* [PATCH 1/9] ttm/radeon/nouveau: Check the DMA address from TTM against known value.
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
@ 2011-09-29 20:33 ` Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 2/9] ttm: Introduce ttm_page_alloc_func structure Konrad Rzeszutek Wilk
` (8 subsequent siblings)
9 siblings, 0 replies; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-29 20:33 UTC (permalink / raw)
To: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher
Cc: xen-devel, Konrad Rzeszutek Wilk
. instead of checking against the DMA_ERROR_CODE value which is
per-platform specific. The zero value is a known invalid value
that the TTM layer sets on the dma_address array if it is not
used (ttm_tt_alloc_page_directory calls drm_calloc_large which
creates a page with GFP_ZERO).
We can't use pci_dma_mapping_error as that is IOMMU
specific (some check for a specific physical address, some
for ranges, some just do a check against zero).
Also update the comments in the header about the true state
of that parameter.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/nouveau/nouveau_sgdma.c | 3 +--
drivers/gpu/drm/radeon/radeon_gart.c | 4 +---
include/drm/ttm/ttm_page_alloc.h | 4 ++--
3 files changed, 4 insertions(+), 7 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_sgdma.c b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
index 82fad91..9b570c3 100644
--- a/drivers/gpu/drm/nouveau/nouveau_sgdma.c
+++ b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
@@ -42,8 +42,7 @@ nouveau_sgdma_populate(struct ttm_backend *be, unsigned long num_pages,
nvbe->nr_pages = 0;
while (num_pages--) {
- /* this code path isn't called and is incorrect anyways */
- if (0) { /*dma_addrs[nvbe->nr_pages] != DMA_ERROR_CODE)*/
+ if (dma_addrs[nvbe->nr_pages] != 0) {
nvbe->pages[nvbe->nr_pages] =
dma_addrs[nvbe->nr_pages];
nvbe->ttm_alloced[nvbe->nr_pages] = true;
diff --git a/drivers/gpu/drm/radeon/radeon_gart.c b/drivers/gpu/drm/radeon/radeon_gart.c
index a533f52..068ba09 100644
--- a/drivers/gpu/drm/radeon/radeon_gart.c
+++ b/drivers/gpu/drm/radeon/radeon_gart.c
@@ -181,9 +181,7 @@ int radeon_gart_bind(struct radeon_device *rdev, unsigned offset,
p = t / (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
for (i = 0; i < pages; i++, p++) {
- /* we reverted the patch using dma_addr in TTM for now but this
- * code stops building on alpha so just comment it out for now */
- if (0) { /*dma_addr[i] != DMA_ERROR_CODE) */
+ if (dma_addr[i] != 0) {
rdev->gart.ttm_alloced[p] = true;
rdev->gart.pages_addr[p] = dma_addr[i];
} else {
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h
index 8062890..0017b17 100644
--- a/include/drm/ttm/ttm_page_alloc.h
+++ b/include/drm/ttm/ttm_page_alloc.h
@@ -36,7 +36,7 @@
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state for the page.
* @count: number of pages to allocate.
- * @dma_address: The DMA (bus) address of pages (if TTM_PAGE_FLAG_DMA32 set).
+ * @dma_address: The DMA (bus) address of pages - (by default zero).
*/
int ttm_get_pages(struct list_head *pages,
int flags,
@@ -51,7 +51,7 @@ int ttm_get_pages(struct list_head *pages,
* count.
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state.
- * @dma_address: The DMA (bus) address of pages (if TTM_PAGE_FLAG_DMA32 set).
+ * @dma_address: The DMA (bus) address of pages (by default zero).
*/
void ttm_put_pages(struct list_head *pages,
unsigned page_count,
--
1.7.4.1
^ permalink raw reply related [flat|nested] 16+ messages in thread* [PATCH 2/9] ttm: Introduce ttm_page_alloc_func structure.
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 1/9] ttm/radeon/nouveau: Check the DMA address from TTM against known value Konrad Rzeszutek Wilk
@ 2011-09-29 20:33 ` Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 3/9] ttm: Pass in 'struct device' to TTM so it can do DMA API on behalf of device Konrad Rzeszutek Wilk
` (7 subsequent siblings)
9 siblings, 0 replies; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-29 20:33 UTC (permalink / raw)
To: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher
Cc: xen-devel, Konrad Rzeszutek Wilk
Which has the function members for all of the current page pool
operations defined. The old calls (ttm_put_pages, ttm_get_pages, etc)
are plumbed through little functions which lookup in the ttm_page_alloc_func
the appropiate implementation and call it.
There is currently only one page pool code so the default registration
goes to 'ttm_page_alloc_default'. The subsequent patch
"ttm: Provide a DMA aware TTM page pool code." introduces the one
to be used when the SWIOTLB code is turned on (that implementation
is a union of the default TTM pool code with the DMA pool code).
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/ttm/ttm_memory.c | 3 ++
drivers/gpu/drm/ttm/ttm_page_alloc.c | 58 ++++++++++++++++++++++++++++----
include/drm/ttm/ttm_page_alloc.h | 60 ++++++++++++++++++++++++++++++++++
3 files changed, 113 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_memory.c b/drivers/gpu/drm/ttm/ttm_memory.c
index e70ddd8..c7d97a5 100644
--- a/drivers/gpu/drm/ttm/ttm_memory.c
+++ b/drivers/gpu/drm/ttm/ttm_memory.c
@@ -356,6 +356,8 @@ static int ttm_mem_init_dma32_zone(struct ttm_mem_global *glob,
}
#endif
+struct ttm_page_alloc_func *ttm_page_alloc;
+
int ttm_mem_global_init(struct ttm_mem_global *glob)
{
struct sysinfo si;
@@ -394,6 +396,7 @@ int ttm_mem_global_init(struct ttm_mem_global *glob)
"Zone %7s: Available graphics memory: %llu kiB.\n",
zone->name, (unsigned long long) zone->max_mem >> 10);
}
+ ttm_page_alloc = &ttm_page_alloc_default;
ttm_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
return 0;
out_no_zone:
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c
index d948575..6a888f8 100644
--- a/drivers/gpu/drm/ttm/ttm_page_alloc.c
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c
@@ -664,9 +664,9 @@ out:
* On success pages list will hold count number of correctly
* cached pages.
*/
-int ttm_get_pages(struct list_head *pages, int flags,
- enum ttm_caching_state cstate, unsigned count,
- dma_addr_t *dma_address)
+int __ttm_get_pages(struct list_head *pages, int flags,
+ enum ttm_caching_state cstate, unsigned count,
+ dma_addr_t *dma_address)
{
struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
struct page *p = NULL;
@@ -734,8 +734,8 @@ int ttm_get_pages(struct list_head *pages, int flags,
}
/* Put all pages in pages list to correct pool to wait for reuse */
-void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
- enum ttm_caching_state cstate, dma_addr_t *dma_address)
+void __ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
+ enum ttm_caching_state cstate, dma_addr_t *dma_address)
{
unsigned long irq_flags;
struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
@@ -785,7 +785,7 @@ static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
pool->name = name;
}
-int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
+int __ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
{
int ret;
@@ -822,7 +822,7 @@ int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
return 0;
}
-void ttm_page_alloc_fini(void)
+void __ttm_page_alloc_fini(void)
{
int i;
@@ -836,7 +836,7 @@ void ttm_page_alloc_fini(void)
_manager = NULL;
}
-int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
+int __ttm_page_alloc_debugfs(struct seq_file *m, void *data)
{
struct ttm_page_pool *p;
unsigned i;
@@ -856,4 +856,46 @@ int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
}
return 0;
}
+
+struct ttm_page_alloc_func ttm_page_alloc_default = {
+ .get_pages = __ttm_get_pages,
+ .put_pages = __ttm_put_pages,
+ .alloc_init = __ttm_page_alloc_init,
+ .alloc_fini = __ttm_page_alloc_fini,
+ .debugfs = __ttm_page_alloc_debugfs,
+};
+
+int ttm_get_pages(struct list_head *pages, int flags,
+ enum ttm_caching_state cstate, unsigned count,
+ dma_addr_t *dma_address)
+{
+ if (ttm_page_alloc && ttm_page_alloc->get_pages)
+ return ttm_page_alloc->get_pages(pages, flags, cstate, count,
+ dma_address);
+ return -1;
+}
+void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
+ enum ttm_caching_state cstate, dma_addr_t *dma_address)
+{
+ if (ttm_page_alloc && ttm_page_alloc->put_pages)
+ ttm_page_alloc->put_pages(pages, page_count, flags, cstate,
+ dma_address);
+}
+int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
+{
+ if (ttm_page_alloc && ttm_page_alloc->alloc_init)
+ return ttm_page_alloc->alloc_init(glob, max_pages);
+ return -1;
+}
+void ttm_page_alloc_fini(void)
+{
+ if (ttm_page_alloc && ttm_page_alloc->alloc_fini)
+ ttm_page_alloc->alloc_fini();
+}
+int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
+{
+ if (ttm_page_alloc && ttm_page_alloc->debugfs)
+ return ttm_page_alloc->debugfs(m, data);
+ return -1;
+}
EXPORT_SYMBOL(ttm_page_alloc_debugfs);
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h
index 0017b17..6e8d73a 100644
--- a/include/drm/ttm/ttm_page_alloc.h
+++ b/include/drm/ttm/ttm_page_alloc.h
@@ -29,6 +29,66 @@
#include "ttm_bo_driver.h"
#include "ttm_memory.h"
+struct ttm_page_alloc_func {
+ /**
+ * struct ttm_page_alloc_func member get_pages
+ * Get count number of pages from pool to pages list.
+ *
+ * @pages: head of empty linked list where pages are filled.
+ * @flags: ttm flags for page allocation.
+ * @cstate: ttm caching state for the page.
+ * @count: number of pages to allocate.
+ * @dma_address: The DMA (bus) address of pages (by default zero).
+ */
+ int (*get_pages) (struct list_head *pages,
+ int flags,
+ enum ttm_caching_state cstate,
+ unsigned count,
+ dma_addr_t *dma_address);
+ /**
+ * struct ttm_page_alloc_func member put_pages.
+ *
+ * Put linked list of pages to pool.
+ *
+ * @pages: list of pages to free.
+ * @page_count: number of pages in the list. Zero can be passed for
+ * unknown count.
+ * @flags: ttm flags for page allocation.
+ * @cstate: ttm caching state.
+ * @dma_address: The DMA (bus) address of pages (by default zero).
+ */
+ void (*put_pages)(struct list_head *pages,
+ unsigned page_count,
+ int flags,
+ enum ttm_caching_state cstate,
+ dma_addr_t *dma_address);
+ /**
+ * struct ttm_page_alloc_func member alloc_init.
+ *
+ * Initialize pool allocator.
+ */
+ int (*alloc_init)(struct ttm_mem_global *glob, unsigned max_pages);
+
+ /**
+ * struct ttm_page_alloc_func member alloc_fini.
+ *
+ * Free pool allocator.
+ */
+ void (*alloc_fini)(void);
+
+ /**
+ * struct ttm_page_alloc_func member debugfs.
+ *
+ * Output the state of pools to debugfs file
+ */
+ int (*debugfs)(struct seq_file *m, void *data);
+};
+
+extern struct ttm_page_alloc_func *ttm_page_alloc;
+
+/* Defined in ttm_page_alloc.c */
+extern struct ttm_page_alloc_func ttm_page_alloc_default;
+
/**
* Get count number of pages from pool to pages list.
*
--
1.7.4.1
^ permalink raw reply related [flat|nested] 16+ messages in thread* [PATCH 3/9] ttm: Pass in 'struct device' to TTM so it can do DMA API on behalf of device.
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 1/9] ttm/radeon/nouveau: Check the DMA address from TTM against known value Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 2/9] ttm: Introduce ttm_page_alloc_func structure Konrad Rzeszutek Wilk
@ 2011-09-29 20:33 ` Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 4/9] swiotlb: Expose swiotlb_nr_tlb function to modules Konrad Rzeszutek Wilk
` (6 subsequent siblings)
9 siblings, 0 replies; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-29 20:33 UTC (permalink / raw)
To: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher
Cc: xen-devel, Konrad Rzeszutek Wilk
We want to pass in the 'struct device' to the TTM layer so that
the TTM DMA pool code (if enabled) can use it. The DMA API code
needs the 'struct device' to do the DMA API operations.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/nouveau/nouveau_mem.c | 3 ++-
drivers/gpu/drm/radeon/radeon_ttm.c | 3 ++-
drivers/gpu/drm/ttm/ttm_bo.c | 4 +++-
drivers/gpu/drm/ttm/ttm_page_alloc.c | 17 ++++++++++-------
drivers/gpu/drm/ttm/ttm_tt.c | 5 +++--
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c | 4 ++--
include/drm/ttm/ttm_bo_driver.h | 7 ++++++-
include/drm/ttm/ttm_page_alloc.h | 16 ++++++++++++----
8 files changed, 40 insertions(+), 19 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_mem.c b/drivers/gpu/drm/nouveau/nouveau_mem.c
index 5ee14d2..a2d7e35 100644
--- a/drivers/gpu/drm/nouveau/nouveau_mem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_mem.c
@@ -417,7 +417,8 @@ nouveau_mem_vram_init(struct drm_device *dev)
ret = ttm_bo_device_init(&dev_priv->ttm.bdev,
dev_priv->ttm.bo_global_ref.ref.object,
&nouveau_bo_driver, DRM_FILE_PAGE_OFFSET,
- dma_bits <= 32 ? true : false);
+ dma_bits <= 32 ? true : false,
+ dev->dev);
if (ret) {
NV_ERROR(dev, "Error initialising bo driver: %d\n", ret);
return ret;
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index 60125dd..dbc6bcb 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -517,7 +517,8 @@ int radeon_ttm_init(struct radeon_device *rdev)
r = ttm_bo_device_init(&rdev->mman.bdev,
rdev->mman.bo_global_ref.ref.object,
&radeon_bo_driver, DRM_FILE_PAGE_OFFSET,
- rdev->need_dma32);
+ rdev->need_dma32,
+ rdev->dev);
if (r) {
DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
return r;
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 2e618b5..0358889 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -1527,12 +1527,14 @@ int ttm_bo_device_init(struct ttm_bo_device *bdev,
struct ttm_bo_global *glob,
struct ttm_bo_driver *driver,
uint64_t file_page_offset,
- bool need_dma32)
+ bool need_dma32,
+ struct device *dev)
{
int ret = -EINVAL;
rwlock_init(&bdev->vm_lock);
bdev->driver = driver;
+ bdev->dev = dev;
memset(bdev->man, 0, sizeof(bdev->man));
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c
index 6a888f8..f9a4d83 100644
--- a/drivers/gpu/drm/ttm/ttm_page_alloc.c
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c
@@ -666,7 +666,7 @@ out:
*/
int __ttm_get_pages(struct list_head *pages, int flags,
enum ttm_caching_state cstate, unsigned count,
- dma_addr_t *dma_address)
+ dma_addr_t *dma_address, struct device *dev)
{
struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
struct page *p = NULL;
@@ -724,7 +724,8 @@ int __ttm_get_pages(struct list_head *pages, int flags,
printk(KERN_ERR TTM_PFX
"Failed to allocate extra pages "
"for large request.");
- ttm_put_pages(pages, 0, flags, cstate, NULL);
+ ttm_put_pages(pages, 0, flags, cstate, dma_address,
+ dev);
return r;
}
}
@@ -735,7 +736,8 @@ int __ttm_get_pages(struct list_head *pages, int flags,
/* Put all pages in pages list to correct pool to wait for reuse */
void __ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
- enum ttm_caching_state cstate, dma_addr_t *dma_address)
+ enum ttm_caching_state cstate, dma_addr_t *dma_address,
+ struct device *dev)
{
unsigned long irq_flags;
struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
@@ -867,19 +869,20 @@ struct ttm_page_alloc_func ttm_page_alloc_default = {
int ttm_get_pages(struct list_head *pages, int flags,
enum ttm_caching_state cstate, unsigned count,
- dma_addr_t *dma_address)
+ dma_addr_t *dma_address, struct device *dev)
{
if (ttm_page_alloc && ttm_page_alloc->get_pages)
return ttm_page_alloc->get_pages(pages, flags, cstate, count,
- dma_address);
+ dma_address, dev);
return -1;
}
void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
- enum ttm_caching_state cstate, dma_addr_t *dma_address)
+ enum ttm_caching_state cstate, dma_addr_t *dma_address,
+ struct device *dev)
{
if (ttm_page_alloc && ttm_page_alloc->put_pages)
ttm_page_alloc->put_pages(pages, page_count, flags, cstate,
- dma_address);
+ dma_address, dev);
}
int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
{
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 58c271e..1f11a33 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -111,7 +111,7 @@ static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
INIT_LIST_HEAD(&h);
ret = ttm_get_pages(&h, ttm->page_flags, ttm->caching_state, 1,
- &ttm->dma_address[index]);
+ &ttm->dma_address[index], ttm->dev);
if (ret != 0)
return NULL;
@@ -305,7 +305,7 @@ static void ttm_tt_free_alloced_pages(struct ttm_tt *ttm)
}
}
ttm_put_pages(&h, count, ttm->page_flags, ttm->caching_state,
- ttm->dma_address);
+ ttm->dma_address, ttm->dev);
ttm->state = tt_unpopulated;
ttm->first_himem_page = ttm->num_pages;
ttm->last_lomem_page = -1;
@@ -398,6 +398,7 @@ struct ttm_tt *ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
ttm->last_lomem_page = -1;
ttm->caching_state = tt_cached;
ttm->page_flags = page_flags;
+ ttm->dev = bdev->dev;
ttm->dummy_read_page = dummy_read_page;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
index 96949b9..d266ae7 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
@@ -322,11 +322,11 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
ttm_lock_set_kill(&dev_priv->fbdev_master.lock, false, SIGTERM);
dev_priv->active_master = &dev_priv->fbdev_master;
-
ret = ttm_bo_device_init(&dev_priv->bdev,
dev_priv->bo_global_ref.ref.object,
&vmw_bo_driver, VMWGFX_FILE_PAGE_OFFSET,
- false);
+ false,
+ dev->dev);
if (unlikely(ret != 0)) {
DRM_ERROR("Failed initializing TTM buffer object driver.\n");
goto out_err1;
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 09af2d7..9a7ecae 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -177,6 +177,7 @@ struct ttm_tt {
tt_unpopulated,
} state;
dma_addr_t *dma_address;
+ struct device *dev;
};
#define TTM_MEMTYPE_FLAG_FIXED (1 << 0) /* Fixed (on-card) PCI memory */
@@ -551,6 +552,7 @@ struct ttm_bo_device {
struct list_head device_list;
struct ttm_bo_global *glob;
struct ttm_bo_driver *driver;
+ struct device *dev;
rwlock_t vm_lock;
struct ttm_mem_type_manager man[TTM_NUM_MEM_TYPES];
spinlock_t fence_lock;
@@ -791,6 +793,8 @@ extern int ttm_bo_device_release(struct ttm_bo_device *bdev);
* @file_page_offset: Offset into the device address space that is available
* for buffer data. This ensures compatibility with other users of the
* address space.
+ * @need_dma32: Allocate pages under 4GB
+ * @dev: 'struct device' of the PCI device.
*
* Initializes a struct ttm_bo_device:
* Returns:
@@ -799,7 +803,8 @@ extern int ttm_bo_device_release(struct ttm_bo_device *bdev);
extern int ttm_bo_device_init(struct ttm_bo_device *bdev,
struct ttm_bo_global *glob,
struct ttm_bo_driver *driver,
- uint64_t file_page_offset, bool need_dma32);
+ uint64_t file_page_offset, bool need_dma32,
+ struct device *dev);
/**
* ttm_bo_unmap_virtual
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h
index 6e8d73a..8fc92f2 100644
--- a/include/drm/ttm/ttm_page_alloc.h
+++ b/include/drm/ttm/ttm_page_alloc.h
@@ -39,12 +39,14 @@ struct ttm_page_alloc_func {
* @cstate: ttm caching state for the page.
* @count: number of pages to allocate.
* @dma_address: The DMA (bus) address of pages (by default zero).
+ * @dev: The device that needs this.
*/
int (*get_pages) (struct list_head *pages,
int flags,
enum ttm_caching_state cstate,
unsigned count,
- dma_addr_t *dma_address);
+ dma_addr_t *dma_address,
+ struct device *dev);
/**
* struct ttm_page_alloc_func member put_pages.
*
@@ -56,12 +58,14 @@ struct ttm_page_alloc_func {
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state.
* @dma_address: The DMA (bus) address of pages (by default zero).
+ * @dev: The device that needs this.
*/
void (*put_pages)(struct list_head *pages,
unsigned page_count,
int flags,
enum ttm_caching_state cstate,
- dma_addr_t *dma_address);
+ dma_addr_t *dma_address,
+ struct device *dev);
/**
* struct ttm_page_alloc_func member alloc_init.
*
@@ -97,12 +101,14 @@ extern struct ttm_page_alloc_func ttm_page_alloc_default;
* @cstate: ttm caching state for the page.
* @count: number of pages to allocate.
* @dma_address: The DMA (bus) address of pages - (by default zero).
+ * @dev: The device that needs this.
*/
int ttm_get_pages(struct list_head *pages,
int flags,
enum ttm_caching_state cstate,
unsigned count,
- dma_addr_t *dma_address);
+ dma_addr_t *dma_address,
+ struct device *dev);
/**
* Put linked list of pages to pool.
*
@@ -112,12 +118,14 @@ int ttm_get_pages(struct list_head *pages,
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state.
* @dma_address: The DMA (bus) address of pages (by default zero).
+ * @dev: The device that needs this.
*/
void ttm_put_pages(struct list_head *pages,
unsigned page_count,
int flags,
enum ttm_caching_state cstate,
- dma_addr_t *dma_address);
+ dma_addr_t *dma_address,
+ struct device *dev);
/**
* Initialize pool allocator.
*/
--
1.7.4.1
^ permalink raw reply related [flat|nested] 16+ messages in thread* [PATCH 4/9] swiotlb: Expose swiotlb_nr_tlb function to modules
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
` (2 preceding siblings ...)
2011-09-29 20:33 ` [PATCH 3/9] ttm: Pass in 'struct device' to TTM so it can do DMA API on behalf of device Konrad Rzeszutek Wilk
@ 2011-09-29 20:33 ` Konrad Rzeszutek Wilk
2011-09-29 23:49 ` FUJITA Tomonori
2011-09-29 20:33 ` [PATCH 5/9] ttm: Provide a DMA aware TTM page pool code Konrad Rzeszutek Wilk
` (5 subsequent siblings)
9 siblings, 1 reply; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-29 20:33 UTC (permalink / raw)
To: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher
Cc: xen-devel, Konrad Rzeszutek Wilk, FUJITA Tomonori
As a mechanism to detect whether SWIOTLB is enabled or not.
We also fix the spelling - it was swioltb instead of
swiotlb.
CC: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
[v1: Ripped out swiotlb_enabled]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/xen/swiotlb-xen.c | 2 +-
include/linux/swiotlb.h | 2 +-
lib/swiotlb.c | 5 +++--
3 files changed, 5 insertions(+), 4 deletions(-)
diff --git a/drivers/xen/swiotlb-xen.c b/drivers/xen/swiotlb-xen.c
index 6e8c15a..cbcd8cc 100644
--- a/drivers/xen/swiotlb-xen.c
+++ b/drivers/xen/swiotlb-xen.c
@@ -149,7 +149,7 @@ void __init xen_swiotlb_init(int verbose)
int rc;
unsigned long nr_tbl;
- nr_tbl = swioltb_nr_tbl();
+ nr_tbl = swiotlb_nr_tbl();
if (nr_tbl)
xen_io_tlb_nslabs = nr_tbl;
else {
diff --git a/include/linux/swiotlb.h b/include/linux/swiotlb.h
index 445702c..e872526 100644
--- a/include/linux/swiotlb.h
+++ b/include/linux/swiotlb.h
@@ -24,7 +24,7 @@ extern int swiotlb_force;
extern void swiotlb_init(int verbose);
extern void swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose);
-extern unsigned long swioltb_nr_tbl(void);
+extern unsigned long swiotlb_nr_tbl(void);
/*
* Enumeration for sync targets
diff --git a/lib/swiotlb.c b/lib/swiotlb.c
index 99093b3..058935e 100644
--- a/lib/swiotlb.c
+++ b/lib/swiotlb.c
@@ -110,11 +110,11 @@ setup_io_tlb_npages(char *str)
__setup("swiotlb=", setup_io_tlb_npages);
/* make io_tlb_overflow tunable too? */
-unsigned long swioltb_nr_tbl(void)
+unsigned long swiotlb_nr_tbl(void)
{
return io_tlb_nslabs;
}
-
+EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
/* Note that this doesn't work with highmem page */
static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
volatile void *address)
@@ -321,6 +321,7 @@ void __init swiotlb_free(void)
free_bootmem_late(__pa(io_tlb_start),
PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
}
+ io_tlb_nslabs = 0;
}
static int is_swiotlb_buffer(phys_addr_t paddr)
--
1.7.4.1
^ permalink raw reply related [flat|nested] 16+ messages in thread* Re: [PATCH 4/9] swiotlb: Expose swiotlb_nr_tlb function to modules
2011-09-29 20:33 ` [PATCH 4/9] swiotlb: Expose swiotlb_nr_tlb function to modules Konrad Rzeszutek Wilk
@ 2011-09-29 23:49 ` FUJITA Tomonori
0 siblings, 0 replies; 16+ messages in thread
From: FUJITA Tomonori @ 2011-09-29 23:49 UTC (permalink / raw)
To: konrad.wilk
Cc: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher, xen-devel, fujita.tomonori
On Thu, 29 Sep 2011 16:33:47 -0400
Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:
> As a mechanism to detect whether SWIOTLB is enabled or not.
> We also fix the spelling - it was swioltb instead of
> swiotlb.
>
> CC: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
> [v1: Ripped out swiotlb_enabled]
> Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> ---
> drivers/xen/swiotlb-xen.c | 2 +-
> include/linux/swiotlb.h | 2 +-
> lib/swiotlb.c | 5 +++--
> 3 files changed, 5 insertions(+), 4 deletions(-)
Acked-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
^ permalink raw reply [flat|nested] 16+ messages in thread
* [PATCH 5/9] ttm: Provide a DMA aware TTM page pool code.
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
` (3 preceding siblings ...)
2011-09-29 20:33 ` [PATCH 4/9] swiotlb: Expose swiotlb_nr_tlb function to modules Konrad Rzeszutek Wilk
@ 2011-09-29 20:33 ` Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 6/9] ttm: Add 'no_dma' parameter to turn the TTM DMA pool off during runtime Konrad Rzeszutek Wilk
` (4 subsequent siblings)
9 siblings, 0 replies; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-29 20:33 UTC (permalink / raw)
To: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher
Cc: xen-devel, Konrad Rzeszutek Wilk
In TTM world the pages for the graphic drivers are kept in three different
pools: write combined, uncached, and cached (write-back). When the pages
are used by the graphic driver the graphic adapter via its built in MMU
(or AGP) programs these pages in. The programming requires the virtual address
(from the graphic adapter perspective) and the physical address (either System RAM
or the memory on the card) which is obtained using the pci_map_* calls (which does the
virtual to physical - or bus address translation). During the graphic application's
"life" those pages can be shuffled around, swapped out to disk, moved from the
VRAM to System RAM or vice-versa. This all works with the existing TTM pool code
- except when we want to use the software IOTLB (SWIOTLB) code to "map" the physical
addresses to the graphic adapter MMU. We end up programming the bounce buffer's
physical address instead of the TTM pool memory's and get a non-worky driver.
There are two solutions:
1) using the DMA API to allocate pages that are screened by the DMA API, or
2) using the pci_sync_* calls to copy the pages from the bounce-buffer and back.
This patch fixes the issue by allocating pages using the DMA API. The second
is a viable option - but it has performance drawbacks and potential correctness
issues - think of the write cache page being bounced (SWIOTLB->TTM), the
WC is set on the TTM page and the copy from SWIOTLB not making it to the TTM
page until the page has been recycled in the pool (and used by another application).
The bounce buffer does not get activated often - only in cases where we have
a 32-bit capable card and we want to use a page that is allocated above the
4GB limit. The bounce buffer offers the solution of copying the contents
of that 4GB page to an location below 4GB and then back when the operation has been
completed (or vice-versa). This is done by using the 'pci_sync_*' calls.
Note: If you look carefully enough in the existing TTM page pool code you will
notice the GFP_DMA32 flag is used - which should guarantee that the provided page
is under 4GB. It certainly is the case, except this gets ignored in two cases:
- If user specifies 'swiotlb=force' which bounces _every_ page.
- If user is using a Xen's PV Linux guest (which uses the SWIOTLB and the
underlaying PFN's aren't necessarily under 4GB).
To not have this extra copying done the other option is to allocate the pages
using the DMA API so that there is not need to map the page and perform the
expensive 'pci_sync_*' calls.
This DMA API capable TTM pool requires for this the 'struct device' to
properly call the DMA API. It also has to track the virtual and bus address of
the page being handed out in case it ends up being swapped out or de-allocated -
to make sure it is de-allocated using the proper's 'struct device'.
Implementation wise the code keeps two lists: one that is attached to the
'struct device' (via the dev->dma_pools list) and a global one to be used when
the 'struct device' is unavailable (think shrinker code). The global list can
iterate over all of the 'struct device' and its associated dma_pool. The list
in dev->dma_pools can only iterate the device's dma_pool.
/[struct device_pool]\
/---------------------------------------------------| dev |
/ +-------| dma_pool |
/-----+------\ / \--------------------/
|struct device| /-->[struct dma_pool for WC]</ /[struct device_pool]\
| dma_pools +----+ /-| dev |
| ... | \--->[struct dma_pool for uncached]<-/--| dma_pool |
\-----+------/ / \--------------------/
\----------------------------------------------/
[Two pools associated with the device (WC and UC), and the parallel list
containing the 'struct dev' and 'struct dma_pool' entries]
The maximum amount of dma pools a device can have is six: write-combined,
uncached, and cached; then there are the DMA32 variants which are:
write-combined dma32, uncached dma32, and cached dma32.
Currently this code only gets activated when any variant of the SWIOTLB IOMMU
code is running (Intel without VT-d, AMD without GART, IBM Calgary and Xen PV
with PCI devices).
Tested-by: Michel Dänzer <michel@daenzer.net>
[v1: Using swiotlb_nr_tbl instead of swiotlb_enabled]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/ttm/Makefile | 3 +
drivers/gpu/drm/ttm/ttm_memory.c | 2 +
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 1313 ++++++++++++++++++++++++++++++
include/drm/ttm/ttm_page_alloc.h | 32 +-
4 files changed, 1346 insertions(+), 4 deletions(-)
create mode 100644 drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
diff --git a/drivers/gpu/drm/ttm/Makefile b/drivers/gpu/drm/ttm/Makefile
index f3cf6f0..8300bc0 100644
--- a/drivers/gpu/drm/ttm/Makefile
+++ b/drivers/gpu/drm/ttm/Makefile
@@ -7,4 +7,7 @@ ttm-y := ttm_agp_backend.o ttm_memory.o ttm_tt.o ttm_bo.o \
ttm_object.o ttm_lock.o ttm_execbuf_util.o ttm_page_alloc.o \
ttm_bo_manager.o
+ifeq ($(CONFIG_SWIOTLB),y)
+ttm-y += ttm_page_alloc_dma.o
+endif
obj-$(CONFIG_DRM_TTM) += ttm.o
diff --git a/drivers/gpu/drm/ttm/ttm_memory.c b/drivers/gpu/drm/ttm/ttm_memory.c
index c7d97a5..57323fe 100644
--- a/drivers/gpu/drm/ttm/ttm_memory.c
+++ b/drivers/gpu/drm/ttm/ttm_memory.c
@@ -397,6 +397,8 @@ int ttm_mem_global_init(struct ttm_mem_global *glob)
zone->name, (unsigned long long) zone->max_mem >> 10);
}
ttm_page_alloc = &ttm_page_alloc_default;
+ if (ttm_page_alloc_need_dma())
+ printk(KERN_INFO TTM_PFX "Using DMA aware pool.\n");
ttm_page_alloc_init(glob, glob->zone_kernel->max_mem/(2*PAGE_SIZE));
return 0;
out_no_zone:
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
new file mode 100644
index 0000000..5a5739c
--- /dev/null
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
@@ -0,0 +1,1313 @@
+/*
+ * Copyright 2011 (c) Oracle Corp.
+
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sub license,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ *
+ * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
+ */
+
+/*
+ * A simple DMA pool losely based on dmapool.c. It has certain advantages
+ * over the DMA pools:
+ * - Pool collects resently freed pages for reuse (and hooks up to
+ * the shrinker).
+ * - Tracks currently in use pages
+ * - Tracks whether the page is UC, WB or cached (and reverts to WB
+ * when freed).
+ */
+
+#include <linux/dma-mapping.h>
+#include <linux/list.h>
+#include <linux/seq_file.h> /* for seq_printf */
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/highmem.h>
+#include <linux/mm_types.h>
+#include <linux/module.h>
+#include <linux/mm.h>
+#include <linux/atomic.h>
+#include <linux/device.h>
+#include <linux/kthread.h>
+#include "ttm/ttm_bo_driver.h"
+#include "ttm/ttm_page_alloc.h"
+#ifdef TTM_HAS_AGP
+#include <asm/agp.h>
+#endif
+
+#define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
+#define SMALL_ALLOCATION 16
+#define FREE_ALL_PAGES (~0U)
+/* times are in msecs */
+#define IS_UNDEFINED (0)
+#define IS_WC (1<<1)
+#define IS_UC (1<<2)
+#define IS_CACHED (1<<3)
+#define IS_DMA32 (1<<4)
+
+enum pool_type {
+ POOL_IS_UNDEFINED,
+ POOL_IS_WC = IS_WC,
+ POOL_IS_UC = IS_UC,
+ POOL_IS_CACHED = IS_CACHED,
+ POOL_IS_WC_DMA32 = IS_WC | IS_DMA32,
+ POOL_IS_UC_DMA32 = IS_UC | IS_DMA32,
+ POOL_IS_CACHED_DMA32 = IS_CACHED | IS_DMA32,
+};
+/*
+ * The pool structure. There are usually six pools:
+ * - generic (not restricted to DMA32):
+ * - write combined, uncached, cached.
+ * - dma32 (up to 2^32 - so up 4GB):
+ * - write combined, uncached, cached.
+ * for each 'struct device'. The 'cached' is for pages that are actively used.
+ * The other ones can be shrunk by the shrinker API if neccessary.
+ * @pools: The 'struct device->dma_pools' link.
+ * @type: Type of the pool
+ * @lock: Protects the page_list from concurrnet access. Must be used with
+ * irqsave/irqrestore variants because pool allocator maybe called from
+ * delayed work.
+ * @page_list: Pool of all pages (both in use and free).
+ * @dev: The device that is associated with these pools.
+ * @size: Size used during DMA allocation.
+ * @npages_free: Count of available pages for re-use.
+ * @npages_in_use: Count of pages that are in use (each of them
+ * is marked in_use.
+ * @nfrees: Stats when pool is shrinking.
+ * @nrefills: Stats when the pool is grown.
+ * @gfp_flags: Flags to pass for alloc_page.
+ * @name: Name of the pool.
+ * @dev_name: Name derieved from dev - similar to how dev_info works.
+ * Used during shutdown as the dev_info during release is unavailable.
+ */
+struct dma_pool {
+ struct list_head pools; /* The 'struct device->dma_pools link */
+ enum pool_type type;
+ spinlock_t lock;
+ struct list_head page_list;
+ struct device *dev;
+ unsigned size;
+ unsigned npages_free;
+ unsigned npages_in_use;
+ unsigned long nfrees; /* Stats when shrunk. */
+ unsigned long nrefills; /* Stats when grown. */
+ gfp_t gfp_flags;
+ char name[13]; /* "cached dma32" */
+ char dev_name[64]; /* Constructed from dev */
+};
+
+/*
+ * The accounting page keeping track of the allocated page along with
+ * the DMA address.
+ * @page_list: The link to the 'page_list' in 'struct dma_pool'.
+ * @vaddr: The virtual address of the page
+ * @dma: The bus address of the page. If the page is not allocated
+ * via the DMA API, it will be -1.
+ * @in_use: Set to true if in use. Should not be freed.
+ */
+struct dma_page {
+ struct list_head page_list;
+ void *vaddr;
+ dma_addr_t dma;
+ unsigned int in_use:1;
+};
+
+/*
+ * Limits for the pool. They are handled without locks because only place where
+ * they may change is in sysfs store. They won't have immediate effect anyway
+ * so forcing serialization to access them is pointless.
+ */
+
+struct ttm_pool_opts {
+ unsigned alloc_size;
+ unsigned max_size;
+ unsigned small;
+};
+
+/*
+ * Contains the list of all of the 'struct device' and their corresponding
+ * DMA pools. Guarded by _mutex->lock.
+ * @pools: The link to 'struct ttm_pool_manager->pools'
+ * @dev: The 'struct device' associated with the 'pool'
+ * @pool: The 'struct dma_pool' associated with the 'dev'
+ */
+struct device_pools {
+ struct list_head pools;
+ struct device *dev;
+ struct dma_pool *pool;
+};
+
+/*
+ * struct ttm_pool_manager - Holds memory pools for fast allocation
+ *
+ * @lock: Lock used when adding/removing from pools
+ * @pools: List of 'struct device' and 'struct dma_pool' tuples.
+ * @options: Limits for the pool.
+ * @npools: Total amount of pools in existence.
+ * @shrinker: The structure used by [un|]register_shrinker
+ */
+struct ttm_pool_manager {
+ struct mutex lock;
+ struct list_head pools;
+ struct ttm_pool_opts options;
+ unsigned npools;
+ struct shrinker mm_shrink;
+ struct kobject kobj;
+};
+
+static struct ttm_pool_manager *_manager;
+
+static struct attribute ttm_page_pool_max = {
+ .name = "pool_max_size",
+ .mode = S_IRUGO | S_IWUSR
+};
+static struct attribute ttm_page_pool_small = {
+ .name = "pool_small_allocation",
+ .mode = S_IRUGO | S_IWUSR
+};
+static struct attribute ttm_page_pool_alloc_size = {
+ .name = "pool_allocation_size",
+ .mode = S_IRUGO | S_IWUSR
+};
+
+static struct attribute *ttm_pool_attrs[] = {
+ &ttm_page_pool_max,
+ &ttm_page_pool_small,
+ &ttm_page_pool_alloc_size,
+ NULL
+};
+
+static void ttm_pool_kobj_release(struct kobject *kobj)
+{
+ struct ttm_pool_manager *m =
+ container_of(kobj, struct ttm_pool_manager, kobj);
+ kfree(m);
+}
+
+static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
+ const char *buffer, size_t size)
+{
+ struct ttm_pool_manager *m =
+ container_of(kobj, struct ttm_pool_manager, kobj);
+ int chars;
+ unsigned val;
+ chars = sscanf(buffer, "%u", &val);
+ if (chars == 0)
+ return size;
+
+ /* Convert kb to number of pages */
+ val = val / (PAGE_SIZE >> 10);
+
+ if (attr == &ttm_page_pool_max)
+ m->options.max_size = val;
+ else if (attr == &ttm_page_pool_small)
+ m->options.small = val;
+ else if (attr == &ttm_page_pool_alloc_size) {
+ if (val > NUM_PAGES_TO_ALLOC*8) {
+ printk(KERN_ERR TTM_PFX
+ "Setting allocation size to %lu "
+ "is not allowed. Recommended size is "
+ "%lu\n",
+ NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
+ NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
+ return size;
+ } else if (val > NUM_PAGES_TO_ALLOC) {
+ printk(KERN_WARNING TTM_PFX
+ "Setting allocation size to "
+ "larger than %lu is not recommended.\n",
+ NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
+ }
+ m->options.alloc_size = val;
+ }
+
+ return size;
+}
+
+static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
+ char *buffer)
+{
+ struct ttm_pool_manager *m =
+ container_of(kobj, struct ttm_pool_manager, kobj);
+ unsigned val = 0;
+
+ if (attr == &ttm_page_pool_max)
+ val = m->options.max_size;
+ else if (attr == &ttm_page_pool_small)
+ val = m->options.small;
+ else if (attr == &ttm_page_pool_alloc_size)
+ val = m->options.alloc_size;
+
+ val = val * (PAGE_SIZE >> 10);
+
+ return snprintf(buffer, PAGE_SIZE, "%u\n", val);
+}
+
+static const struct sysfs_ops ttm_pool_sysfs_ops = {
+ .show = &ttm_pool_show,
+ .store = &ttm_pool_store,
+};
+
+static struct kobj_type ttm_pool_kobj_type = {
+ .release = &ttm_pool_kobj_release,
+ .sysfs_ops = &ttm_pool_sysfs_ops,
+ .default_attrs = ttm_pool_attrs,
+};
+
+#ifndef CONFIG_X86
+static int set_pages_array_wb(struct page **pages, int addrinarray)
+{
+#ifdef TTM_HAS_AGP
+ int i;
+
+ for (i = 0; i < addrinarray; i++)
+ unmap_page_from_agp(pages[i]);
+#endif
+ return 0;
+}
+
+static int set_pages_array_wc(struct page **pages, int addrinarray)
+{
+#ifdef TTM_HAS_AGP
+ int i;
+
+ for (i = 0; i < addrinarray; i++)
+ map_page_into_agp(pages[i]);
+#endif
+ return 0;
+}
+
+static int set_pages_array_uc(struct page **pages, int addrinarray)
+{
+#ifdef TTM_HAS_AGP
+ int i;
+
+ for (i = 0; i < addrinarray; i++)
+ map_page_into_agp(pages[i]);
+#endif
+ return 0;
+}
+#endif /* for !CONFIG_X86 */
+
+static int ttm_set_pages_caching(struct dma_pool *pool,
+ struct page **pages, unsigned cpages)
+{
+ int r = 0;
+ /* Set page caching */
+ if (pool->type & IS_UC) {
+ r = set_pages_array_uc(pages, cpages);
+ if (r)
+ pr_err(TTM_PFX
+ "%s: Failed to set %d pages to uc!\n",
+ pool->dev_name, cpages);
+ }
+ if (pool->type & IS_WC) {
+ r = set_pages_array_wc(pages, cpages);
+ if (r)
+ pr_err(TTM_PFX
+ "%s: Failed to set %d pages to wc!\n",
+ pool->dev_name, cpages);
+ }
+ return r;
+}
+
+static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
+{
+ dma_addr_t dma = d_page->dma;
+
+ pr_debug("%s: (%s:%d) Freeing %p (%p) (DMA:0x%lx)\n",
+ pool->dev_name, pool->name, current->pid, d_page->vaddr,
+ virt_to_page(d_page->vaddr), (unsigned long)dma);
+
+ dma_free_coherent(pool->dev, pool->size, d_page->vaddr, dma);
+
+ kfree(d_page);
+ d_page = NULL;
+}
+static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
+{
+ struct dma_page *d_page;
+
+ d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
+ if (!d_page)
+ return NULL;
+
+ d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size,
+ &d_page->dma,
+ pool->gfp_flags);
+ if (d_page->vaddr) {
+ pr_debug("%s: (%s:%d) Allocated %p (%p) (DMA:0x%lx)\n",
+ pool->dev_name, pool->name, current->pid, d_page->vaddr,
+ virt_to_page(d_page->vaddr),
+ (unsigned long)d_page->dma);
+ d_page->in_use = 0;
+ } else {
+ kfree(d_page);
+ d_page = NULL;
+ }
+
+ return d_page;
+}
+static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
+{
+ enum pool_type type = IS_UNDEFINED;
+
+ if (flags & TTM_PAGE_FLAG_DMA32)
+ type |= IS_DMA32;
+ if (cstate == tt_cached)
+ type |= IS_CACHED;
+ else if (cstate == tt_uncached)
+ type |= IS_UC;
+ else
+ type |= IS_WC;
+
+ return type;
+}
+static void ttm_pool_update_inuse(struct dma_pool *pool, unsigned count)
+{
+ unsigned long irq_flags;
+
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ pool->npages_free += count;
+ pool->npages_in_use -= count;
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+}
+static void ttm_pool_update_free_locked(struct dma_pool *pool,
+ unsigned freed_pages)
+{
+ pool->npages_free -= freed_pages;
+ pool->nfrees += freed_pages;
+
+}
+/* set memory back to wb and free the pages. */
+static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
+ struct page *pages[], unsigned npages)
+{
+ struct dma_page *d_page, *tmp;
+
+ if (npages && set_pages_array_wb(pages, npages))
+ pr_err(TTM_PFX "%s: Failed to set %d pages to wb!\n",
+ pool->dev_name, npages);
+
+ pr_debug("%s: (%s:%d) Freeing %d pages at once.\n",
+ pool->dev_name, pool->name, current->pid, npages);
+
+ list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
+ list_del(&d_page->page_list);
+ __ttm_dma_free_page(pool, d_page);
+ }
+}
+/*
+ * Free pages from pool.
+ *
+ * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
+ * number of pages in one go.
+ *
+ * @pool: to free the pages from
+ * @nr_free: If set to true will free all pages in pool
+ **/
+static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free)
+{
+ unsigned long irq_flags;
+ struct dma_page *dma_p, *tmp;
+ struct page **pages_to_free;
+ struct list_head d_pages;
+ unsigned freed_pages = 0,
+ npages_to_free = nr_free;
+
+ if (NUM_PAGES_TO_ALLOC < nr_free)
+ npages_to_free = NUM_PAGES_TO_ALLOC;
+
+ pages_to_free = kmalloc(npages_to_free * sizeof(struct dma_page *),
+ GFP_KERNEL);
+
+ if (!pages_to_free) {
+ pr_err(TTM_PFX
+ "%s: Failed to allocate memory for pool free operation.\n",
+ pool->dev_name);
+ return 0;
+ }
+ INIT_LIST_HEAD(&d_pages);
+restart:
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ list_for_each_entry_safe_reverse(dma_p, tmp, &pool->page_list,
+ page_list) {
+ if (freed_pages >= npages_to_free)
+ break;
+
+ if (dma_p->in_use)
+ continue;
+
+ pr_debug("%s: (%s:%d) %p (%p) (DMA:0x%lx) is expunged.\n",
+ pool->dev_name, pool->name, current->pid, dma_p->vaddr,
+ virt_to_page(dma_p->vaddr),
+ (unsigned long)dma_p->dma);
+
+ /* Move the dma_page from one list to another. */
+ list_del(&dma_p->page_list);
+ list_add(&dma_p->page_list, &d_pages);
+
+ pages_to_free[freed_pages++] = virt_to_page(dma_p->vaddr);
+ /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
+ if (freed_pages >= NUM_PAGES_TO_ALLOC) {
+
+ ttm_pool_update_free_locked(pool, freed_pages);
+ /**
+ * Because changing page caching is costly
+ * we unlock the pool to prevent stalling.
+ */
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+
+ ttm_dma_pages_put(pool, &d_pages, pages_to_free,
+ freed_pages);
+
+ INIT_LIST_HEAD(&d_pages);
+
+ if (likely(nr_free != FREE_ALL_PAGES))
+ nr_free -= freed_pages;
+
+ if (NUM_PAGES_TO_ALLOC >= nr_free)
+ npages_to_free = nr_free;
+ else
+ npages_to_free = NUM_PAGES_TO_ALLOC;
+
+ freed_pages = 0;
+
+ /* free all so restart the processing */
+ if (nr_free)
+ goto restart;
+
+ /* Not allowed to fall tough or break because
+ * following context is inside spinlock while we are
+ * outside here.
+ */
+ goto out;
+
+ }
+ }
+
+ /* remove range of pages from the pool */
+ if (freed_pages) {
+ ttm_pool_update_free_locked(pool, freed_pages);
+ nr_free -= freed_pages;
+ }
+
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+
+ if (freed_pages)
+ ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
+out:
+ kfree(pages_to_free);
+ return nr_free;
+}
+
+static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
+{
+ struct device_pools *p;
+ struct dma_pool *pool;
+ struct dma_page *d_page, *d_tmp;
+
+ if (!dev)
+ return;
+
+ mutex_lock(&_manager->lock);
+ list_for_each_entry_reverse(p, &_manager->pools, pools) {
+ if (p->dev != dev)
+ continue;
+ pool = p->pool;
+ if (pool->type != type)
+ continue;
+ pr_debug("%s: (%s:%d) of device pool freed "\
+ "(has %d free, and %d in use).\n",
+ pool->dev_name, pool->name, current->pid,
+ pool->npages_free, pool->npages_in_use);
+ list_del(&p->pools);
+ kfree(p);
+ _manager->npools--;
+ break;
+ }
+ list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
+ unsigned long irq_save;
+ if (pool->type != type)
+ continue;
+ /* Takes a spinlock.. */
+ ttm_dma_page_pool_free(pool, FREE_ALL_PAGES);
+ /* .. but afterwards we can take it too */
+ spin_lock_irqsave(&pool->lock, irq_save);
+ list_for_each_entry_safe(d_page, d_tmp, &pool->page_list,
+ page_list) {
+ if (d_page->in_use) {
+ pr_err("%s: (%s:%d) %p (%p DMA:0x%lx) busy!\n",
+ pool->dev_name, pool->name,
+ current->pid, d_page->vaddr,
+ virt_to_page(d_page->vaddr),
+ (unsigned long)d_page->dma);
+ list_del(&d_page->page_list);
+ kfree(d_page);
+ pool->npages_in_use--;
+ }
+ }
+ spin_unlock_irqrestore(&pool->lock, irq_save);
+ WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
+ /* This code path is called after _all_ references to the
+ * struct device has been dropped - so nobody should be
+ * touching it. In case somebody is trying to _add_ we are
+ * guarded by the mutex. */
+ list_del(&pool->pools);
+ kfree(pool);
+ break;
+ }
+ mutex_unlock(&_manager->lock);
+}
+/*
+ * On free-ing of the 'struct device' this deconstructor is run.
+ * Albeit the pool might have already been freed earlier.
+ */
+static void ttm_dma_pool_release(struct device *dev, void *res)
+{
+ struct dma_pool *pool = *(struct dma_pool **)res;
+
+ if (pool)
+ ttm_dma_free_pool(dev, pool->type);
+}
+
+static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
+{
+ return *(struct dma_pool **)res == match_data;
+}
+
+static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
+ enum pool_type type)
+{
+ char *n[] = {"wc", "uc", "cached", " dma32", "unknown",};
+ enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED};
+ struct device_pools *sec_pool = NULL;
+ struct dma_pool *pool = NULL, **ptr;
+ unsigned i;
+ int ret = -ENODEV;
+ char *p;
+
+ if (!dev)
+ return NULL;
+
+ ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
+ if (!ptr)
+ return NULL;
+
+ ret = -ENOMEM;
+
+ pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
+ dev_to_node(dev));
+ if (!pool)
+ goto err_mem;
+
+ sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
+ dev_to_node(dev));
+ if (!sec_pool)
+ goto err_mem;
+
+ INIT_LIST_HEAD(&sec_pool->pools);
+ sec_pool->dev = dev;
+ sec_pool->pool = pool;
+
+ INIT_LIST_HEAD(&pool->page_list);
+ INIT_LIST_HEAD(&pool->pools);
+ spin_lock_init(&pool->lock);
+ pool->dev = dev;
+ pool->npages_free = pool->npages_in_use = 0;
+ pool->nfrees = 0;
+ pool->gfp_flags = flags;
+ pool->size = PAGE_SIZE;
+ pool->type = type;
+ pool->nrefills = 0;
+ p = pool->name;
+ for (i = 0; i < 5; i++) {
+ if (type & t[i]) {
+ p += snprintf(p, sizeof(pool->name) - (p - pool->name),
+ "%s", n[i]);
+ }
+ }
+ *p = 0;
+ /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
+ * - the kobj->name has already been deallocated.*/
+ snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
+ dev_driver_string(dev), dev_name(dev));
+ mutex_lock(&_manager->lock);
+ /* You can get the dma_pool from either the global: */
+ list_add(&sec_pool->pools, &_manager->pools);
+ _manager->npools++;
+ /* or from 'struct device': */
+ list_add(&pool->pools, &dev->dma_pools);
+ mutex_unlock(&_manager->lock);
+
+ *ptr = pool;
+ devres_add(dev, ptr);
+
+ return pool;
+err_mem:
+ devres_free(ptr);
+ kfree(sec_pool);
+ kfree(pool);
+ return ERR_PTR(ret);
+}
+static struct dma_pool *ttm_dma_find_pool(struct device *dev,
+ enum pool_type type)
+{
+ struct dma_pool *pool, *tmp, *found = NULL;
+
+ if (type == IS_UNDEFINED)
+ return found;
+ /* NB: We iterate on the 'struct dev' which has no spinlock, but
+ * it does have a kref which we have taken. */
+ list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) {
+ if (pool->type != type)
+ continue;
+ found = pool;
+ }
+ if (found)
+ pr_debug("%s: (%s:%d) Found. It has %d free pages (%d in use)\n",
+ found->dev_name, found->name, current->pid,
+ found->npages_free, found->npages_in_use);
+ return found;
+}
+
+/*
+ * Free pages the pages that failed to change the caching state. If there is
+ * any pages that have changed their caching state already put them to the
+ * pool.
+ */
+static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
+ struct page **failed_pages,
+ unsigned cpages)
+{
+ unsigned long irq_flags;
+ unsigned i;
+ struct dma_page *dma_p, *t;
+ struct list_head d_pages;
+
+ /* Failed pages have to be freed */
+ i = cpages;
+
+ INIT_LIST_HEAD(&d_pages);
+
+ /* To make it faster we only take the spinlock on list
+ * removal, and later on do the free-ing at our leisure. */
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ list_for_each_entry_safe(dma_p, t, &pool->page_list, page_list) {
+ struct page *p = failed_pages[i];
+ if (virt_to_page(dma_p->vaddr) != p) {
+ pr_debug("%s: (%s:%d) Skipping %p (%p) (DMA:0x%lx)\n",
+ pool->dev_name, pool->name, current->pid,
+ dma_p->vaddr,
+ virt_to_page(dma_p->vaddr),
+ (unsigned long)dma_p->dma);
+ continue;
+ }
+ list_del(&dma_p->page_list);
+ list_add(&dma_p->page_list, &d_pages);
+ list_del(&failed_pages[i]->lru);
+ if (--i == 0)
+ break;
+ }
+ ttm_pool_update_free_locked(pool, (cpages - i));
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+
+ ttm_dma_pages_put(pool, &d_pages, NULL/* Don't try to set WB on them */,
+ cpages - i);
+}
+
+static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
+ struct list_head *pages,
+ dma_addr_t *dma_address,
+ unsigned dma_offset, unsigned count)
+{
+ struct page **caching_array;
+ struct dma_page *dma_p;
+ struct page *p;
+ int r = 0;
+ unsigned i, cpages;
+ unsigned long irq_flags;
+ unsigned max_cpages = min(count,
+ (unsigned)(PAGE_SIZE/sizeof(struct page *)));
+
+ /* allocate array for page caching change */
+ caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
+
+ if (!caching_array) {
+ pr_err(TTM_PFX
+ "%s: Unable to allocate table for new pages.",
+ pool->dev_name);
+ return -ENOMEM;
+ }
+ pr_debug("%s: (%s:%d) Getting %d pages @ %d idx\n",
+ pool->dev_name, pool->name, current->pid,
+ count, dma_offset);
+ for (i = 0, cpages = 0; i < count; ++i) {
+ dma_p = __ttm_dma_alloc_page(pool);
+ if (!dma_p) {
+ pr_err(TTM_PFX "%s: Unable to get page %u.\n",
+ pool->dev_name, i);
+
+ /* store already allocated pages in the pool after
+ * setting the caching state */
+ if (cpages) {
+ r = ttm_set_pages_caching(pool, caching_array,
+ cpages);
+ if (r)
+ ttm_dma_handle_caching_state_failure(
+ pool, caching_array, cpages);
+ }
+ r = -ENOMEM;
+ goto out;
+ }
+ p = virt_to_page(dma_p->vaddr);
+ /* Take ownership of that page. */
+ dma_p->in_use = true;
+ /* And now add it in without having to worry about it being
+ * immediately picked up by another thread. */
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ list_add(&dma_p->page_list, &pool->page_list);
+ pool->npages_in_use++;
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+#ifdef CONFIG_HIGHMEM
+ /* gfp flags of highmem page should never be dma32 so we
+ * we should be fine in such case
+ */
+ if (!PageHighMem(p))
+#endif
+ {
+ caching_array[cpages++] = p;
+ if (cpages == max_cpages) {
+
+ r = ttm_set_pages_caching(pool, caching_array,
+ cpages);
+ if (r) {
+ ttm_dma_handle_caching_state_failure(
+ pool, caching_array, cpages);
+ goto out;
+ }
+ cpages = 0;
+ }
+ }
+ /* Note: We do _not_ add the pages to the cached pool here. */
+ list_add_tail(&p->lru, pages);
+ dma_address[dma_offset + i] = dma_p->dma;
+ }
+
+ if (cpages) {
+ r = ttm_set_pages_caching(pool, caching_array, cpages);
+ if (r)
+ ttm_dma_handle_caching_state_failure(pool,
+ caching_array, cpages);
+ }
+out:
+ kfree(caching_array);
+ return r;
+}
+/*
+ * Recycle (or delete) the 'pages' that are on the 'pool'.
+ * @pool: The pool that the pages are associated with.
+ * @pages: The list of pages we are done with.
+ * @page_count: Count of how many pages (or zero if all).
+ * @erase: Instead of recycling - just free them.
+ */
+static int ttm_dma_put_pages_in_pool(struct dma_pool *pool,
+ struct list_head *pages,
+ unsigned page_count,
+ bool erase)
+{
+ unsigned long uninitialized_var(irq_flags);
+ struct list_head uninitialized_var(d_pages);
+ struct page **uninitialized_var(pages_to_free);
+ unsigned uninitialized_var(freed_pages);
+ struct dma_page *d_page, *d_tmp;
+ struct page *p, *tmp;
+ bool found = false;
+ unsigned count = 0;
+
+ if (list_empty(pages))
+ return 0;
+
+ if (page_count == 0) {
+ list_for_each_entry_safe(p, tmp, pages, lru)
+ ++page_count;
+ }
+ pr_debug("%s: (%s:%d) %s %d pages\n",
+ pool->dev_name, pool->name, current->pid,
+ erase ? "Destroying" : "Recycling", page_count);
+
+ if (erase) {
+ INIT_LIST_HEAD(&d_pages);
+ pages_to_free = kmalloc(page_count * sizeof(struct dma_page *),
+ GFP_KERNEL);
+ if (!pages_to_free) {
+ dev_err(pool->dev, TTM_PFX
+ "Failed to allocate memory for pool free operation.\n");
+ return 0;
+ }
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ freed_pages = 0;
+ }
+ list_for_each_entry_safe_reverse(p, tmp, pages, lru) {
+ found = false;
+ /* We only hold the lock when erasing. Otherwise we just
+ * set the d_page->in_use bit. */
+ list_for_each_entry_safe(d_page, d_tmp, &pool->page_list,
+ page_list) {
+ if (p != virt_to_page(d_page->vaddr))
+ continue;
+ found = true;
+ break;
+ }
+ if (!found)
+ break; /* We could continue, but why bother..*/
+
+ WARN_ON(!d_page->in_use);
+ d_page->in_use = false;
+ count++;
+ list_del_init(&p->lru);
+ if (erase) {
+ list_del(&d_page->page_list);
+ list_add(&d_page->page_list, &d_pages);
+ pages_to_free[freed_pages++] =
+ virt_to_page(d_page->vaddr);
+ }
+ /* Do not advance past what we were asked to delete. */
+ if (count == page_count)
+ break;
+ }
+ if (erase) {
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+ ttm_dma_pages_put(pool, &d_pages, pages_to_free /* to set WB */,
+ freed_pages);
+ kfree(pages_to_free);
+ }
+ pr_debug("%s: (%s:%d) Put %d/%d pages in the pool.\n",
+ pool->dev_name, pool->name, current->pid, count, page_count);
+ return count;
+}
+/*
+ * @return count of pages still required to fulfill the request.
+*/
+static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
+ struct list_head *pages,
+ dma_addr_t *dma_address,
+ unsigned count,
+ unsigned long *irq_flags)
+{
+ int r = count;
+
+ if (count < _manager->options.small &&
+ count > pool->npages_free) {
+ /* Do NOT try to get more than count. This is b/c
+ * dma_address[count++] will fail. */
+ unsigned alloc_size = min(count, _manager->options.alloc_size);
+
+ spin_unlock_irqrestore(&pool->lock, *irq_flags);
+ /* Returns how many more are neccessary to fulfill the
+ * request. */
+ r = ttm_dma_pool_alloc_new_pages(pool, pages, dma_address,
+ 0 /* no offset */, alloc_size);
+ spin_lock_irqsave(&pool->lock, *irq_flags);
+
+ if (!r) {
+ ++pool->nrefills;
+ } else {
+ pr_err(TTM_PFX "%s: Failed to fill %s pool (r:%d)!\n",
+ pool->dev_name, pool->name, r);
+ spin_unlock_irqrestore(&pool->lock, *irq_flags);
+ count = ttm_dma_put_pages_in_pool(pool, pages,
+ 0 /* no WB */,
+ false /* recycle */);
+ spin_lock_irqsave(&pool->lock, *irq_flags);
+ pool->npages_free += count;
+ pool->npages_in_use -= count;
+ }
+ }
+ return r;
+
+}
+
+/*
+ * @return count of pages still required to fulfill the request.
+ */
+static int ttm_dma_pool_get_pages(struct dma_pool *pool,
+ struct list_head *pages,
+ dma_addr_t *dma_address, unsigned count)
+{
+ unsigned long irq_flags;
+ int r = 0;
+ unsigned i;
+ struct page *p;
+ struct dma_page *dma_p;
+
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ r = ttm_dma_page_pool_fill_locked(pool, pages, dma_address,
+ count, &irq_flags);
+ pr_debug("%s: (%s:%d) Asked for %d, got %d %s.\n",
+ pool->dev_name, pool->name, current->pid, count, r,
+ (r < 0) ? "err:" : "pages");
+
+ if (r < 0)
+ goto out;
+
+ if (r > count) {
+ /* This should never happen. */
+ WARN_ON(1);
+ /* But just in case, limit it to what we requested. */
+ r = count;
+ }
+ /* How many "left" we need to pick off the free list */
+ count = r;
+ /* And in case we have gotten all the pages we need - we exit. */
+ if (count == 0)
+ goto out;
+ /* NB: Don't set r=0 at the start of the loop, otherwise you will
+ * overwrite the previously dma_address[x] fields. */
+ r = count - r;
+ i = 0;
+ pr_debug("%s: (%s:%d) Scavenging for %d pages - inserting @ %d idx " \
+ "(have %d pages free)\n",
+ pool->dev_name, pool->name, current->pid, count, r,
+ pool->npages_free);
+ /* Copy as many as we need from the pool to fulfill the request.*/
+ list_for_each_entry(dma_p, &pool->page_list, page_list) {
+ if (dma_p->in_use)
+ continue;
+ p = virt_to_page(dma_p->vaddr);
+ list_add_tail(&p->lru, pages);
+ dma_address[r++] = dma_p->dma;
+ pr_debug("%s: (%s:%d) Salvaged %p (%p) (DMA:0x%lx)\n",
+ pool->dev_name, pool->name, current->pid, dma_p->vaddr,
+ virt_to_page(dma_p->vaddr),
+ (unsigned long)dma_p->dma);
+
+ /* Take ownership of that page. */
+ dma_p->in_use = true;
+ if (++i == count)
+ break;
+ }
+ pool->npages_in_use += i;
+ pool->npages_free -= i;
+ count -= i;
+ pr_debug("%s: (%s:%d) Have taken %d pages, need %d more.\n",
+ pool->dev_name, pool->name, current->pid, r, count);
+out:
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+ return count;
+}
+/*
+ * On success pages list will hold count number of correctly
+ * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
+ */
+static int ttm_dma_get_pages(struct list_head *pages, int flags,
+ enum ttm_caching_state cstate, unsigned count,
+ dma_addr_t *dma_address, struct device *dev)
+
+{
+ int r = -ENOMEM;
+ struct dma_pool *pool;
+ gfp_t gfp_flags;
+ enum pool_type type;
+ unsigned pages_got = count;
+
+ type = ttm_to_type(flags, cstate);
+
+ if (flags & TTM_PAGE_FLAG_DMA32)
+ gfp_flags = GFP_USER | GFP_DMA32;
+ else
+ gfp_flags = GFP_HIGHUSER;
+
+ if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
+ gfp_flags |= __GFP_ZERO;
+
+ pool = ttm_dma_find_pool(dev, type);
+ if (!pool) {
+ pool = ttm_dma_pool_init(dev, gfp_flags, type);
+ if (IS_ERR_OR_NULL(pool))
+ return -ENOMEM;
+ }
+ /* Take pages out of a pool (if applicable) */
+ r = ttm_dma_pool_get_pages(pool, pages, dma_address, count);
+ /* clear the pages coming from the pool if requested */
+ if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
+ struct page *p;
+ list_for_each_entry(p, pages, lru) {
+ clear_page(page_address(p));
+ }
+ }
+ pages_got = count - r;
+ /* If pool didn't have enough pages allocate new one. */
+ if (r > 0) {
+ unsigned pages_need = r;
+ r = ttm_dma_pool_alloc_new_pages(pool, pages, dma_address,
+ pages_got /* offset in dma_address*/,
+ pages_need);
+ if (r >= 0)
+ pages_got += pages_need - r;
+
+ pr_debug("%s: (%s:%d) have %d pages, %s %d.\n",
+ pool->dev_name,
+ pool->name, current->pid, pages_got,
+ pages_got == count ?
+ "got them all" : "need more - (err):", r);
+ if (r) {
+ /* If there is any pages in the list put them back to
+ * the pool. */
+ pr_err(TTM_PFX
+ "%s: Failed to allocate extra pages "
+ "for large request.",
+ pool->dev_name);
+ count = ttm_dma_put_pages_in_pool(pool, pages,
+ 0 /* no WB */,
+ false /* recycle */);
+ INIT_LIST_HEAD(pages);
+ ttm_pool_update_inuse(pool, count);
+ return count;
+ }
+ }
+ return r;
+}
+
+/* Put all pages in pages list to correct pool to wait for reuse */
+static void ttm_dma_put_pages(struct list_head *pages, unsigned page_count,
+ int flags, enum ttm_caching_state cstate,
+ dma_addr_t *dma_address, struct device *dev)
+{
+ struct dma_pool *pool;
+ enum pool_type type;
+ bool is_cached = false;
+ unsigned count, i;
+ unsigned long irq_flags;
+
+ if (list_empty(pages))
+ return;
+
+ type = ttm_to_type(flags, cstate);
+ pool = ttm_dma_find_pool(dev, type);
+ if (!pool) {
+ WARN_ON(!pool);
+ return;
+ }
+ is_cached = (ttm_dma_find_pool(pool->dev,
+ ttm_to_type(flags, tt_cached)) == pool);
+
+ dev_dbg(pool->dev, "(%s:%d) %s %d pages.\n", pool->name, current->pid,
+ (is_cached) ? "Destroying" : "Recycling", page_count);
+
+ count = ttm_dma_put_pages_in_pool(pool, pages, page_count, is_cached);
+
+ for (i = 0; i < count; i++)
+ dma_address[i] = 0;
+
+ spin_lock_irqsave(&pool->lock, irq_flags);
+ pool->npages_in_use -= count;
+ if (!is_cached)
+ pool->npages_free += count;
+ spin_unlock_irqrestore(&pool->lock, irq_flags);
+
+ page_count -= count;
+ WARN(page_count != 0,
+ "Only freed %d page(s) in %s. Could not free %d rest!\n",
+ count, pool->name, page_count);
+
+ if (pool->npages_free > _manager->options.max_size) {
+ page_count = pool->npages_free - _manager->options.max_size;
+ if (page_count < NUM_PAGES_TO_ALLOC)
+ page_count = NUM_PAGES_TO_ALLOC;
+ }
+ if (page_count)
+ ttm_dma_page_pool_free(pool, page_count);
+}
+
+/* Get good estimation how many pages are free in pools */
+static int ttm_pool_get_num_unused_pages(void)
+{
+ struct device_pools *p;
+ unsigned total = 0;
+
+ mutex_lock(&_manager->lock);
+ list_for_each_entry(p, &_manager->pools, pools)
+ total += p->pool->npages_free;
+ mutex_unlock(&_manager->lock);
+ return total;
+}
+
+/**
+ * Callback for mm to request pool to reduce number of page held.
+ */
+static int ttm_pool_mm_shrink(struct shrinker *shrink,
+ struct shrink_control *sc)
+{
+ static atomic_t start_pool = ATOMIC_INIT(0);
+ unsigned idx = 0;
+ unsigned pool_offset = atomic_add_return(1, &start_pool);
+ unsigned shrink_pages = sc->nr_to_scan;
+ struct device_pools *p;
+
+ mutex_lock(&_manager->lock);
+ pool_offset = pool_offset % _manager->npools;
+
+ list_for_each_entry(p, &_manager->pools, pools) {
+ unsigned nr_free;
+
+ if (!p->dev)
+ continue;
+ if (shrink_pages == 0)
+ break;
+ /* Do it in round-robin fashion. */
+ if (++idx < pool_offset)
+ continue;
+ nr_free = shrink_pages;
+ shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free);
+ pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
+ p->pool->dev_name, p->pool->name, current->pid, nr_free,
+ shrink_pages);
+ }
+ mutex_unlock(&_manager->lock);
+ /* return estimated number of unused pages in pool */
+ return ttm_pool_get_num_unused_pages();
+}
+
+static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
+{
+ manager->mm_shrink.shrink = &ttm_pool_mm_shrink;
+ manager->mm_shrink.seeks = 1;
+ register_shrinker(&manager->mm_shrink);
+}
+static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
+{
+ unregister_shrinker(&manager->mm_shrink);
+}
+
+static int ttm_dma_page_alloc_init(struct ttm_mem_global *glob,
+ unsigned max_pages)
+{
+ int ret = -ENOMEM;
+
+ WARN_ON(_manager);
+
+ printk(KERN_INFO TTM_PFX "Initializing DMA pool allocator.\n");
+
+ _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
+ if (!_manager)
+ goto err_manager;
+
+ mutex_init(&_manager->lock);
+ INIT_LIST_HEAD(&_manager->pools);
+
+ _manager->options.max_size = max_pages;
+ _manager->options.small = SMALL_ALLOCATION;
+ _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
+
+ /* This takes care of auto-freeing the _manager */
+ ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
+ &glob->kobj, "pool");
+ if (unlikely(ret != 0)) {
+ kobject_put(&_manager->kobj);
+ goto err;
+ }
+ ttm_pool_mm_shrink_init(_manager);
+
+ return 0;
+err_manager:
+ kfree(_manager);
+ _manager = NULL;
+err:
+ return ret;
+}
+static void ttm_dma_page_alloc_fini(void)
+{
+ struct device_pools *p, *t;
+
+ printk(KERN_INFO TTM_PFX "Finalizing DMA pool allocator.\n");
+
+ ttm_pool_mm_shrink_fini(_manager);
+
+ list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
+ dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
+ current->pid);
+ WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
+ ttm_dma_pool_match, p->pool));
+ ttm_dma_free_pool(p->dev, p->pool->type);
+ }
+ kobject_put(&_manager->kobj);
+ _manager = NULL;
+}
+
+static int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
+{
+ struct device_pools *p;
+ struct dma_pool *pool = NULL;
+ char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
+ "name", "virt", "busaddr"};
+
+ if (!_manager) {
+ seq_printf(m, "No pool allocator running.\n");
+ return 0;
+ }
+ seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
+ h[0], h[1], h[2], h[3], h[4], h[5]);
+ mutex_lock(&_manager->lock);
+ list_for_each_entry(p, &_manager->pools, pools) {
+ struct device *dev = p->dev;
+ if (!dev)
+ continue;
+ pool = p->pool;
+ seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
+ pool->name, pool->nrefills,
+ pool->nfrees, pool->npages_in_use,
+ pool->npages_free,
+ pool->dev_name);
+ }
+#ifdef DEBUG
+ seq_printf(m, "%13s %8s %12s %12s %8s\n",
+ h[0], h[3], h[6], h[7], h[5]);
+ list_for_each_entry(p, &_manager->pools, pools) {
+ struct dma_page *d_page;
+ struct device *dev = p->dev;
+ if (!dev)
+ continue;
+ pool = p->pool;
+
+ if ((pool->npages_free + pool->npages_in_use) == 0)
+ continue;
+
+ spin_lock(&pool->lock);
+ list_for_each_entry(d_page, &pool->page_list, page_list) {
+ seq_printf(m,
+ "%13s %8s %12lx %12lx %8s\n",
+ pool->name, d_page->in_use ? "Busy" : "Free",
+ (unsigned long)d_page->vaddr,
+ (unsigned long)d_page->dma,
+ pool->dev_name);
+ }
+ spin_unlock(&pool->lock);
+ }
+#endif
+ mutex_unlock(&_manager->lock);
+ return 0;
+}
+
+struct ttm_page_alloc_func ttm_page_alloc_dma = {
+ .get_pages = ttm_dma_get_pages,
+ .put_pages = ttm_dma_put_pages,
+ .alloc_init = ttm_dma_page_alloc_init,
+ .alloc_fini = ttm_dma_page_alloc_fini,
+ .debugfs = ttm_dma_page_alloc_debugfs,
+};
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h
index 8fc92f2..a58331d 100644
--- a/include/drm/ttm/ttm_page_alloc.h
+++ b/include/drm/ttm/ttm_page_alloc.h
@@ -29,6 +29,11 @@
#include "ttm_bo_driver.h"
#include "ttm_memory.h"
+#ifdef CONFIG_SWIOTLB
+#include <linux/dma-mapping.h>
+#include <linux/swiotlb.h>
+#endif
+
struct ttm_page_alloc_func {
/**
* struct ttm_page_alloc_func member get_pages
@@ -38,7 +43,8 @@ struct ttm_page_alloc_func {
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state for the page.
* @count: number of pages to allocate.
- * @dma_address: The DMA (bus) address of pages (by default zero).
+ * @dma_address: The DMA (bus) address of pages (if
+ * TTM DMA pool is used - otherwise it is zero).
* @dev: The device that needs this.
*/
int (*get_pages) (struct list_head *pages,
@@ -57,7 +63,8 @@ struct ttm_page_alloc_func {
* unknown count.
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state.
- * @dma_address: The DMA (bus) address of pages (by default zero).
+ * @dma_address: The DMA (bus) address of pages (if
+ * TTM DMA pool is used - otherwise it is zero).
* @dev: The device that needs this.
*/
void (*put_pages)(struct list_head *pages,
@@ -93,6 +100,21 @@ extern struct ttm_page_alloc_func *ttm_page_alloc;
/* Defined in ttm_page_alloc.c */
extern struct ttm_page_alloc_func ttm_page_alloc_default;
+#ifdef CONFIG_SWIOTLB
+/* Defined in ttm_page_alloc_dma.c */
+extern struct ttm_page_alloc_func ttm_page_alloc_dma;
+
+static inline bool ttm_page_alloc_need_dma(void)
+{
+ if (swiotlb_nr_tbl()) {
+ ttm_page_alloc = &ttm_page_alloc_dma;
+ return true;
+ }
+ return false;
+}
+#else
+static inline bool ttm_page_alloc_need_dma(void) { return false; }
+#endif
/**
* Get count number of pages from pool to pages list.
*
@@ -100,7 +122,8 @@ extern struct ttm_page_alloc_func ttm_page_alloc_default;
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state for the page.
* @count: number of pages to allocate.
- * @dma_address: The DMA (bus) address of pages - (by default zero).
+ * @dma_address: The DMA (bus) address of pages (if TTM DMA pool is used -
+ * otherwise the value is zero).
* @dev: The device that needs this.
*/
int ttm_get_pages(struct list_head *pages,
@@ -117,7 +140,8 @@ int ttm_get_pages(struct list_head *pages,
* count.
* @flags: ttm flags for page allocation.
* @cstate: ttm caching state.
- * @dma_address: The DMA (bus) address of pages (by default zero).
+ * @dma_address: The DMA (bus) address of pages (if TTM DMA pool is used -
+ * otherwise the value is zero).
* @dev: The device that needs this.
*/
void ttm_put_pages(struct list_head *pages,
--
1.7.4.1
^ permalink raw reply related [flat|nested] 16+ messages in thread* [PATCH 6/9] ttm: Add 'no_dma' parameter to turn the TTM DMA pool off during runtime.
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
` (4 preceding siblings ...)
2011-09-29 20:33 ` [PATCH 5/9] ttm: Provide a DMA aware TTM page pool code Konrad Rzeszutek Wilk
@ 2011-09-29 20:33 ` Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 7/9] nouveau/radeon: Set coherent DMA mask Konrad Rzeszutek Wilk
` (3 subsequent siblings)
9 siblings, 0 replies; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-29 20:33 UTC (permalink / raw)
To: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher
Cc: xen-devel, Konrad Rzeszutek Wilk
The TTM DMA only gets turned on when the SWIOTLB is enabled - but
we might also want to turn it off when SWIOTLB is on to
use the non-DMA TTM pool code.
In the future this parameter can be removed.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 4 ++++
include/drm/ttm/ttm_page_alloc.h | 4 ++--
2 files changed, 6 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
index 5a5739c..5909d28 100644
--- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
@@ -51,6 +51,10 @@
#include <asm/agp.h>
#endif
+int __read_mostly dma_ttm_disable;
+MODULE_PARM_DESC(no_dma, "Disable TTM DMA pool");
+module_param_named(no_dma, dma_ttm_disable, bool, S_IRUGO);
+
#define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
#define SMALL_ALLOCATION 16
#define FREE_ALL_PAGES (~0U)
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h
index a58331d..1dde3bd 100644
--- a/include/drm/ttm/ttm_page_alloc.h
+++ b/include/drm/ttm/ttm_page_alloc.h
@@ -103,10 +103,10 @@ extern struct ttm_page_alloc_func ttm_page_alloc_default;
#ifdef CONFIG_SWIOTLB
/* Defined in ttm_page_alloc_dma.c */
extern struct ttm_page_alloc_func ttm_page_alloc_dma;
-
+extern int dma_ttm_disable;
static inline bool ttm_page_alloc_need_dma(void)
{
- if (swiotlb_nr_tbl()) {
+ if (!dma_ttm_disable && swiotlb_nr_tbl()) {
ttm_page_alloc = &ttm_page_alloc_dma;
return true;
}
--
1.7.4.1
^ permalink raw reply related [flat|nested] 16+ messages in thread* [PATCH 7/9] nouveau/radeon: Set coherent DMA mask
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
` (5 preceding siblings ...)
2011-09-29 20:33 ` [PATCH 6/9] ttm: Add 'no_dma' parameter to turn the TTM DMA pool off during runtime Konrad Rzeszutek Wilk
@ 2011-09-29 20:33 ` Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 8/9] ttm/tt: Move ttm_tt_set_page_caching implementation in TTM page pool code Konrad Rzeszutek Wilk
` (2 subsequent siblings)
9 siblings, 0 replies; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-29 20:33 UTC (permalink / raw)
To: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher
Cc: xen-devel, Konrad Rzeszutek Wilk
All the storage devices that use the dmapool set the coherent DMA
mask so they can properly use the dmapool. Since the TTM DMA pool
code is based on that and dma_alloc_coherent checks the
'coherent_dma_mask' and not 'dma_mask' we want to set it.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/nouveau/nouveau_mem.c | 5 +++++
drivers/gpu/drm/radeon/radeon_device.c | 6 ++++++
2 files changed, 11 insertions(+), 0 deletions(-)
diff --git a/drivers/gpu/drm/nouveau/nouveau_mem.c b/drivers/gpu/drm/nouveau/nouveau_mem.c
index a2d7e35..bb6ccbd 100644
--- a/drivers/gpu/drm/nouveau/nouveau_mem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_mem.c
@@ -408,6 +408,11 @@ nouveau_mem_vram_init(struct drm_device *dev)
if (ret)
return ret;
+ ret = pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(dma_bits));
+ if (ret) {
+ /* Reset to default value. */
+ pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(32));
+ }
dev_priv->fb_phys = pci_resource_start(dev->pdev, 1);
ret = nouveau_ttm_global_init(dev_priv);
diff --git a/drivers/gpu/drm/radeon/radeon_device.c b/drivers/gpu/drm/radeon/radeon_device.c
index 7cfaa7e..0c0a970 100644
--- a/drivers/gpu/drm/radeon/radeon_device.c
+++ b/drivers/gpu/drm/radeon/radeon_device.c
@@ -757,8 +757,14 @@ int radeon_device_init(struct radeon_device *rdev,
r = pci_set_dma_mask(rdev->pdev, DMA_BIT_MASK(dma_bits));
if (r) {
rdev->need_dma32 = true;
+ dma_bits = 32;
printk(KERN_WARNING "radeon: No suitable DMA available.\n");
}
+ r = pci_set_consistent_dma_mask(rdev->pdev, DMA_BIT_MASK(dma_bits));
+ if (r) {
+ pci_set_consistent_dma_mask(rdev->pdev, DMA_BIT_MASK(32));
+ printk(KERN_WARNING "radeon: No coherent DMA available.\n");
+ }
/* Registers mapping */
/* TODO: block userspace mapping of io register */
--
1.7.4.1
^ permalink raw reply related [flat|nested] 16+ messages in thread* [PATCH 8/9] ttm/tt: Move ttm_tt_set_page_caching implementation in TTM page pool code.
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
` (6 preceding siblings ...)
2011-09-29 20:33 ` [PATCH 7/9] nouveau/radeon: Set coherent DMA mask Konrad Rzeszutek Wilk
@ 2011-09-29 20:33 ` Konrad Rzeszutek Wilk
2011-09-29 20:33 ` [PATCH 9/9] ttm/dma: Implement set_page_caching implementation in the TTM DMA " Konrad Rzeszutek Wilk
2011-09-30 6:59 ` [PATCH] TTM DMA pool v1.8 Thomas Hellstrom
9 siblings, 0 replies; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-29 20:33 UTC (permalink / raw)
To: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher
Cc: xen-devel, Konrad Rzeszutek Wilk
. and provide in the function callback a method to register
with the backend a new (*set_caching).
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/ttm/ttm_page_alloc.c | 46 ++++++++++++++++++++++++++++++++++
drivers/gpu/drm/ttm/ttm_tt.c | 46 +++++----------------------------
include/drm/ttm/ttm_page_alloc.h | 29 +++++++++++++++++++++
3 files changed, 82 insertions(+), 39 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c
index f9a4d83..002f64f 100644
--- a/drivers/gpu/drm/ttm/ttm_page_alloc.c
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c
@@ -455,6 +455,36 @@ static int ttm_set_pages_caching(struct page **pages,
return r;
}
+#ifdef CONFIG_X86
+static int __ttm_page_set_page_caching(struct page *p,
+ int flags,
+ enum ttm_caching_state c_old,
+ enum ttm_caching_state c_new,
+ struct device *dev)
+{
+ int ret = 0;
+
+ if (PageHighMem(p))
+ return 0;
+
+ if (c_old != tt_cached) {
+ /* p isn't in the default caching state, set it to
+ * writeback first to free its current memtype. */
+
+ ret = set_pages_wb(p, 1);
+ if (ret)
+ return ret;
+ }
+
+ if (c_new == tt_wc)
+ ret = set_memory_wc((unsigned long) page_address(p), 1);
+ else if (c_new == tt_uncached)
+ ret = set_pages_uc(p, 1);
+
+ return ret;
+}
+#endif
+
/**
* Free pages the pages that failed to change the caching state. If there is
* any pages that have changed their caching state already put them to the
@@ -865,6 +895,9 @@ struct ttm_page_alloc_func ttm_page_alloc_default = {
.alloc_init = __ttm_page_alloc_init,
.alloc_fini = __ttm_page_alloc_fini,
.debugfs = __ttm_page_alloc_debugfs,
+#ifdef CONFIG_X86
+ .set_caching = __ttm_page_set_page_caching,
+#endif
};
int ttm_get_pages(struct list_head *pages, int flags,
@@ -902,3 +935,16 @@ int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
return -1;
}
EXPORT_SYMBOL(ttm_page_alloc_debugfs);
+
+#ifdef CONFIG_X86
+int ttm_tt_set_page_caching(struct page *p,
+ int flags,
+ enum ttm_caching_state c_old,
+ enum ttm_caching_state c_new,
+ struct device *dev)
+{
+ if (ttm_page_alloc && ttm_page_alloc->set_caching)
+ return ttm_page_alloc->set_caching(p, flags, c_old, c_new, dev);
+ return -1;
+}
+#endif
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 1f11a33..0f5ce97 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -176,41 +176,6 @@ int ttm_tt_populate(struct ttm_tt *ttm)
}
EXPORT_SYMBOL(ttm_tt_populate);
-#ifdef CONFIG_X86
-static inline int ttm_tt_set_page_caching(struct page *p,
- enum ttm_caching_state c_old,
- enum ttm_caching_state c_new)
-{
- int ret = 0;
-
- if (PageHighMem(p))
- return 0;
-
- if (c_old != tt_cached) {
- /* p isn't in the default caching state, set it to
- * writeback first to free its current memtype. */
-
- ret = set_pages_wb(p, 1);
- if (ret)
- return ret;
- }
-
- if (c_new == tt_wc)
- ret = set_memory_wc((unsigned long) page_address(p), 1);
- else if (c_new == tt_uncached)
- ret = set_pages_uc(p, 1);
-
- return ret;
-}
-#else /* CONFIG_X86 */
-static inline int ttm_tt_set_page_caching(struct page *p,
- enum ttm_caching_state c_old,
- enum ttm_caching_state c_new)
-{
- return 0;
-}
-#endif /* CONFIG_X86 */
-
/*
* Change caching policy for the linear kernel map
* for range of pages in a ttm.
@@ -238,9 +203,9 @@ static int ttm_tt_set_caching(struct ttm_tt *ttm,
for (i = 0; i < ttm->num_pages; ++i) {
cur_page = ttm->pages[i];
if (likely(cur_page != NULL)) {
- ret = ttm_tt_set_page_caching(cur_page,
+ ret = ttm_tt_set_page_caching(cur_page, ttm->page_flags,
ttm->caching_state,
- c_state);
+ c_state, ttm->dev);
if (unlikely(ret != 0))
goto out_err;
}
@@ -254,8 +219,11 @@ out_err:
for (j = 0; j < i; ++j) {
cur_page = ttm->pages[j];
if (likely(cur_page != NULL)) {
- (void)ttm_tt_set_page_caching(cur_page, c_state,
- ttm->caching_state);
+ (void)ttm_tt_set_page_caching(cur_page,
+ ttm->page_flags,
+ c_state,
+ ttm->caching_state,
+ ttm->dev);
}
}
diff --git a/include/drm/ttm/ttm_page_alloc.h b/include/drm/ttm/ttm_page_alloc.h
index 1dde3bd..73ad03c 100644
--- a/include/drm/ttm/ttm_page_alloc.h
+++ b/include/drm/ttm/ttm_page_alloc.h
@@ -93,6 +93,18 @@ struct ttm_page_alloc_func {
* Output the state of pools to debugfs file
*/
int (*debugfs)(struct seq_file *m, void *data);
+
+ /**
+ *
+ * struct ttm_page_alloc_func member set_caching
+ *
+ * Set the caching on the page retrieved from get_pages.
+ */
+ int (*set_caching)(struct page *p,
+ int flags,
+ enum ttm_caching_state c_old,
+ enum ttm_caching_state c_new,
+ struct device *dev);
};
extern struct ttm_page_alloc_func *ttm_page_alloc;
@@ -163,4 +175,21 @@ void ttm_page_alloc_fini(void);
* Output the state of pools to debugfs file
*/
extern int ttm_page_alloc_debugfs(struct seq_file *m, void *data);
+
+#ifdef CONFIG_X86
+int ttm_tt_set_page_caching(struct page *p,
+ int flags,
+ enum ttm_caching_state c_old,
+ enum ttm_caching_state c_new,
+ struct device *dev);
+#else
+static inline int ttm_tt_set_page_caching(struct page *p,
+ int flags,
+ enum ttm_caching_state c_old,
+ enum ttm_caching_state c_new,
+ struct device *dev)
+{
+ return 0;
+}
+#endif
#endif
--
1.7.4.1
^ permalink raw reply related [flat|nested] 16+ messages in thread* [PATCH 9/9] ttm/dma: Implement set_page_caching implementation in the TTM DMA pool code.
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
` (7 preceding siblings ...)
2011-09-29 20:33 ` [PATCH 8/9] ttm/tt: Move ttm_tt_set_page_caching implementation in TTM page pool code Konrad Rzeszutek Wilk
@ 2011-09-29 20:33 ` Konrad Rzeszutek Wilk
2011-09-30 6:59 ` [PATCH] TTM DMA pool v1.8 Thomas Hellstrom
9 siblings, 0 replies; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-29 20:33 UTC (permalink / raw)
To: linux-kernel, thellstrom, dri-devel, bskeggs, j.glisse, thomas,
airlied, airlied, alexdeucher
Cc: xen-devel, Konrad Rzeszutek Wilk
. which is pretty much like the other TTM pool except it
also handles moving the page to another pool list.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 96 ++++++++++++++++++++++++++++++
1 files changed, 96 insertions(+), 0 deletions(-)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
index 5909d28..cea031e 100644
--- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
@@ -1307,11 +1307,107 @@ static int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
mutex_unlock(&_manager->lock);
return 0;
}
+#ifdef CONFIG_X86
+static int ttm_dma_page_set_page_caching(struct page *p,
+ int flags,
+ enum ttm_caching_state c_old,
+ enum ttm_caching_state c_new,
+ struct device *dev)
+{
+ struct dma_pool *src, *dst;
+ enum pool_type type;
+ struct dma_page *dma_p;
+ bool found = false;
+ unsigned long irq_flags;
+ int ret = 0;
+
+ if (!p)
+ return 0;
+
+ if (PageHighMem(p))
+ return 0;
+
+ type = ttm_to_type(flags, c_old);
+ src = ttm_dma_find_pool(dev, type);
+ if (!src) {
+ WARN_ON(!src);
+ return -ENOMEM;
+ }
+ type = ttm_to_type(flags, c_new);
+ dst = ttm_dma_find_pool(dev, type);
+ if (!dst) {
+ gfp_t gfp_flags;
+ if (flags & TTM_PAGE_FLAG_DMA32)
+ gfp_flags = GFP_USER | GFP_DMA32;
+ else
+ gfp_flags = GFP_HIGHUSER;
+
+ if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
+ gfp_flags |= __GFP_ZERO;
+
+ dst = ttm_dma_pool_init(dev, gfp_flags, type);
+ if (IS_ERR_OR_NULL(dst))
+ return -ENOMEM;
+ }
+
+ dev_dbg(dev, "(%d) Caching %p (%p) from %x to %x.\n", current->pid,
+ p, page_address(p), c_old, c_new);
+
+ if (c_old != tt_cached) {
+ /* p isn't in the default caching state, set it to
+ * writeback first to free its current memtype. */
+
+ ret = set_pages_wb(p, 1);
+ if (ret)
+ return ret;
+ }
+ if (c_new == tt_wc)
+ ret = set_memory_wc((unsigned long) page_address(p), 1);
+ else if (c_new == tt_uncached)
+ ret = set_pages_uc(p, 1);
+
+ if (ret)
+ return ret;
+
+ dev_dbg(src->dev, "(%s:%d) Moving %p (%p) to %s.\n", src->name,
+ current->pid, p, page_address(p), dst->name);
+
+ /* To make it faster we only take the spinlock on list
+ * removal, and later on adding the page to the destination pool. */
+ spin_lock_irqsave(&src->lock, irq_flags);
+ list_for_each_entry(dma_p, &src->page_list, page_list) {
+ if (virt_to_page(dma_p->vaddr) != p) {
+ pr_debug("%s: (%s:%d) Skipping %p (%p) (DMA:0x%lx)\n",
+ src->dev_name, src->name, current->pid,
+ dma_p->vaddr,
+ virt_to_page(dma_p->vaddr),
+ (unsigned long)dma_p->dma);
+ continue;
+ }
+ list_del(&dma_p->page_list);
+ src->npages_in_use -= 1;
+ found = true;
+ break;
+ }
+ spin_lock_irqsave(&src->lock, irq_flags);
+ if (!found)
+ return -ENODEV;
+
+ spin_lock_irqsave(&dst->lock, irq_flags);
+ list_add(&dma_p->page_list, &dst->page_list);
+ dst->npages_in_use++;
+ spin_unlock_irqrestore(&dst->lock, irq_flags);
+ return 0;
+};
+#endif
struct ttm_page_alloc_func ttm_page_alloc_dma = {
.get_pages = ttm_dma_get_pages,
.put_pages = ttm_dma_put_pages,
.alloc_init = ttm_dma_page_alloc_init,
.alloc_fini = ttm_dma_page_alloc_fini,
.debugfs = ttm_dma_page_alloc_debugfs,
+#ifdef CONFIG_X86
+ .set_caching = ttm_dma_page_set_page_caching,
+#endif
};
--
1.7.4.1
^ permalink raw reply related [flat|nested] 16+ messages in thread* Re: [PATCH] TTM DMA pool v1.8
2011-09-29 20:33 [PATCH] TTM DMA pool v1.8 Konrad Rzeszutek Wilk
` (8 preceding siblings ...)
2011-09-29 20:33 ` [PATCH 9/9] ttm/dma: Implement set_page_caching implementation in the TTM DMA " Konrad Rzeszutek Wilk
@ 2011-09-30 6:59 ` Thomas Hellstrom
2011-09-30 14:09 ` Konrad Rzeszutek Wilk
9 siblings, 1 reply; 16+ messages in thread
From: Thomas Hellstrom @ 2011-09-30 6:59 UTC (permalink / raw)
To: Konrad Rzeszutek Wilk
Cc: linux-kernel, dri-devel, bskeggs, j.glisse, thomas, airlied,
airlied, alexdeucher, xen-devel
Konrad,
I'm really sorry for taking so long to review this.
I'd like to go through a couple of high-level things first before
reviewing the coding itself.
The page_alloc_func structure looks nice, but I'd like to have it per
ttm backend,
we would just need to make sure that the backend is alive when we alloc
/ free pages.
The reason for this is that there may be backends that want to allocate
dma memory running simultaneously with those who don't. When the backend
fires up, it can determine whether to use DMA memory or not.
This also eliminates the need for patch 3/9. and is in line with patch 8/9.
2) Memory accounting: If the number DMA pages are limited in a way that
the ttm memory global routines are not aware of. How do we handle memory
accounting? (How do we avoid exhausting IOMMU space)?
3) Page swapping. Currently we just copy pages to shmem pages and then
free device pages. In the future we'd probably like to insert non-dma
pages directly into the swap cache. Is it possible to differentiate dma
pages from pages that are directly insertable?
Thanks
Thomas
On 09/29/2011 10:33 PM, Konrad Rzeszutek Wilk wrote:
> [.. and this is what I said in v1 post]:
>
> Way back in January this patchset:
> http://lists.freedesktop.org/archives/dri-devel/2011-January/006905.html
> was merged in, but pieces of it had to be reverted b/c they did not
> work properly under PowerPC, ARM, and when swapping out pages to disk.
>
> After a bit of discussion on the mailing list
> http://marc.info/?i=4D769726.2030307@shipmail.org I started working on it, but
> got waylaid by other things .. and finally I am able to post the RFC patches.
>
> There was a lot of discussion about it and I am not sure if I captured
> everybody's thoughts - if I did not - that is _not_ intentional - it has just
> been quite some time..
>
> Anyhow .. the patches explore what the "lib/dmapool.c" does - which is to have a
> DMA pool that the device has associated with. I kind of married that code
> along with drivers/gpu/drm/ttm/ttm_page_alloc.c to create a TTM DMA pool code.
> The end result is DMA pool with extra features: can do write-combine, uncached,
> writeback (and tracks them and sets back to WB when freed); tracks "cached"
> pages that don't really need to be returned to a pool; and hooks up to
> the shrinker code so that the pools can be shrunk.
>
> If you guys think this set of patches make sense - my future plans were
> 1) Get this in large crowd of testing .. and if it works for a kernel release
> 2) to move a bulk of this in the lib/dmapool.c (I spoke with Matthew Wilcox
> about it and he is OK as long as I don't introduce performance regressions).
>
> But before I do any of that a second set of eyes taking a look at these
> patches would be most welcome.
>
> In regards to testing, I've been running them non-stop for the last month
> (and found some issues which I've fixed up) - and been quite happy with how
> they work.
>
> Michel (thanks!) took a spin of the patches on his PowerPC and they did not
> cause any regressions (wheew).
>
> The patches are also located in a git tree:
>
> git://oss.oracle.com/git/kwilk/xen.git devel/ttm.dma_pool.v1.8
>
> drivers/gpu/drm/nouveau/nouveau_mem.c | 8 +-
> drivers/gpu/drm/nouveau/nouveau_sgdma.c | 3 +-
> drivers/gpu/drm/radeon/radeon_device.c | 6 +
> drivers/gpu/drm/radeon/radeon_gart.c | 4 +-
> drivers/gpu/drm/radeon/radeon_ttm.c | 3 +-
> drivers/gpu/drm/ttm/Makefile | 3 +
> drivers/gpu/drm/ttm/ttm_bo.c | 4 +-
> drivers/gpu/drm/ttm/ttm_memory.c | 5 +
> drivers/gpu/drm/ttm/ttm_page_alloc.c | 63 ++-
> drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 1317 ++++++++++++++++++++++++++++++
> drivers/gpu/drm/ttm/ttm_tt.c | 5 +-
> drivers/gpu/drm/vmwgfx/vmwgfx_drv.c | 4 +-
> drivers/xen/swiotlb-xen.c | 2 +-
> include/drm/ttm/ttm_bo_driver.h | 7 +-
> include/drm/ttm/ttm_page_alloc.h | 100 +++-
> include/linux/swiotlb.h | 7 +-
> lib/swiotlb.c | 5 +-
> 17 files changed, 1516 insertions(+), 30 deletions(-)
>
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH] TTM DMA pool v1.8
2011-09-30 6:59 ` [PATCH] TTM DMA pool v1.8 Thomas Hellstrom
@ 2011-09-30 14:09 ` Konrad Rzeszutek Wilk
2011-10-03 16:35 ` Thomas Hellstrom
0 siblings, 1 reply; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-09-30 14:09 UTC (permalink / raw)
To: Thomas Hellstrom
Cc: linux-kernel, dri-devel, bskeggs, j.glisse, thomas, airlied,
airlied, alexdeucher, xen-devel
On Fri, Sep 30, 2011 at 08:59:52AM +0200, Thomas Hellstrom wrote:
> Konrad,
>
> I'm really sorry for taking so long to review this.
That is OK. We all are busy - and it gave me some time to pretty
up the code even more.
>
> I'd like to go through a couple of high-level things first before
> reviewing the coding itself.
>
> The page_alloc_func structure looks nice, but I'd like to have it
> per ttm backend,
Meaning the 'struct ttm_backend_func' right?
> we would just need to make sure that the backend is alive when we
> alloc / free pages.
> The reason for this is that there may be backends that want to
> allocate dma memory running simultaneously with those who don't.
As in for some TTM manager use the non-DMA, and for other DMA
(is_dma32 is set?) Or say two cards - one that has the concept
of MMU and one that does not and both of them are readeon?
> When the backend fires up, it can determine whether to use DMA
> memory or not.
Or more of backends (and drivers) that do not have any concept
of MMU and just use framebuffers and such?
I think we would just have to stick in a pointer to the
appropiate 'struct ttm_page_alloc_func' (and the 'struct device')
in the 'struct ttm_backend_func'. And this would be done by
backend that would decided which one to use.
And the TTM could find out which page_alloc_func to use straight
from the ttm_backend_func and call that (along with the 'struct device'
also gathered from that structure). Rough idea (on top of the patches):
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c
index dd05db3..e7a0c3c 100644
--- a/drivers/gpu/drm/ttm/ttm_page_alloc.c
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c
@@ -902,12 +902,12 @@ struct ttm_page_alloc_func ttm_page_alloc_default = {
int ttm_get_pages(struct list_head *pages, int flags,
enum ttm_caching_state cstate, unsigned count,
- dma_addr_t *dma_address, struct device *dev)
+ dma_addr_t *dma_address, struct ttm_backend *be)
{
- if (ttm_page_alloc && ttm_page_alloc->get_pages)
- return ttm_page_alloc->get_pages(pages, flags, cstate, count,
- dma_address, dev);
- return -1;
+ if (be->page_backend && be->page_backend->get_pages)
+ return be->page_backend->get_pages(pages, flags, cstate, count,
+ dma_address, be->dev);
+ return __ttm_get_pages(pages, flags, cstate, count, dma_address, NULL);
}
void ttm_put_pages(struct list_head *pages, unsigned page_count, int flags,
enum ttm_caching_state cstate, dma_addr_t *dma_address,
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 0f5ce97..0d75213 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -111,7 +111,7 @@ static struct page *__ttm_tt_get_page(struct ttm_tt *ttm, int index)
INIT_LIST_HEAD(&h);
ret = ttm_get_pages(&h, ttm->page_flags, ttm->caching_state, 1,
- &ttm->dma_address[index], ttm->dev);
+ &ttm->dma_address[index], ttm->be);
if (ret != 0)
return NULL;
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 6509544..9729753 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -100,6 +100,18 @@ struct ttm_backend_func {
* Destroy the backend.
*/
void (*destroy) (struct ttm_backend *backend);
+
+ /*
+ * Pointer to the struct ttm_page_alloc_func this backend wants
+ * to use.
+ */
+ struct ttm_page_alloc_func *page_backend;
+
+ /*
+ * The device that the ttm_page_alloc_func should use for this
+ * backend.
+ */
+ struct device *dev;
};
/**
>
> This also eliminates the need for patch 3/9. and is in line with patch 8/9.
<nods> That would be " ttm: Pass in 'struct device' to TTM so it can do
DMA API on behalf of device." as now the driver would be responsible for
saving that.
And "ttm/tt: Move ttm_tt_set_page_caching implementation in TTM page pool code."
would still be there, except it would be done per ttm-backend. Well
by choosing which TTM page pool the TTM backend would use.
>
> 2) Memory accounting: If the number DMA pages are limited in a way
> that the ttm memory global routines are not aware of. How do we
> handle memory accounting? (How do we avoid exhausting IOMMU space)?
Good question. Not entirely sure about that. I know that on
SWIOTLB there is no limit (as you do not use the bounce buffer)
but not sure about VT-D, AMD-VI, GART, or when there is no IOMMU.
Let me get back to you on that.
Granted, the code _only_ gets activated when we use SWIOTLB right
now so the answer is "no exhausting" currently. Either way let me
dig a bit deeper.
>
> 3) Page swapping. Currently we just copy pages to shmem pages and
> then free device pages. In the future we'd probably like to insert
> non-dma pages directly into the swap cache. Is it possible to
> differentiate dma pages from pages that are directly insertable?
Yes. The TTM DMA pool keeps track of which pages belong to which
pool and will reject non-dma pages (or pages which belong to
a different pool). It is fairly easy to expose that check
so that the generic TTM code can make the proper distinction.
But also - once you free a device page ('cached') it gets freed.
The other pages (writecombine, writeback, uncached) end up sitting
in a recycle pool to be used. This is believe how the current
TTM page code works right now (and this TTM DMA pool follows).
The swapping code back (so from swap to pool) does not seem to
distinguish it that much - it just allocates a new page - and
then copies from whatever was in the swap cache?
This is something you were thinking to do in the future I presume?
^ permalink raw reply related [flat|nested] 16+ messages in thread* Re: [PATCH] TTM DMA pool v1.8
2011-09-30 14:09 ` Konrad Rzeszutek Wilk
@ 2011-10-03 16:35 ` Thomas Hellstrom
2011-10-03 16:46 ` Konrad Rzeszutek Wilk
0 siblings, 1 reply; 16+ messages in thread
From: Thomas Hellstrom @ 2011-10-03 16:35 UTC (permalink / raw)
To: Konrad Rzeszutek Wilk
Cc: linux-kernel, dri-devel, bskeggs, j.glisse, thomas, airlied,
airlied, alexdeucher, xen-devel
On 09/30/2011 04:09 PM, Konrad Rzeszutek Wilk wrote:
> On Fri, Sep 30, 2011 at 08:59:52AM +0200, Thomas Hellstrom wrote:
>
>> Konrad,
>>
>> I'm really sorry for taking so long to review this.
>>
> That is OK. We all are busy - and it gave me some time to pretty
> up the code even more.
>
>
>> I'd like to go through a couple of high-level things first before
>> reviewing the coding itself.
>>
>> The page_alloc_func structure looks nice, but I'd like to have it
>> per ttm backend,
>>
> Meaning the 'struct ttm_backend_func' right?
>
>
Yes, that's right.
>> we would just need to make sure that the backend is alive when we
>> alloc / free pages.
>> The reason for this is that there may be backends that want to
>> allocate dma memory running simultaneously with those who don't.
>>
> As in for some TTM manager use the non-DMA, and for other DMA
> (is_dma32 is set?) Or say two cards - one that has the concept
> of MMU and one that does not and both of them are readeon?
>
For example, or let's say you have a low-end system that in the future
needs to
allocate DMA memory, and want to plug in a high-end graphics card, like
Radeon.
>
>> When the backend fires up, it can determine whether to use DMA
>> memory or not.
>>
> Or more of backends (and drivers) that do not have any concept
> of MMU and just use framebuffers and such?
>
> I think we would just have to stick in a pointer to the
> appropiate 'struct ttm_page_alloc_func' (and the 'struct device')
> in the 'struct ttm_backend_func'. And this would be done by
> backend that would decided which one to use.
>
Yes, either that, or merge the two structs.
> And the TTM could find out which page_alloc_func to use straight
> from the ttm_backend_func and call that (along with the 'struct device'
> also gathered from that structure). Rough idea (on top of the patches):
>
> diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c
> index dd05db3..e7a0c3c 100644
> --- a/drivers/gpu/drm/ttm/ttm_page_alloc.c
> +++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c
> @@ -902,12 +902,12 @@ struct ttm_page_alloc_func ttm_page_alloc_default = {
>
> int ttm_get_pages(struct list_head *pages, int flags,
> enum ttm_caching_state cstate, unsigned count,
> - dma_addr_t *dma_address, struct device *dev)
> + dma_addr_t *dma_address, struct ttm_backend *be)
>
I'd like to see it even more simple. If the ttm_backend_func is
responsible for allocating pages,
ttm_get_pages would be called by the backend code, and the dma_addr_t
pointer can be kept
in the backend object. No need to expose neither device nor dma address
to core ttm that
really doesn't want to care. The dma_address is then available in the
backend only
for binding / unbinding, and ttm_get_pages will be called exclusively by
the backend, and its
interface can pass struct device.
> And "ttm/tt: Move ttm_tt_set_page_caching implementation in TTM page pool code."
> would still be there, except it would be done per ttm-backend. Well
> by choosing which TTM page pool the TTM backend would use.
>
>
Yes.
>> 2) Memory accounting: If the number DMA pages are limited in a way
>> that the ttm memory global routines are not aware of. How do we
>> handle memory accounting? (How do we avoid exhausting IOMMU space)?
>>
> Good question. Not entirely sure about that. I know that on
> SWIOTLB there is no limit (as you do not use the bounce buffer)
> but not sure about VT-D, AMD-VI, GART, or when there is no IOMMU.
>
> Let me get back to you on that.
>
> Granted, the code _only_ gets activated when we use SWIOTLB right
> now so the answer is "no exhausting" currently. Either way let me
> dig a bit deeper.
>
I'm fine with it working OK with SWIOTLB now.
When we need to handle other situations, let's find out how to do it then.
>> 3) Page swapping. Currently we just copy pages to shmem pages and
>> then free device pages. In the future we'd probably like to insert
>> non-dma pages directly into the swap cache. Is it possible to
>> differentiate dma pages from pages that are directly insertable?
>>
> Yes. The TTM DMA pool keeps track of which pages belong to which
> pool and will reject non-dma pages (or pages which belong to
> a different pool). It is fairly easy to expose that check
> so that the generic TTM code can make the proper distinction.
>
> But also - once you free a device page ('cached') it gets freed.
> The other pages (writecombine, writeback, uncached) end up sitting
> in a recycle pool to be used. This is believe how the current
> TTM page code works right now (and this TTM DMA pool follows).
>
Yes, that's OK, as long as the system shrinker can free those pages.
> The swapping code back (so from swap to pool) does not seem to
> distinguish it that much - it just allocates a new page - and
> then copies from whatever was in the swap cache?
>
> This is something you were thinking to do in the future I presume?
>
Yes. If / when I do that, I might be adding a new backend function to
put a ttm in an
"anonymous state", that is using only pages that can be inserted in the
swap cache or passed
around to other devices, and to put a ttm in a "device" state, that
copies it to device mappable pages.
Thanks,
/Thomas
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH] TTM DMA pool v1.8
2011-10-03 16:35 ` Thomas Hellstrom
@ 2011-10-03 16:46 ` Konrad Rzeszutek Wilk
2011-10-03 16:56 ` Thomas Hellstrom
0 siblings, 1 reply; 16+ messages in thread
From: Konrad Rzeszutek Wilk @ 2011-10-03 16:46 UTC (permalink / raw)
To: Thomas Hellstrom
Cc: linux-kernel, dri-devel, bskeggs, j.glisse, thomas, airlied,
airlied, alexdeucher, xen-devel
On Mon, Oct 03, 2011 at 06:35:42PM +0200, Thomas Hellstrom wrote:
> On 09/30/2011 04:09 PM, Konrad Rzeszutek Wilk wrote:
> >On Fri, Sep 30, 2011 at 08:59:52AM +0200, Thomas Hellstrom wrote:
> >>Konrad,
> >>
> >>I'm really sorry for taking so long to review this.
> >That is OK. We all are busy - and it gave me some time to pretty
> >up the code even more.
> >
> >>I'd like to go through a couple of high-level things first before
> >>reviewing the coding itself.
> >>
> >>The page_alloc_func structure looks nice, but I'd like to have it
> >>per ttm backend,
> >Meaning the 'struct ttm_backend_func' right?
> >
>
> Yes, that's right.
>
> >>we would just need to make sure that the backend is alive when we
> >>alloc / free pages.
> >>The reason for this is that there may be backends that want to
> >>allocate dma memory running simultaneously with those who don't.
> >As in for some TTM manager use the non-DMA, and for other DMA
> >(is_dma32 is set?) Or say two cards - one that has the concept
> >of MMU and one that does not and both of them are readeon?
>
> For example, or let's say you have a low-end system that in the
> future needs to
> allocate DMA memory, and want to plug in a high-end graphics card,
> like Radeon.
>
>
>
> >>When the backend fires up, it can determine whether to use DMA
> >>memory or not.
> >Or more of backends (and drivers) that do not have any concept
> >of MMU and just use framebuffers and such?
> >
> >I think we would just have to stick in a pointer to the
> >appropiate 'struct ttm_page_alloc_func' (and the 'struct device')
> >in the 'struct ttm_backend_func'. And this would be done by
> >backend that would decided which one to use.
>
> Yes, either that, or merge the two structs.
>
> >And the TTM could find out which page_alloc_func to use straight
> >from the ttm_backend_func and call that (along with the 'struct device'
> >also gathered from that structure). Rough idea (on top of the patches):
> >
> >diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc.c b/drivers/gpu/drm/ttm/ttm_page_alloc.c
> >index dd05db3..e7a0c3c 100644
> >--- a/drivers/gpu/drm/ttm/ttm_page_alloc.c
> >+++ b/drivers/gpu/drm/ttm/ttm_page_alloc.c
> >@@ -902,12 +902,12 @@ struct ttm_page_alloc_func ttm_page_alloc_default = {
> >
> > int ttm_get_pages(struct list_head *pages, int flags,
> > enum ttm_caching_state cstate, unsigned count,
> >- dma_addr_t *dma_address, struct device *dev)
> >+ dma_addr_t *dma_address, struct ttm_backend *be)
>
> I'd like to see it even more simple. If the ttm_backend_func is
> responsible for allocating pages,
> ttm_get_pages would be called by the backend code, and the
> dma_addr_t pointer can be kept
> in the backend object. No need to expose neither device nor dma
> address to core ttm that
> really doesn't want to care. The dma_address is then available in
> the backend only
> for binding / unbinding, and ttm_get_pages will be called
> exclusively by the backend, and its
> interface can pass struct device.
<nods> OK, let me start cracking on that.
>
> >And "ttm/tt: Move ttm_tt_set_page_caching implementation in TTM page pool code."
> >would still be there, except it would be done per ttm-backend. Well
> >by choosing which TTM page pool the TTM backend would use.
> >
>
> Yes.
>
> >>2) Memory accounting: If the number DMA pages are limited in a way
> >>that the ttm memory global routines are not aware of. How do we
> >>handle memory accounting? (How do we avoid exhausting IOMMU space)?
> >Good question. Not entirely sure about that. I know that on
> >SWIOTLB there is no limit (as you do not use the bounce buffer)
> >but not sure about VT-D, AMD-VI, GART, or when there is no IOMMU.
> >
> >Let me get back to you on that.
> >
> >Granted, the code _only_ gets activated when we use SWIOTLB right
> >now so the answer is "no exhausting" currently. Either way let me
> >dig a bit deeper.
>
> I'm fine with it working OK with SWIOTLB now.
> When we need to handle other situations, let's find out how to do it then.
<nods>
>
> >>3) Page swapping. Currently we just copy pages to shmem pages and
> >>then free device pages. In the future we'd probably like to insert
> >>non-dma pages directly into the swap cache. Is it possible to
> >>differentiate dma pages from pages that are directly insertable?
> >Yes. The TTM DMA pool keeps track of which pages belong to which
> >pool and will reject non-dma pages (or pages which belong to
> >a different pool). It is fairly easy to expose that check
> >so that the generic TTM code can make the proper distinction.
> >
> >But also - once you free a device page ('cached') it gets freed.
> >The other pages (writecombine, writeback, uncached) end up sitting
> >in a recycle pool to be used. This is believe how the current
> >TTM page code works right now (and this TTM DMA pool follows).
>
> Yes, that's OK, as long as the system shrinker can free those pages.
It does now (I had a spinlock mishap).. which reminds me - how
do I test these patches with your vmwgfx driver? I've an old version
of VMWare Workstation 8, would that do?
>
> >The swapping code back (so from swap to pool) does not seem to
> >distinguish it that much - it just allocates a new page - and
> >then copies from whatever was in the swap cache?
> >
> >This is something you were thinking to do in the future I presume?
>
> Yes. If / when I do that, I might be adding a new backend function
> to put a ttm in an
> "anonymous state", that is using only pages that can be inserted in
> the swap cache or passed
> around to other devices, and to put a ttm in a "device" state, that
> copies it to device mappable pages.
OK, that should be no trouble - we would need to expose a function
call to "detach" the page from the TTM pool (which could mean
actually allocating a new page for the "other" device, and copying
it from the "source" to "other" and then freeing the "source).
I am thinking ... you hotplug an high-end radeon while the machine has
an ATI ES1000 in it, and want to move those pages to the new
card. The ATI ES1000 can only do up to 4GB, while the new fancy card
has no such limits (and perhaps does not want to use the TTM DMA
pool).
Is this what you had in mind?
>
> Thanks,
> /Thomas
^ permalink raw reply [flat|nested] 16+ messages in thread* Re: [PATCH] TTM DMA pool v1.8
2011-10-03 16:46 ` Konrad Rzeszutek Wilk
@ 2011-10-03 16:56 ` Thomas Hellstrom
0 siblings, 0 replies; 16+ messages in thread
From: Thomas Hellstrom @ 2011-10-03 16:56 UTC (permalink / raw)
To: Konrad Rzeszutek Wilk
Cc: linux-kernel, dri-devel, bskeggs, j.glisse, thomas, airlied,
airlied, alexdeucher, xen-devel
On 10/03/2011 06:46 PM, Konrad Rzeszutek Wilk wrote:
>
> It does now (I had a spinlock mishap).. which reminds me - how
> do I test these patches with your vmwgfx driver? I've an old version
> of VMWare Workstation 8, would that do?
>
VMware workstation 8 is OK (it's actually the latest version of
workstation).
You'd need the vmwgfx kernel, driver,
latest mesa master compiled with the "svga" driver and the "dri" and
"xa" state trackers.
xf86-video-vmware, the "vmwgfx" branch.
And you should be fine running 3D.
>>
>>> The swapping code back (so from swap to pool) does not seem to
>>> distinguish it that much - it just allocates a new page - and
>>> then copies from whatever was in the swap cache?
>>>
>>> This is something you were thinking to do in the future I presume?
>>>
>> Yes. If / when I do that, I might be adding a new backend function
>> to put a ttm in an
>> "anonymous state", that is using only pages that can be inserted in
>> the swap cache or passed
>> around to other devices, and to put a ttm in a "device" state, that
>> copies it to device mappable pages.
>>
> OK, that should be no trouble - we would need to expose a function
> call to "detach" the page from the TTM pool (which could mean
> actually allocating a new page for the "other" device, and copying
> it from the "source" to "other" and then freeing the "source).
>
> I am thinking ... you hotplug an high-end radeon while the machine has
> an ATI ES1000 in it, and want to move those pages to the new
> card. The ATI ES1000 can only do up to 4GB, while the new fancy card
> has no such limits (and perhaps does not want to use the TTM DMA
> pool).
>
> Is this what you had in mind?
>
Yes, that's a typical use-case. Or passing pages between an array of
GPGPUs...
>> Thanks,
>> /Thomas
>>
/Thomas
^ permalink raw reply [flat|nested] 16+ messages in thread