vfio: fix VFIO mapping failures in ppc64le
Checks
Commit Message
ppc64le failed when using large physical memory. I found problems in my two
commits in the past.
In commit e072d16f8920 ("vfio: fix expanding DMA area in ppc64le"), I added
a sanity check using a mapped address to resolve an issue around expanding
IOMMU window, but this was not enough, since memory allocation can return
memory anywhere dependent on memory fragmentation. DPDK may still skip DMA
mapping and attempts to unmap non-mapped DMA during expanding IOMMU window.
As a result, SPDK apps using large physical memory frequently failed to
proceed the communication with NVMe and/or went into an infinite loop.
The root cause of the bug was in a gap between memory segments managed by
DPDK and firmware-level DMA mapping. DPDK's memory segments don't contain
the state of DMA mapping, and so, the memesg_walk cannot determine if an
iterated memory segment is mapped or not. This resulted in incorrect DMA
maps and unmaps.
At this time, I added the code to avoid iterating non-mapped memory
segments during DMA mapping. The memseg_walk iterates over memory segments
marked as "used", and so, the code sets memory segments that will be
mapped or unmapped as "free" transiently.
The commit db90b4969e2e ("vfio: retry creating sPAPR DMA window") allows
retring different page levels and sizes to create DMA window. However, this
allows page sizes different from hugepage sizes. This inconsistency caused
failures at the time of DMA mapping after the window creation. This patch
fixes to retry only different page levels.
Fixes: e072d16f8920 ("vfio: fix expanding DMA area in ppc64le")
Fixes: db90b4969e2e ("vfio: retry creating sPAPR DMA window")
Signed-off-by: Takeshi Yoshimura <tyos@jp.ibm.com>
---
lib/librte_eal/linux/eal/eal_vfio.c | 76 +++++++++++++----------------
1 file changed, 33 insertions(+), 43 deletions(-)
Comments
> ppc64le failed when using large physical memory. I found problems in my two
> commits in the past.
>
> In commit e072d16f8920 ("vfio: fix expanding DMA area in ppc64le"), I added
> a sanity check using a mapped address to resolve an issue around expanding
> IOMMU window, but this was not enough, since memory allocation can return
> memory anywhere dependent on memory fragmentation. DPDK may still skip DMA
> mapping and attempts to unmap non-mapped DMA during expanding IOMMU window.
> As a result, SPDK apps using large physical memory frequently failed to
> proceed the communication with NVMe and/or went into an infinite loop.
>
> The root cause of the bug was in a gap between memory segments managed by
> DPDK and firmware-level DMA mapping. DPDK's memory segments don't contain
> the state of DMA mapping, and so, the memesg_walk cannot determine if an
> iterated memory segment is mapped or not. This resulted in incorrect DMA
> maps and unmaps.
>
> At this time, I added the code to avoid iterating non-mapped memory
> segments during DMA mapping. The memseg_walk iterates over memory segments
> marked as "used", and so, the code sets memory segments that will be
> mapped or unmapped as "free" transiently.
>
> The commit db90b4969e2e ("vfio: retry creating sPAPR DMA window") allows
> retring different page levels and sizes to create DMA window. However, this
> allows page sizes different from hugepage sizes. This inconsistency caused
> failures at the time of DMA mapping after the window creation. This patch
> fixes to retry only different page levels.
>
> Fixes: e072d16f8920 ("vfio: fix expanding DMA area in ppc64le")
> Fixes: db90b4969e2e ("vfio: retry creating sPAPR DMA window")
>
> Signed-off-by: Takeshi Yoshimura <tyos@jp.ibm.com>
> ---
Reviewed-by: David Christensen <drc@linux.vnet.ibm.com>
On Fri, Jan 17, 2020 at 5:26 AM Takeshi Yoshimura <tyos@jp.ibm.com> wrote:
>
> ppc64le failed when using large physical memory. I found problems in my two
> commits in the past.
Let's hope this is not a "Jamais deux sans trois" ;-).
Since this problem is linked to your usage in SPDK, please test
against -rc2 and confirm this is ok.
>
> In commit e072d16f8920 ("vfio: fix expanding DMA area in ppc64le"), I added
> a sanity check using a mapped address to resolve an issue around expanding
> IOMMU window, but this was not enough, since memory allocation can return
> memory anywhere dependent on memory fragmentation. DPDK may still skip DMA
> mapping and attempts to unmap non-mapped DMA during expanding IOMMU window.
> As a result, SPDK apps using large physical memory frequently failed to
> proceed the communication with NVMe and/or went into an infinite loop.
>
> The root cause of the bug was in a gap between memory segments managed by
> DPDK and firmware-level DMA mapping. DPDK's memory segments don't contain
> the state of DMA mapping, and so, the memesg_walk cannot determine if an
> iterated memory segment is mapped or not. This resulted in incorrect DMA
> maps and unmaps.
>
> At this time, I added the code to avoid iterating non-mapped memory
> segments during DMA mapping. The memseg_walk iterates over memory segments
> marked as "used", and so, the code sets memory segments that will be
> mapped or unmapped as "free" transiently.
>
> The commit db90b4969e2e ("vfio: retry creating sPAPR DMA window") allows
> retring different page levels and sizes to create DMA window. However, this
> allows page sizes different from hugepage sizes. This inconsistency caused
> failures at the time of DMA mapping after the window creation. This patch
> fixes to retry only different page levels.
>
> Fixes: e072d16f8920 ("vfio: fix expanding DMA area in ppc64le")
> Fixes: db90b4969e2e ("vfio: retry creating sPAPR DMA window")
Cc: stable@dpdk.org
>
> Signed-off-by: Takeshi Yoshimura <tyos@jp.ibm.com>
Reviewed-by: David Christensen <drc@linux.vnet.ibm.com>
Applied, thanks.
--
David Marchand
@@ -532,6 +532,17 @@ vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len,
return;
}
+#ifdef RTE_ARCH_PPC_64
+ ms = rte_mem_virt2memseg(addr, msl);
+ while (cur_len < len) {
+ int idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
+
+ rte_fbarray_set_free(&msl->memseg_arr, idx);
+ cur_len += ms->len;
+ ++ms;
+ }
+ cur_len = 0;
+#endif
/* memsegs are contiguous in memory */
ms = rte_mem_virt2memseg(addr, msl);
while (cur_len < len) {
@@ -551,6 +562,17 @@ vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len,
cur_len += ms->len;
++ms;
}
+#ifdef RTE_ARCH_PPC_64
+ cur_len = 0;
+ ms = rte_mem_virt2memseg(addr, msl);
+ while (cur_len < len) {
+ int idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
+
+ rte_fbarray_set_used(&msl->memseg_arr, idx);
+ cur_len += ms->len;
+ ++ms;
+ }
+#endif
}
static int
@@ -1416,16 +1438,11 @@ vfio_spapr_dma_do_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
return 0;
}
-struct spapr_remap_walk_param {
- int vfio_container_fd;
- uint64_t addr_64;
-};
-
static int
vfio_spapr_map_walk(const struct rte_memseg_list *msl,
const struct rte_memseg *ms, void *arg)
{
- struct spapr_remap_walk_param *param = arg;
+ int *vfio_container_fd = arg;
/* skip external memory that isn't a heap */
if (msl->external && !msl->heap)
@@ -1435,10 +1452,7 @@ vfio_spapr_map_walk(const struct rte_memseg_list *msl,
if (ms->iova == RTE_BAD_IOVA)
return 0;
- if (ms->addr_64 == param->addr_64)
- return 0;
-
- return vfio_spapr_dma_do_map(param->vfio_container_fd, ms->addr_64, ms->iova,
+ return vfio_spapr_dma_do_map(*vfio_container_fd, ms->addr_64, ms->iova,
ms->len, 1);
}
@@ -1446,7 +1460,7 @@ static int
vfio_spapr_unmap_walk(const struct rte_memseg_list *msl,
const struct rte_memseg *ms, void *arg)
{
- struct spapr_remap_walk_param *param = arg;
+ int *vfio_container_fd = arg;
/* skip external memory that isn't a heap */
if (msl->external && !msl->heap)
@@ -1456,17 +1470,13 @@ vfio_spapr_unmap_walk(const struct rte_memseg_list *msl,
if (ms->iova == RTE_BAD_IOVA)
return 0;
- if (ms->addr_64 == param->addr_64)
- return 0;
-
- return vfio_spapr_dma_do_map(param->vfio_container_fd, ms->addr_64, ms->iova,
+ return vfio_spapr_dma_do_map(*vfio_container_fd, ms->addr_64, ms->iova,
ms->len, 0);
}
struct spapr_walk_param {
uint64_t window_size;
uint64_t hugepage_sz;
- uint64_t addr_64;
};
static int
@@ -1484,10 +1494,6 @@ vfio_spapr_window_size_walk(const struct rte_memseg_list *msl,
if (ms->iova == RTE_BAD_IOVA)
return 0;
- /* do not iterate ms we haven't mapped yet */
- if (param->addr_64 && ms->addr_64 == param->addr_64)
- return 0;
-
if (max > param->window_size) {
param->hugepage_sz = ms->hugepage_sz;
param->window_size = max;
@@ -1531,20 +1537,11 @@ vfio_spapr_create_new_dma_window(int vfio_container_fd,
/* try possible page_shift and levels for workaround */
uint32_t levels;
- for (levels = 1; levels <= info.ddw.levels; levels++) {
- uint32_t pgsizes = info.ddw.pgsizes;
-
- while (pgsizes != 0) {
- create->page_shift = 31 - __builtin_clz(pgsizes);
- create->levels = levels;
- ret = ioctl(vfio_container_fd,
- VFIO_IOMMU_SPAPR_TCE_CREATE, create);
- if (!ret)
- break;
- pgsizes &= ~(1 << create->page_shift);
- }
- if (!ret)
- break;
+ for (levels = create->levels + 1;
+ ret && levels <= info.ddw.levels; levels++) {
+ create->levels = levels;
+ ret = ioctl(vfio_container_fd,
+ VFIO_IOMMU_SPAPR_TCE_CREATE, create);
}
#endif
if (ret) {
@@ -1585,7 +1582,6 @@ vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
/* check if window size needs to be adjusted */
memset(¶m, 0, sizeof(param));
- param.addr_64 = vaddr;
/* we're inside a callback so use thread-unsafe version */
if (rte_memseg_walk_thread_unsafe(vfio_spapr_window_size_walk,
@@ -1610,14 +1606,9 @@ vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
if (do_map) {
/* re-create window and remap the entire memory */
if (iova + len > create.window_size) {
- struct spapr_remap_walk_param remap_param = {
- .vfio_container_fd = vfio_container_fd,
- .addr_64 = vaddr,
- };
-
/* release all maps before recreating the window */
if (rte_memseg_walk_thread_unsafe(vfio_spapr_unmap_walk,
- &remap_param) < 0) {
+ &vfio_container_fd) < 0) {
RTE_LOG(ERR, EAL, "Could not release DMA maps\n");
ret = -1;
goto out;
@@ -1644,7 +1635,7 @@ vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
/* we're inside a callback, so use thread-unsafe version
*/
if (rte_memseg_walk_thread_unsafe(vfio_spapr_map_walk,
- &remap_param) < 0) {
+ &vfio_container_fd) < 0) {
RTE_LOG(ERR, EAL, "Could not recreate DMA maps\n");
ret = -1;
goto out;
@@ -1691,7 +1682,6 @@ vfio_spapr_dma_map(int vfio_container_fd)
struct spapr_walk_param param;
memset(¶m, 0, sizeof(param));
- param.addr_64 = 0UL;
/* create DMA window from 0 to max(phys_addr + len) */
rte_memseg_walk(vfio_spapr_window_size_walk, ¶m);