@@ -202,36 +173,43 @@ int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
                return -EINVAL;
        }
        /* initialize parser */
-       memset(&parser, 0, sizeof(struct radeon_cs_parser));
+       INIT_LIST_HEAD(&parser.validated);
+       parser.idx = 0;
+       parser.ibc.kdata = NULL;
+       parser.ibc.length_dw = 0;
+       parser.relocs = NULL;
+       parser.relocs_ptr = NULL;
+       parser.ib = NULL;
        parser.filp = filp;
        parser.rdev = rdev;
-       r = radeon_cs_parser_init(&parser, data);
+       r =  radeon_ib_get(rdev, &parser.ib);
        if (r) {
-               DRM_ERROR("Failed to initialize parser !\n");
+               DRM_ERROR("Failed to get ib !\n");
                radeon_cs_parser_fini(&parser, r);
                mutex_unlock(&rdev->cs_mutex);
                return r;
        }
-       r =  radeon_ib_get(rdev, &parser.ib);
+       parser.ibc.kdata = parser.ib->ib;
+       parser.relocs = parser.ib->relocs;
+       parser.relocs_ptr = parser.ib->relocs_ptr;
+       r = radeon_cs_parser_init(&parser, data);
        if (r) {
-               DRM_ERROR("Failed to get ib !\n");
+               DRM_ERROR("Failed to initialize parser !\n");
                radeon_cs_parser_fini(&parser, r);
                mutex_unlock(&rdev->cs_mutex);
                return r;
        }
-       r = radeon_cs_parser_relocs(&parser);
-       if (r) {
-               DRM_ERROR("Failed to parse relocation !\n");
+       if (parser.ibc.kdata == NULL) {
+               DRM_INFO("No IB chunk in cs, nothing to do\n");
                radeon_cs_parser_fini(&parser, r);
                mutex_unlock(&rdev->cs_mutex);
-               return r;
+               return 0;
        }
        /* Copy the packet into the IB, the parser will read from the
         * input memory (cached) and write to the IB (which can be
         * uncached). */
-       ib_chunk = &parser.chunks[parser.chunk_ib_idx];
-       parser.ib->length_dw = ib_chunk->length_dw;
-       memcpy((void *)parser.ib->ptr, ib_chunk->kdata, ib_chunk->length_dw*4);
+       parser.ib->length_dw = parser.ibc.length_dw;
+       memcpy((void *)parser.ib->ptr, parser.ibc.kdata, parser.ibc.length_dw * 4);
        r = radeon_cs_parse(&parser);
        if (r) {
                DRM_ERROR("Invalid command stream !\n");
drivers/gpu/drm/radeon/radeon_ring.c +23 -1
index a853261..0026a11 100644
@@ -177,6 +177,20 @@ int radeon_ib_pool_init(struct radeon_device *rdev)
 
        /* Allocate 1M object buffer */
        INIT_LIST_HEAD(&rdev->ib_pool.scheduled_ibs);
+       rdev->ib_pool.ib = drm_calloc_large(RADEON_IB_POOL_SIZE, 64 * 1024);
+       if (rdev->ib_pool.ib == NULL) {
+               return -ENOMEM;
+       }
+       rdev->ib_pool.relocs = drm_calloc_large(RADEON_IB_POOL_SIZE * 1024,
+                                               sizeof(struct radeon_cs_reloc));
+       if (rdev->ib_pool.relocs == NULL) {
+               return -ENOMEM;
+       }
+       rdev->ib_pool.relocs_ptr = drm_calloc_large(RADEON_IB_POOL_SIZE * 1024,
+                                                   sizeof(void *));
+       if (rdev->ib_pool.relocs == NULL) {
+               return -ENOMEM;
+       }
        r = radeon_object_create(rdev, NULL,  RADEON_IB_POOL_SIZE*64*1024,
                                 true, RADEON_GEM_DOMAIN_GTT,
                                 false, &rdev->ib_pool.robj);
@@ -202,6 +216,9 @@ int radeon_ib_pool_init(struct radeon_device *rdev)
                rdev->ib_pool.ibs[i].ptr = ptr + offset;
                rdev->ib_pool.ibs[i].idx = i;
                rdev->ib_pool.ibs[i].length_dw = 0;
+               rdev->ib_pool.ibs[i].ib = &rdev->ib_pool.ib[i * 16 * 1024];
+               rdev->ib_pool.ibs[i].relocs = &rdev->ib_pool.relocs[i * 1024];
+               rdev->ib_pool.ibs[i].relocs_ptr = &rdev->ib_pool.relocs_ptr[i * 1024];
                INIT_LIST_HEAD(&rdev->ib_pool.ibs[i].list);
        }
        bitmap_zero(rdev->ib_pool.alloc_bm, RADEON_IB_POOL_SIZE);
@@ -225,6 +242,12 @@ void radeon_ib_pool_fini(struct radeon_device *rdev)
                radeon_object_unref(&rdev->ib_pool.robj);
                rdev->ib_pool.robj = NULL;
        }
+       if (rdev->ib_pool.ib)
+               drm_free_large(rdev->ib_pool.ib);
+       if (rdev->ib_pool.relocs)
+               drm_free_large(rdev->ib_pool.relocs);
+       if (rdev->ib_pool.relocs_ptr)
+               drm_free_large(rdev->ib_pool.relocs_ptr);
        mutex_unlock(&rdev->ib_pool.mutex);
  }
 
--

1.6.2.2

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

#2 lore raw
Quoted 9 lines · Pekka Enberg

Hi Jerome,

Quoted 6 lines · Jerome Glisse

Command stream parsing is the most common operation and can happen hundred of times per second, we don't want to allocate/free memory each time this ioctl is call. This rework the ioctl to avoid doing so by allocating temporary memory along the ib pool.

Quoted 6 lines
Quoted 1 line

So how much does this help (i.e. where are the numbers)? I am bit surprised "hundred of times per second" is an issue for our slab allocators. Hmm?

I didn't have real number but the vmap path was really slower, quake3 fps goes from ~20 to ~40 on average when going from vmap to preallocated. When using kmalloc i don't thing there was so much performance hit. But i think the biggest hit was that in previous code i asked for zeroed memory so i am pretty sure kernel spend a bit of time clearing page. I reworked the code to avoid needing cleared memory and so avoid memset, this is likely why we get a performance boost.

Cheers,
Jerome

#3 lore raw

Hi Jerome,

Quoted 12 lines · Pekka Enberg
Quoted 12 lines · Jerome Glisse
Quoted 7 lines

Command stream parsing is the most common operation and can happen hundred of times per second, we don't want to allocate/free memory each time this ioctl is call. This rework the ioctl to avoid doing so by allocating temporary memory along the ib pool.

So how much does this help (i.e. where are the numbers)? I am bit surprised "hundred of times per second" is an issue for our slab allocators. Hmm?

Quoted 8 lines · Jerome Glisse

I didn't have real number but the vmap path was really slower, quake3 fps goes from ~20 to ~40 on average when going from vmap to preallocated. When using kmalloc i don't thing there was so much performance hit. But i think the biggest hit was that in previous code i asked for zeroed memory so i am pretty sure kernel spend a bit of time clearing page. I reworked the code to avoid needing cleared memory and so avoid memset, this is likely why we get a performance boost.

OK. If kmalloc() (without memset) really was too slow for your case, I'd be interested in looking at it in more detail. I'm not completely convinced the memory pool is needed here but I'm not a DRM expert so I'm not NAK'ing this either...

Pekka

#4 lore raw

Pekka Enberg skrev:

Quoted 31 lines

Hi Jerome,

Quoted 14 lines · Pekka Enberg
Quoted 14 lines · Jerome Glisse
Quoted 9 lines

Command stream parsing is the most common operation and can happen hundred of times per second, we don't want to allocate/free memory each time this ioctl is call. This rework the ioctl to avoid doing so by allocating temporary memory along the ib pool.

So how much does this help (i.e. where are the numbers)? I am bit surprised "hundred of times per second" is an issue for our slab allocators. Hmm?

Quoted 9 lines · Jerome Glisse

I didn't have real number but the vmap path was really slower, quake3 fps goes from ~20 to ~40 on average when going from vmap to preallocated. When using kmalloc i don't thing there was so much performance hit. But i think the biggest hit was that in previous code i asked for zeroed memory so i am pretty sure kernel spend a bit of time clearing page. I reworked the code to avoid needing cleared memory and so avoid memset, this is likely why we get a performance boost.

Quoted 8 lines

OK. If kmalloc() (without memset) really was too slow for your case, I'd be interested in looking at it in more detail. I'm not completely convinced the memory pool is needed here but I'm not a DRM expert so I'm not NAK'ing this either...

Pekka

Hi!
From previous experience with other drivers kmalloc() is just fine performance-wise.
I've also never seen memsetting pages turn up on the profile. It would be interesting to see an oprofile timing of this to try and pinpoint what's happening.

However, in this case, I believe Jerome was forced to use vmalloc to guarantee that the allocation would succeed, and frequent vmallocs seem to be a performance killer.

One should also be careful about frame-rates. Tuning memory manager / command submission operation is usually a matter of how much CPU is consumed for a given framerate. If one compares framerates one must make sure that the CPU is at nearly 100% while benchmarking.

/Thomas

Quoted 7 lines

------------------------------------------------------------------------------
--
_______________________________________________
Dri-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/dri-devel

#5 lore raw
Quoted 51 lines · Thomas Hellström

Pekka Enberg skrev:

Quoted 31 lines

Hi Jerome,

Quoted 14 lines · Pekka Enberg
Quoted 14 lines · Jerome Glisse
Quoted 9 lines

Command stream parsing is the most common operation and can happen hundred of times per second, we don't want to allocate/free memory each time this ioctl is call. This rework the ioctl to avoid doing so by allocating temporary memory along the ib pool.

So how much does this help (i.e. where are the numbers)? I am bit surprised "hundred of times per second" is an issue for our slab allocators. Hmm?

Quoted 9 lines · Jerome Glisse

I didn't have real number but the vmap path was really slower, quake3 fps goes from ~20 to ~40 on average when going from vmap to preallocated. When using kmalloc i don't thing there was so much performance hit. But i think the biggest hit was that in previous code i asked for zeroed memory so i am pretty sure kernel spend a bit of time clearing page. I reworked the code to avoid needing cleared memory and so avoid memset, this is likely why we get a performance boost.

Quoted 8 lines

OK. If kmalloc() (without memset) really was too slow for your case, I'd be interested in looking at it in more detail. I'm not completely convinced the memory pool is needed here but I'm not a DRM expert so I'm not NAK'ing this either...

Pekka

Hi!
From previous experience with other drivers kmalloc() is just fine
performance-wise.
I've also never seen memsetting pages turn up on the profile. It would be interesting to see an oprofile timing of this to try and pinpoint what's happening.

However, in this case, I believe Jerome was forced to use vmalloc to guarantee that the allocation would succeed, and frequent vmallocs seem to be a performance killer.

Quoted 7 lines

One should also be careful about frame-rates. Tuning memory manager / command submission operation is usually a matter of how much CPU is consumed for a given framerate. If one compares framerates one must make sure that the CPU is at nearly 100% while benchmarking.

/Thomas

To give a more correct rough estimate, at 60fps we will issue somethings like 10 to 50 cs ioctl per frame so it's several thousands of cs ioctl so several thousands of 64K allocation, and memory clearing, i believe such things would show up on profile. I am not running kernel without my patch as i am working on other stuff now, but i will lower the pool size so that we don't waste too much memory right now i think the preallocation use somethings around 8M of memory. I think i can get it down to 1M without impacting performance (even less if we are on pcie).

Cheers,
Jerome