<html> <head> <meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> </head> <body> Am 16.02.22 um 14:11 schrieb Somalapuram, Amaranath: <blockquote type="cite" cite="mid:ed1103b1-835c-e56a-3118-17bd60f0f5f9@amd.com"> <meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> On 2/16/2022 3:41 PM, Christian König wrote: <blockquote type="cite" cite="mid:e9f1e95f-5aac-4a25-51f1-b971a8189d79@amd.com">Am 16.02.22 um 10:49 schrieb Somalapuram Amaranath: <blockquote type="cite">List of register populated for dump collection during the GPU reset. Signed-off-by: Somalapuram Amaranath <a class="moz-txt-link-rfc2396E" href="mailto:Amaranath.Somalapuram@amd.com" moz-do-not-send="true"><Amaranath.Somalapuram@amd.com></a> --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 5 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 95 +++++++++++++++++++++ 2 files changed, 100 insertions(+) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index b85b67a88a3d..57965316873b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -1097,6 +1097,11 @@ struct amdgpu_device { struct amdgpu_reset_control *reset_cntl; uint32_t ip_versions[HW_ID_MAX][HWIP_MAX_INSTANCE]; + + /* reset dump register */ + uint32_t *reset_dump_reg_list; + int n_regs; + struct mutex reset_dump_mutex; </blockquote> I think we should rather use the reset lock for this instead of introducing just another mutex. <blockquote type="cite"> }; static inline struct amdgpu_device *drm_to_adev(struct drm_device *ddev) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c index 164d6a9e9fbb..faf985c7cb93 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c @@ -1609,6 +1609,98 @@ DEFINE_DEBUGFS_ATTRIBUTE(fops_ib_preempt, NULL, DEFINE_DEBUGFS_ATTRIBUTE(fops_sclk_set, NULL, amdgpu_debugfs_sclk_set, "%llu\n"); +static ssize_t amdgpu_reset_dump_register_list_read(struct file *f, + char __user *buf, size_t size, loff_t *pos) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private; + char reg_offset[11]; + int i, r, len = 0; + + if (*pos) + return 0; + + if (adev->n_regs == 0) + return 0; + + for (i = 0; i < adev->n_regs; i++) { + sprintf(reg_offset, "0x%x ", adev->reset_dump_reg_list[i]); + r = copy_to_user(buf + len, reg_offset, strlen(reg_offset)); + + if (r) + return -EFAULT; + + len += strlen(reg_offset); + } </blockquote> You need to hold the lock protecting adev->reset_dump_reg_list and adev->n_regs while accessing those. (BTW: num_regs instead of n_regs would match more what we use elsewhere, but is not a must have). </blockquote> This is read function for user and returns only list of reg offsets, I did not understand correctly ! <blockquote type="cite" cite="mid:e9f1e95f-5aac-4a25-51f1-b971a8189d79@amd.com"> <blockquote type="cite">+ + r = copy_to_user(buf + len, "\n", 1); + + if (r) + return -EFAULT; + + len++; + *pos += len; + + return len; +} + +static ssize_t amdgpu_reset_dump_register_list_write(struct file *f, + const char __user *buf, size_t size, loff_t *pos) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private; + char *reg_offset, *reg, reg_temp[11]; + static int alloc_count; + int ret, i = 0, len = 0; + + do { + reg_offset = reg_temp; + memset(reg_offset, 0, 11); + ret = copy_from_user(reg_offset, buf + len, min(11, ((int)size-len))); + + if (ret) + goto failed; + + reg = strsep(&reg_offset, " "); + + if (alloc_count <= i) { </blockquote> <blockquote type="cite">+ adev->reset_dump_reg_list = krealloc_array( + adev->reset_dump_reg_list, 1, + sizeof(uint32_t), GFP_KERNEL); + alloc_count++; + } + + ret = kstrtouint(reg, 16, &adev->reset_dump_reg_list[i]); </blockquote> This here is modifying adev->reset_dump_reg_list as well and so must be protected by a lock as well. The tricky part is that we can't allocate memory while holding this lock (because we need it during reset as well). One solution for this is to read the register list into a local array first and when that's done swap the local array with the one in adev->reset_dump_reg_list while holding the lock. Regards, Christian. </blockquote> There are 2 situations: 1st time creating list n_regs will be 0 and trace event will not be triggered 2nd time while updating list n_regs is already set and adev->reset_dump_reg_list will have some offsets address (hypothetically speaking during reset + update read values from RREG32 will mix up of old list and new list) its only critical when its freed and n_regs is not 0 </blockquote> No, that won't work like this. See you *must* always hold a lock when reading or writing the array. Otherwise it is perfectly possible that one thread sees only halve of the updates of another thread. The only alternative would be RCU, atomic replace and manual barrier handling, but that would be complete overkill for that feature. Regards, Christian. <blockquote type="cite" cite="mid:ed1103b1-835c-e56a-3118-17bd60f0f5f9@amd.com"> Regards, S.Amarnath <blockquote type="cite" cite="mid:e9f1e95f-5aac-4a25-51f1-b971a8189d79@amd.com"> <blockquote type="cite">+ + if (ret) + goto failed; + + len += strlen(reg) + 1; + i++; + + } while (len < size); + + adev->n_regs = i; + + return size; + +failed: + mutex_lock(&adev->reset_dump_mutex); + kfree(adev->reset_dump_reg_list); + adev->reset_dump_reg_list = NULL; + alloc_count = 0; + adev->n_regs = 0; + mutex_unlock(&adev->reset_dump_mutex); + return -EFAULT; +} + + + +static const struct file_operations amdgpu_reset_dump_register_list = { + .owner = THIS_MODULE, + .read = amdgpu_reset_dump_register_list_read, + .write = amdgpu_reset_dump_register_list_write, + .llseek = default_llseek +}; + int amdgpu_debugfs_init(struct amdgpu_device *adev) { struct dentry *root = adev_to_drm(adev)->primary->debugfs_root; @@ -1618,6 +1710,7 @@ int amdgpu_debugfs_init(struct amdgpu_device *adev) if (!debugfs_initialized()) return 0; + mutex_init(&adev->reset_dump_mutex); ent = debugfs_create_file("amdgpu_preempt_ib", 0600, root, adev, &fops_ib_preempt); if (IS_ERR(ent)) { @@ -1672,6 +1765,8 @@ int amdgpu_debugfs_init(struct amdgpu_device *adev) &amdgpu_debugfs_test_ib_fops); debugfs_create_file("amdgpu_vm_info", 0444, root, adev, &amdgpu_debugfs_vm_info_fops); + debugfs_create_file("amdgpu_reset_dump_register_list", 0644, root, adev, + &amdgpu_reset_dump_register_list); adev->debugfs_vbios_blob.data = adev->bios; adev->debugfs_vbios_blob.size = adev->bios_size; </blockquote> </blockquote> </blockquote> </body> </html>