lkml.org 
[lkml]   [2014]   [Jul]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Subject[PATCH 69/83] hsa/radeon: Fix initialization of sh_mem registers
Date
This patch fixes a bug in the code flow that caused an override of the sh_mem
registers.
The bug resulted in sh_mem registers being not initialized properly and
overwrite over sh_mem registers for vmid 0 (which is a vmid of non-HSA processes).

Reviewed-by: Ben Goz <ben.goz@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@amd.com>
---
drivers/gpu/hsa/radeon/kfd_device_queue_manager.c | 48 ++++++++++++-----------
1 file changed, 26 insertions(+), 22 deletions(-)

diff --git a/drivers/gpu/hsa/radeon/kfd_device_queue_manager.c b/drivers/gpu/hsa/radeon/kfd_device_queue_manager.c
index 5ec8da7..56875f9 100644
--- a/drivers/gpu/hsa/radeon/kfd_device_queue_manager.c
+++ b/drivers/gpu/hsa/radeon/kfd_device_queue_manager.c
@@ -87,21 +87,25 @@ static void init_process_memory(struct device_queue_manager *dqm, struct qcm_pro
unsigned int temp;
BUG_ON(!dqm || !qpd);

+ /* check if sh_mem_config register already configured */
+ if (qpd->sh_mem_config == 0) {
+ qpd->sh_mem_config =
+ ALIGNMENT_MODE(SH_MEM_ALIGNMENT_MODE_UNALIGNED) |
+ DEFAULT_MTYPE(MTYPE_NONCACHED) |
+ APE1_MTYPE(MTYPE_NONCACHED);
+ qpd->sh_mem_ape1_limit = 0;
+ qpd->sh_mem_ape1_base = 0;
+ }
+
if (qpd->pqm->process->is_32bit_user_mode) {
temp = get_sh_mem_bases_32(qpd->pqm->process, dqm->dev);
qpd->sh_mem_bases = SHARED_BASE(temp);
- qpd->sh_mem_config = PTR32;
+ qpd->sh_mem_config |= PTR32;
} else {
temp = get_sh_mem_bases_nybble_64(qpd->pqm->process, dqm->dev);
qpd->sh_mem_bases = compute_sh_mem_bases_64bit(temp);
- qpd->sh_mem_config = 0;
}

- qpd->sh_mem_config |= ALIGNMENT_MODE(SH_MEM_ALIGNMENT_MODE_UNALIGNED);
- qpd->sh_mem_config |= DEFAULT_MTYPE(MTYPE_NONCACHED);
- qpd->sh_mem_ape1_limit = 0;
- qpd->sh_mem_ape1_base = 0;
-
pr_debug("kfd: is32bit process: %d sh_mem_bases nybble: 0x%X and register 0x%X\n",
qpd->pqm->process->is_32bit_user_mode, temp, qpd->sh_mem_bases);
}
@@ -110,6 +114,8 @@ static void program_sh_mem_settings(struct device_queue_manager *dqm, struct qcm
{
struct mqd_manager *mqd;

+ BUG_ON(qpd->vmid < KFD_VMID_START_OFFSET);
+
mqd = dqm->get_mqd_manager(dqm, KFD_MQD_TYPE_CIK_COMPUTE);
if (mqd == NULL)
return;
@@ -139,12 +145,6 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm, struct queue *
print_queue(q);

mutex_lock(&dqm->lock);
- /* later memory apertures should be initialized in lazy mode */
- if (!is_mem_initialized)
- if (init_memory(dqm) != 0) {
- retval = -ENODATA;
- goto init_memory_failed;
- }

if (dqm->vmid_bitmap == 0 && qpd->vmid == 0) {
retval = -ENOMEM;
@@ -217,7 +217,6 @@ no_hqd:
*allocate_vmid = qpd->vmid = q->properties.vmid = 0;
}
no_vmid:
-init_memory_failed:
mutex_unlock(&dqm->lock);
return retval;
}
@@ -951,20 +950,25 @@ static bool set_cache_memory_policy(struct device_queue_manager *dqm,
qpd->sh_mem_ape1_limit = limit >> 16;
}

- default_mtype = (default_policy == cache_policy_coherent) ? MTYPE_NONCACHED : MTYPE_CACHED;
- ape1_mtype = (alternate_policy == cache_policy_coherent) ? MTYPE_NONCACHED : MTYPE_CACHED;
+ default_mtype = (default_policy == cache_policy_coherent) ?
+ MTYPE_NONCACHED :
+ MTYPE_CACHED;
+
+ ape1_mtype = (alternate_policy == cache_policy_coherent) ?
+ MTYPE_NONCACHED :
+ MTYPE_CACHED;

- qpd->sh_mem_config = ALIGNMENT_MODE(SH_MEM_ALIGNMENT_MODE_UNALIGNED)
+ qpd->sh_mem_config = (qpd->sh_mem_config & PTR32)
+ | ALIGNMENT_MODE(SH_MEM_ALIGNMENT_MODE_UNALIGNED)
| DEFAULT_MTYPE(default_mtype)
| APE1_MTYPE(ape1_mtype);

-
- if (sched_policy == KFD_SCHED_POLICY_NO_HWS)
+ if ((sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
program_sh_mem_settings(dqm, qpd);

-
- pr_debug("kfd: sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n", qpd->sh_mem_config,
- qpd->sh_mem_ape1_base, qpd->sh_mem_ape1_limit);
+ pr_debug("kfd: sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
+ qpd->sh_mem_config, qpd->sh_mem_ape1_base,
+ qpd->sh_mem_ape1_limit);

mutex_unlock(&dqm->lock);
return true;
--
1.9.1


\
 
 \ /
  Last update: 2014-07-11 05:41    [W:0.249 / U:22.792 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site