Hi all,
Just a quick repost of v2[1] with a small fix for the bug reported by Nate.
To recap, whilst this mostly only improves worst-case performance, those
worst-cases have a tendency to be pathologically bad:
Ard reports general desktop performance with Chromium on AMD Seattle going
from ~1-2FPS to perfectly usable.
Leizhen reports gigabit ethernet throughput going from ~6.5Mbit/s to line
speed.
I also inadvertantly found that the HiSilicon hns_dsaf driver was taking ~35s
to probe simply becuase of the number of DMA buffers it maps on startup (perf
shows around 76% of that was spent under the lock in alloc_iova()). With this
series applied it takes a mere ~1s, mostly of unrelated mdelay()s, with
alloc_iova() entirely lost in the noise.
Robin.
[1] https://www.mail-archive.com/[email protected]/msg19139.html
Robin Murphy (1):
iommu/iova: Extend rbtree node caching
Zhen Lei (3):
iommu/iova: Optimise rbtree searching
iommu/iova: Optimise the padding calculation
iommu/iova: Make dma_32bit_pfn implicit
drivers/gpu/drm/tegra/drm.c | 3 +-
drivers/gpu/host1x/dev.c | 3 +-
drivers/iommu/amd_iommu.c | 7 +--
drivers/iommu/dma-iommu.c | 18 +------
drivers/iommu/intel-iommu.c | 11 ++--
drivers/iommu/iova.c | 114 +++++++++++++++++----------------------
drivers/misc/mic/scif/scif_rma.c | 3 +-
include/linux/iova.h | 8 +--
8 files changed, 62 insertions(+), 105 deletions(-)
--
2.13.4.dirty
From: Zhen Lei <[email protected]>
Checking the IOVA bounds separately before deciding which direction to
continue the search (if necessary) results in redundantly comparing both
pfns twice each. GCC can already determine that the final comparison op
is redundant and optimise it down to 3 in total, but we can go one
further with a little tweak of the ordering (which makes the intent of
the code that much cleaner as a bonus).
Signed-off-by: Zhen Lei <[email protected]>
Tested-by: Ard Biesheuvel <[email protected]>
Tested-by: Zhen Lei <[email protected]>
[rm: rewrote commit message to clarify]
Signed-off-by: Robin Murphy <[email protected]>
---
v3: No change
drivers/iommu/iova.c | 9 +++------
1 file changed, 3 insertions(+), 6 deletions(-)
diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
index 246f14c83944..8f7552dc4e04 100644
--- a/drivers/iommu/iova.c
+++ b/drivers/iommu/iova.c
@@ -289,15 +289,12 @@ private_find_iova(struct iova_domain *iovad, unsigned long pfn)
while (node) {
struct iova *iova = rb_entry(node, struct iova, node);
- /* If pfn falls within iova's range, return iova */
- if ((pfn >= iova->pfn_lo) && (pfn <= iova->pfn_hi)) {
- return iova;
- }
-
if (pfn < iova->pfn_lo)
node = node->rb_left;
- else if (pfn > iova->pfn_lo)
+ else if (pfn > iova->pfn_hi)
node = node->rb_right;
+ else
+ return iova; /* pfn falls within iova's range */
}
return NULL;
--
2.13.4.dirty
The cached node mechanism provides a significant performance benefit for
allocations using a 32-bit DMA mask, but in the case of non-PCI devices
or where the 32-bit space is full, the loss of this benefit can be
significant - on large systems there can be many thousands of entries in
the tree, such that traversing to the end then walking all the way down
to find free space every time becomes increasingly awful.
Maintain a similar cached node for the whole IOVA space as a superset of
the 32-bit space so that performance can remain much more consistent.
Inspired by work by Zhen Lei <[email protected]>.
Tested-by: Ard Biesheuvel <[email protected]>
Tested-by: Zhen Lei <[email protected]>
Signed-off-by: Robin Murphy <[email protected]>
---
v3: Fix potential null dereference on deletion
drivers/iommu/iova.c | 61 ++++++++++++++++++++++++++--------------------------
include/linux/iova.h | 3 ++-
2 files changed, 32 insertions(+), 32 deletions(-)
diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
index d094d1ca8f23..b208f656f6a4 100644
--- a/drivers/iommu/iova.c
+++ b/drivers/iommu/iova.c
@@ -46,6 +46,7 @@ init_iova_domain(struct iova_domain *iovad, unsigned long granule,
spin_lock_init(&iovad->iova_rbtree_lock);
iovad->rbroot = RB_ROOT;
+ iovad->cached_node = NULL;
iovad->cached32_node = NULL;
iovad->granule = granule;
iovad->start_pfn = start_pfn;
@@ -57,48 +58,48 @@ EXPORT_SYMBOL_GPL(init_iova_domain);
static struct rb_node *
__get_cached_rbnode(struct iova_domain *iovad, unsigned long *limit_pfn)
{
- if ((*limit_pfn > iovad->dma_32bit_pfn) ||
- (iovad->cached32_node == NULL))
+ struct rb_node *cached_node = NULL;
+ struct iova *curr_iova;
+
+ if (*limit_pfn <= iovad->dma_32bit_pfn)
+ cached_node = iovad->cached32_node;
+ if (!cached_node)
+ cached_node = iovad->cached_node;
+ if (!cached_node)
return rb_last(&iovad->rbroot);
- else {
- struct rb_node *prev_node = rb_prev(iovad->cached32_node);
- struct iova *curr_iova =
- rb_entry(iovad->cached32_node, struct iova, node);
- *limit_pfn = curr_iova->pfn_lo;
- return prev_node;
- }
+
+ curr_iova = rb_entry(cached_node, struct iova, node);
+ *limit_pfn = min(*limit_pfn, curr_iova->pfn_lo);
+
+ return rb_prev(cached_node);
}
static void
-__cached_rbnode_insert_update(struct iova_domain *iovad,
- unsigned long limit_pfn, struct iova *new)
+__cached_rbnode_insert_update(struct iova_domain *iovad, struct iova *new)
{
- if (limit_pfn != iovad->dma_32bit_pfn)
- return;
- iovad->cached32_node = &new->node;
+ if (new->pfn_lo > iovad->dma_32bit_pfn)
+ iovad->cached_node = &new->node;
+ else
+ iovad->cached32_node = &new->node;
}
static void
__cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
{
struct iova *cached_iova;
- struct rb_node *curr;
+ struct rb_node **curr = NULL;
- if (!iovad->cached32_node)
+ if (free->pfn_hi < iovad->dma_32bit_pfn)
+ curr = &iovad->cached32_node;
+ if (!curr)
+ curr = &iovad->cached_node;
+
+ if (!*curr)
return;
- curr = iovad->cached32_node;
- cached_iova = rb_entry(curr, struct iova, node);
+ cached_iova = rb_entry(*curr, struct iova, node);
- if (free->pfn_lo >= cached_iova->pfn_lo) {
- struct rb_node *node = rb_next(&free->node);
- struct iova *iova = rb_entry(node, struct iova, node);
-
- /* only cache if it's below 32bit pfn */
- if (node && iova->pfn_lo < iovad->dma_32bit_pfn)
- iovad->cached32_node = node;
- else
- iovad->cached32_node = NULL;
- }
+ if (free->pfn_lo >= cached_iova->pfn_lo)
+ *curr = rb_next(&free->node);
}
/* Insert the iova into domain rbtree by holding writer lock */
@@ -135,7 +136,6 @@ static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
{
struct rb_node *prev, *curr = NULL;
unsigned long flags;
- unsigned long saved_pfn;
unsigned long align_mask = ~0UL;
if (size_aligned)
@@ -143,7 +143,6 @@ static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
/* Walk the tree backwards */
spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
- saved_pfn = limit_pfn;
curr = __get_cached_rbnode(iovad, &limit_pfn);
prev = curr;
while (curr) {
@@ -173,7 +172,7 @@ static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
/* If we have 'prev', it's a valid place to start the insertion. */
iova_insert_rbtree(&iovad->rbroot, new, prev);
- __cached_rbnode_insert_update(iovad, saved_pfn, new);
+ __cached_rbnode_insert_update(iovad, new);
spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
diff --git a/include/linux/iova.h b/include/linux/iova.h
index e0a892ae45c0..0bb8df43b393 100644
--- a/include/linux/iova.h
+++ b/include/linux/iova.h
@@ -40,7 +40,8 @@ struct iova_rcache {
struct iova_domain {
spinlock_t iova_rbtree_lock; /* Lock to protect update of rbtree */
struct rb_root rbroot; /* iova domain rbtree root */
- struct rb_node *cached32_node; /* Save last alloced node */
+ struct rb_node *cached_node; /* Save last alloced node */
+ struct rb_node *cached32_node; /* Save last 32-bit alloced node */
unsigned long granule; /* pfn granularity for this domain */
unsigned long start_pfn; /* Lower limit for this domain */
unsigned long dma_32bit_pfn;
--
2.13.4.dirty
From: Zhen Lei <[email protected]>
Now that the cached node optimisation can apply to all allocations, the
couple of users which were playing tricks with dma_32bit_pfn in order to
benefit from it can stop doing so. Conversely, there is also no need for
all the other users to explicitly calculate a 'real' 32-bit PFN, when
init_iova_domain() can happily do that itself from the page granularity.
CC: Thierry Reding <[email protected]>
CC: Jonathan Hunter <[email protected]>
CC: David Airlie <[email protected]>
CC: Sudeep Dutt <[email protected]>
CC: Ashutosh Dixit <[email protected]>
Signed-off-by: Zhen Lei <[email protected]>
Tested-by: Ard Biesheuvel <[email protected]>
Tested-by: Zhen Lei <[email protected]>
[rm: use iova_shift(), rewrote commit message]
Signed-off-by: Robin Murphy <[email protected]>
---
v3: No change
drivers/gpu/drm/tegra/drm.c | 3 +--
drivers/gpu/host1x/dev.c | 3 +--
drivers/iommu/amd_iommu.c | 7 ++-----
drivers/iommu/dma-iommu.c | 18 +-----------------
drivers/iommu/intel-iommu.c | 11 +++--------
drivers/iommu/iova.c | 4 ++--
drivers/misc/mic/scif/scif_rma.c | 3 +--
include/linux/iova.h | 5 ++---
8 files changed, 13 insertions(+), 41 deletions(-)
diff --git a/drivers/gpu/drm/tegra/drm.c b/drivers/gpu/drm/tegra/drm.c
index 518f4b69ea53..81e9ae1ee90b 100644
--- a/drivers/gpu/drm/tegra/drm.c
+++ b/drivers/gpu/drm/tegra/drm.c
@@ -150,8 +150,7 @@ static int tegra_drm_load(struct drm_device *drm, unsigned long flags)
order = __ffs(tegra->domain->pgsize_bitmap);
init_iova_domain(&tegra->carveout.domain, 1UL << order,
- carveout_start >> order,
- carveout_end >> order);
+ carveout_start >> order);
tegra->carveout.shift = iova_shift(&tegra->carveout.domain);
tegra->carveout.limit = carveout_end >> tegra->carveout.shift;
diff --git a/drivers/gpu/host1x/dev.c b/drivers/gpu/host1x/dev.c
index 2c58a390123a..57c8eed0ed71 100644
--- a/drivers/gpu/host1x/dev.c
+++ b/drivers/gpu/host1x/dev.c
@@ -193,8 +193,7 @@ static int host1x_probe(struct platform_device *pdev)
order = __ffs(host->domain->pgsize_bitmap);
init_iova_domain(&host->iova, 1UL << order,
- geometry->aperture_start >> order,
- geometry->aperture_end >> order);
+ geometry->aperture_start >> order);
host->iova_end = geometry->aperture_end;
}
diff --git a/drivers/iommu/amd_iommu.c b/drivers/iommu/amd_iommu.c
index 688e77576e5a..a12e3e12014a 100644
--- a/drivers/iommu/amd_iommu.c
+++ b/drivers/iommu/amd_iommu.c
@@ -63,7 +63,6 @@
/* IO virtual address start page frame number */
#define IOVA_START_PFN (1)
#define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
-#define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
/* Reserved IOVA ranges */
#define MSI_RANGE_START (0xfee00000)
@@ -2010,8 +2009,7 @@ static struct dma_ops_domain *dma_ops_domain_alloc(void)
if (!dma_dom->domain.pt_root)
goto free_dma_dom;
- init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
- IOVA_START_PFN, DMA_32BIT_PFN);
+ init_iova_domain(&dma_dom->iovad, PAGE_SIZE, IOVA_START_PFN);
/* Initialize reserved ranges */
copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
@@ -2912,8 +2910,7 @@ static int init_reserved_iova_ranges(void)
struct pci_dev *pdev = NULL;
struct iova *val;
- init_iova_domain(&reserved_iova_ranges, PAGE_SIZE,
- IOVA_START_PFN, DMA_32BIT_PFN);
+ init_iova_domain(&reserved_iova_ranges, PAGE_SIZE, IOVA_START_PFN);
lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
&reserved_rbtree_key);
diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index 9d1cebe7f6cb..191be9c80a8a 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -292,18 +292,7 @@ int iommu_dma_init_domain(struct iommu_domain *domain, dma_addr_t base,
/* ...then finally give it a kicking to make sure it fits */
base_pfn = max_t(unsigned long, base_pfn,
domain->geometry.aperture_start >> order);
- end_pfn = min_t(unsigned long, end_pfn,
- domain->geometry.aperture_end >> order);
}
- /*
- * PCI devices may have larger DMA masks, but still prefer allocating
- * within a 32-bit mask to avoid DAC addressing. Such limitations don't
- * apply to the typical platform device, so for those we may as well
- * leave the cache limit at the top of their range to save an rb_last()
- * traversal on every allocation.
- */
- if (dev && dev_is_pci(dev))
- end_pfn &= DMA_BIT_MASK(32) >> order;
/* start_pfn is always nonzero for an already-initialised domain */
if (iovad->start_pfn) {
@@ -312,16 +301,11 @@ int iommu_dma_init_domain(struct iommu_domain *domain, dma_addr_t base,
pr_warn("Incompatible range for DMA domain\n");
return -EFAULT;
}
- /*
- * If we have devices with different DMA masks, move the free
- * area cache limit down for the benefit of the smaller one.
- */
- iovad->dma_32bit_pfn = min(end_pfn + 1, iovad->dma_32bit_pfn);
return 0;
}
- init_iova_domain(iovad, 1UL << order, base_pfn, end_pfn);
+ init_iova_domain(iovad, 1UL << order, base_pfn);
if (!dev)
return 0;
diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
index 687f18f65cea..afa3b4e765e7 100644
--- a/drivers/iommu/intel-iommu.c
+++ b/drivers/iommu/intel-iommu.c
@@ -82,8 +82,6 @@
#define IOVA_START_PFN (1)
#define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
-#define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
-#define DMA_64BIT_PFN IOVA_PFN(DMA_BIT_MASK(64))
/* page table handling */
#define LEVEL_STRIDE (9)
@@ -1874,8 +1872,7 @@ static int dmar_init_reserved_ranges(void)
struct iova *iova;
int i;
- init_iova_domain(&reserved_iova_list, VTD_PAGE_SIZE, IOVA_START_PFN,
- DMA_32BIT_PFN);
+ init_iova_domain(&reserved_iova_list, VTD_PAGE_SIZE, IOVA_START_PFN);
lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
&reserved_rbtree_key);
@@ -1933,8 +1930,7 @@ static int domain_init(struct dmar_domain *domain, struct intel_iommu *iommu,
int adjust_width, agaw;
unsigned long sagaw;
- init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN,
- DMA_32BIT_PFN);
+ init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN);
domain_reserve_special_ranges(domain);
/* calculate AGAW */
@@ -4989,8 +4985,7 @@ static int md_domain_init(struct dmar_domain *domain, int guest_width)
{
int adjust_width;
- init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN,
- DMA_32BIT_PFN);
+ init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN);
domain_reserve_special_ranges(domain);
/* calculate AGAW */
diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
index b208f656f6a4..612b8f6ce34d 100644
--- a/drivers/iommu/iova.c
+++ b/drivers/iommu/iova.c
@@ -35,7 +35,7 @@ static void free_iova_rcaches(struct iova_domain *iovad);
void
init_iova_domain(struct iova_domain *iovad, unsigned long granule,
- unsigned long start_pfn, unsigned long pfn_32bit)
+ unsigned long start_pfn)
{
/*
* IOVA granularity will normally be equal to the smallest
@@ -50,7 +50,7 @@ init_iova_domain(struct iova_domain *iovad, unsigned long granule,
iovad->cached32_node = NULL;
iovad->granule = granule;
iovad->start_pfn = start_pfn;
- iovad->dma_32bit_pfn = pfn_32bit + 1;
+ iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad));
init_iova_rcaches(iovad);
}
EXPORT_SYMBOL_GPL(init_iova_domain);
diff --git a/drivers/misc/mic/scif/scif_rma.c b/drivers/misc/mic/scif/scif_rma.c
index 329727e00e97..c824329f7012 100644
--- a/drivers/misc/mic/scif/scif_rma.c
+++ b/drivers/misc/mic/scif/scif_rma.c
@@ -39,8 +39,7 @@ void scif_rma_ep_init(struct scif_endpt *ep)
struct scif_endpt_rma_info *rma = &ep->rma_info;
mutex_init(&rma->rma_lock);
- init_iova_domain(&rma->iovad, PAGE_SIZE, SCIF_IOVA_START_PFN,
- SCIF_DMA_64BIT_PFN);
+ init_iova_domain(&rma->iovad, PAGE_SIZE, SCIF_IOVA_START_PFN);
spin_lock_init(&rma->tc_lock);
mutex_init(&rma->mmn_lock);
INIT_LIST_HEAD(&rma->reg_list);
diff --git a/include/linux/iova.h b/include/linux/iova.h
index 0bb8df43b393..58c2a365c45f 100644
--- a/include/linux/iova.h
+++ b/include/linux/iova.h
@@ -102,7 +102,7 @@ struct iova *reserve_iova(struct iova_domain *iovad, unsigned long pfn_lo,
unsigned long pfn_hi);
void copy_reserved_iova(struct iova_domain *from, struct iova_domain *to);
void init_iova_domain(struct iova_domain *iovad, unsigned long granule,
- unsigned long start_pfn, unsigned long pfn_32bit);
+ unsigned long start_pfn);
struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn);
void put_iova_domain(struct iova_domain *iovad);
struct iova *split_and_remove_iova(struct iova_domain *iovad,
@@ -170,8 +170,7 @@ static inline void copy_reserved_iova(struct iova_domain *from,
static inline void init_iova_domain(struct iova_domain *iovad,
unsigned long granule,
- unsigned long start_pfn,
- unsigned long pfn_32bit)
+ unsigned long start_pfn)
{
}
--
2.13.4.dirty
From: Zhen Lei <[email protected]>
The mask for calculating the padding size doesn't change, so there's no
need to recalculate it every loop iteration. Furthermore, Once we've
done that, it becomes clear that we don't actually need to calculate a
padding size at all - by flipping the arithmetic around, we can just
combine the upper limit, size, and mask directly to check against the
lower limit.
For an arm64 build, this alone knocks 16% off the size of the entire
alloc_iova() function!
Signed-off-by: Zhen Lei <[email protected]>
Tested-by: Ard Biesheuvel <[email protected]>
Tested-by: Zhen Lei <[email protected]>
[rm: simplified more of the arithmetic, rewrote commit message]
Signed-off-by: Robin Murphy <[email protected]>
---
v3: No change
drivers/iommu/iova.c | 40 ++++++++++++++--------------------------
1 file changed, 14 insertions(+), 26 deletions(-)
diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
index 8f7552dc4e04..d094d1ca8f23 100644
--- a/drivers/iommu/iova.c
+++ b/drivers/iommu/iova.c
@@ -129,16 +129,6 @@ iova_insert_rbtree(struct rb_root *root, struct iova *iova,
rb_insert_color(&iova->node, root);
}
-/*
- * Computes the padding size required, to make the start address
- * naturally aligned on the power-of-two order of its size
- */
-static unsigned int
-iova_get_pad_size(unsigned int size, unsigned int limit_pfn)
-{
- return (limit_pfn - size) & (__roundup_pow_of_two(size) - 1);
-}
-
static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
unsigned long size, unsigned long limit_pfn,
struct iova *new, bool size_aligned)
@@ -146,7 +136,10 @@ static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
struct rb_node *prev, *curr = NULL;
unsigned long flags;
unsigned long saved_pfn;
- unsigned int pad_size = 0;
+ unsigned long align_mask = ~0UL;
+
+ if (size_aligned)
+ align_mask <<= __fls(size);
/* Walk the tree backwards */
spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
@@ -156,31 +149,26 @@ static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
while (curr) {
struct iova *curr_iova = rb_entry(curr, struct iova, node);
- if (limit_pfn <= curr_iova->pfn_lo) {
+ if (limit_pfn <= curr_iova->pfn_lo)
goto move_left;
- } else if (limit_pfn > curr_iova->pfn_hi) {
- if (size_aligned)
- pad_size = iova_get_pad_size(size, limit_pfn);
- if ((curr_iova->pfn_hi + size + pad_size) < limit_pfn)
- break; /* found a free slot */
- }
+
+ if (((limit_pfn - size) & align_mask) > curr_iova->pfn_hi)
+ break; /* found a free slot */
+
limit_pfn = curr_iova->pfn_lo;
move_left:
prev = curr;
curr = rb_prev(curr);
}
- if (!curr) {
- if (size_aligned)
- pad_size = iova_get_pad_size(size, limit_pfn);
- if ((iovad->start_pfn + size + pad_size) > limit_pfn) {
- spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
- return -ENOMEM;
- }
+ if (limit_pfn < size ||
+ (!curr && ((limit_pfn - size) & align_mask) < iovad->start_pfn)) {
+ spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
+ return -ENOMEM;
}
/* pfn_lo will point to size aligned address if size_aligned is set */
- new->pfn_lo = limit_pfn - (size + pad_size);
+ new->pfn_lo = (limit_pfn - size) & align_mask;
new->pfn_hi = new->pfn_lo + size - 1;
/* If we have 'prev', it's a valid place to start the insertion. */
--
2.13.4.dirty
Hi Robin,
On 8/22/2017 11:17 AM, Robin Murphy wrote:
> Hi all,
>
> Just a quick repost of v2[1] with a small fix for the bug reported by Nate.
I tested the series and can confirm that the crash I reported on v2
no longer occurs with this version.
> To recap, whilst this mostly only improves worst-case performance, those
> worst-cases have a tendency to be pathologically bad:
>
> Ard reports general desktop performance with Chromium on AMD Seattle going
> from ~1-2FPS to perfectly usable.
>
> Leizhen reports gigabit ethernet throughput going from ~6.5Mbit/s to line
> speed.
>
> I also inadvertantly found that the HiSilicon hns_dsaf driver was taking ~35s
> to probe simply becuase of the number of DMA buffers it maps on startup (perf
> shows around 76% of that was spent under the lock in alloc_iova()). With this
> series applied it takes a mere ~1s, mostly of unrelated mdelay()s, with
> alloc_iova() entirely lost in the noise.
Are any of these cases PCI devices attached to domains that have run
out of 32-bit IOVAs and have to retry the allocation using dma_limit?
iommu_dma_alloc_iova() {
[...]
if (dma_limit > DMA_BIT_MASK(32) && dev_is_pci(dev)) [<- TRUE]
iova = alloc_iova_fast(DMA_BIT_MASK(32)); [<- NULL]
if (!iova)
iova = alloc_iova_fast(dma_limit); [<- OK ]
[...]
}
I am asking because, when using 64k pages, the Mellanox CX4 adapter
exhausts the supply 32-bit IOVAs simply allocating per-cpu IOVA space
during 'ifconfig up' so the code path outlined above is taken for
nearly all subsequent allocations. Although I do see a notable (~2x)
performance improvement with this series, I would still characterize it
as "pathologically bad" at < 10% of iommu passthrough performance.
This was a bit surprising to me as I thought the iova_rcache would
have eliminated the need to walk the rbtree for runtime allocations.
Unfortunately, it looks like the failed attempt to allocate a 32-bit
IOVA actually drops the cached IOVAs that we could have used when
subsequently performing the allocation at dma_limit.
alloc_iova_fast() {
[...]
iova_pfn = iova_rcache_get(...); [<- Fail, no 32-bit IOVAs]
if (iova_pfn)
return iova_pfn;
retry:
new_iova = alloc_iova(...); [<- Fail, no 32-bit IOVAs]
if (!new_iova) {
unsigned int cpu;
if (flushed_rcache)
return 0;
/* Try replenishing IOVAs by flushing rcache. */
flushed_rcache = true;
for_each_online_cpu(cpu)
free_cpu_cached_iovas(cpu, iovad); [<- :( ]
goto retry;
}
}
As an experiment, I added code to skip the rcache flushing/retry for
the 32-bit allocations. In this configuration, 100% of passthrough mode
performance was achieved. I made the same change in the baseline and
measured performance at ~95% of passthrough mode.
I also got similar results by altogether removing the 32-bit allocation
from iommu_dma_alloc_iova() which makes me wonder why we even bother.
What (PCIe) workloads have been shown to actually benefit from it?
Tested-by: Nate Watterson <[email protected]>
-Nate
>
> Robin.
>
> [1] https://www.mail-archive.com/[email protected]/msg19139.html
>
> Robin Murphy (1):
> iommu/iova: Extend rbtree node caching
>
> Zhen Lei (3):
> iommu/iova: Optimise rbtree searching
> iommu/iova: Optimise the padding calculation
> iommu/iova: Make dma_32bit_pfn implicit
>
> drivers/gpu/drm/tegra/drm.c | 3 +-
> drivers/gpu/host1x/dev.c | 3 +-
> drivers/iommu/amd_iommu.c | 7 +--
> drivers/iommu/dma-iommu.c | 18 +------
> drivers/iommu/intel-iommu.c | 11 ++--
> drivers/iommu/iova.c | 114 +++++++++++++++++----------------------
> drivers/misc/mic/scif/scif_rma.c | 3 +-
> include/linux/iova.h | 8 +--
> 8 files changed, 62 insertions(+), 105 deletions(-)
>
--
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project.
Hey Robin, Nate,
On Fri, Aug 25, 2017 at 02:52:41PM -0400, Nate Watterson wrote:
> Tested-by: Nate Watterson <[email protected]>
If nobody has objections here anymore I would merge these patches when
v4.14-rc1 is released. Given that these changes are a bit more intrusive
and the code is shared between a couple of drivers, I'd like to give it
more testing in linux-next.
Robin, can you rebase these patches to v4.14-rc1 when it is released,
add Nate's Tested-by, and resend them to me please?
Thanks,
Joerg